# forck.live — 9–15 November 2026

> AI lab announcements, news reports and podcasts.

forck.live follows what companies announce about their own products. Press may
corroborate an official story; it never leads and never appears as an item of its
own. A missing date stays missing. A title is the publisher's words — where a
company does not publish in English, an English rendering is listed separately
as `Title in English` and named as this site's translation, never as theirs.

This file is one issue of the weekly archive: every confirmed
first-party announcement of ISO week 2026-W46, 9–15 November 2026, newest
first. The window never moves, so this URL always names the same seven days —
but a post published inside them and detected later is added when it arrives,
so a closed week can still gain entries.

Every entry is one first-party announcement, and carries the same eight fields
in this order before any others:

    Company · Announced · Category · Coverage · Announcement · Group · Source · Record

`Source` is the company's own page the announcement was traced to and `Record`
is this site's page for it. Both are safe to cite.

`Coverage` is a number of outlets, or `not counted` — which means this record
reached no count of its own, never that nobody wrote about it.

`Category` is this tracker's own analysis, and is authoritative wherever it is
stated. `not stated` means either that nothing has analysed the post yet or
that the analysis placed it outside this taxonomy; neither is a verdict about
the announcement.

`Announcement` is whether the post is news of something the company did, rather
than a how-to, a customer story, a changelog listing or a staffing note.

`Group` is which of this record's groups the entry falls in, carried on the
entry itself so it holds where the groups are headings and where they are not.
One of: `models`, `covered`, `announcements`, `routine`.

Anything after `Record` is held detail that only some announcements carry.

- This week's page: https://forck.live/week/2026-W46
- The current issue, and the catalogue of the other weeks: https://forck.live/briefing.md
- Catalogue: https://forck.live/llms.txt

## Announcements

### OpenAI — Our new blog for science

- Company: OpenAI (openai.com)
- Announced: 2026-11-15T22:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/our-new-blog-for-science
- Record: https://forck.live/items/36426-our-new-blog-for-science
- Subject: GPT / ChatGPT / API

OpenAI is starting a science blog where scientists and academic researchers can share what they discover with AI and how they do the work. The company says accelerating scientific discovery is one of the most impactful ways to work toward its mission of ensuring AGI benefits all of humanity. Posts will cover new scientific results from OpenAI, its collaborators and the wider community, practical guidance on using its models for tasks such as scientific computing, data analysis, literature review and experiment planning, and firsthand accounts from researchers. OpenAI points to recent work as evidence of the expanding range of problems its models address: in September an internal system produced a proof addressing the Navier-Stokes existence and smoothness problem, and in early October it released mathematical results including one on the quasi-Riemann hypothesis. In biology, thousands of agents powered by GPT-Rosalind analyzed existing experimental data from millions of mouse brain cells in 32 minutes with PerturbAI and Scripps Research, identifying possible biological mechanisms to test. OpenAI says the capabilities raise questions about how proofs are checked and attributed, and that scientific communities should help shape how the tools are used.

### Tencent — ⚡Meet Hy4 flash, the cheapest model for everyday productivity.

- Company: Tencent (tencent.com)
- Announced: 2026-11-15T16:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://x.com/TencentHunyuan/status/2122217770314301469
- Record: https://forck.live/items/36516-meet-hy4-flash-the-cheapest-model-for-everyday-productivity
- Subject: Hunyuan

Matches Claude Haiku 5.5 (high) on daily work (human blind tests), yet 18x task cost reduction. API pricing (per 1M tokens): $0.003 cached · $0.070 input · $0.281 output. 🔗：https://t.co/hqTDu8QjBH

### Harvey — Wotton Kearney Adopts Harvey Firmwide Across Asia Pacific

- Company: Harvey (harvey.ai)
- Announced: 2026-11-15
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.harvey.ai/blog/wotton-kearney-adopts-harvey-firmwide-across-asia-pacific
- Record: https://forck.live/items/36436-wotton-kearney-adopts-harvey-firmwide-across-asia-pacific
- Subject: Harvey

Wotton Kearney is rolling Harvey out firmwide across its 13 offices in Australia, New Zealand, Singapore and Thailand, after a six-month pilot showed strong engagement. The pilot involved 200 users across the firm's Asia Pacific operations and reached a 94% adoption rate, generating more than 30,000 queries. WK says the rollout is the next phase of its AI, Innovation and Transformation strategy, aimed at helping its people work more effectively and delivering greater value for clients. Harvey sits inside a broader AI ecosystem at the firm that supports productivity, legal research, drafting, document review and analysis. As part of the next phase, WK will explore working more closely with clients through Shared Spaces, which the firm describes as secure AI-enabled environments for information sharing and legal workflow innovation with governance and confidentiality protections. Iain McGuire, WK's Chief Technology & AI Officer, said the pilot showed that when AI is implemented responsibly, people embrace it, and that the focus is now on practical outcomes and strong governance.

### Ideogram — How to Turn an Image into an Editable SVG: Workflow and Examples

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-15
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/image-to-editable-svg
- Record: https://forck.live/items/36386-how-to-turn-an-image-into-an-editable-svg-workflow-and-examples
- Subject: Ideogram image models

Ideogram published a walkthrough of turning a PNG or JPG into an editable SVG using its Vectorizer app, following two AI-generated images through the conversion. The post says the hotel illustration became 62 vector paths, and after a navy swatch of #1E3140 was replaced with a plum #6B3F5B, the downloaded file kept every path's geometry unchanged while four path fills changed. Vectorizer offers Simple, Detailed and Auto detail settings; the bookshop conversion showed 10 credits with either Auto or Detailed selected, so Detailed added no cost for that image. The source selector accepts up to 100 images at once, and images already in Ideogram can be sent to Vectorizer from the arrow beside Edit. The post recommends starting with a clean, large source, since enlarging a blurry screenshot will not recover missing contours, and reviewing the result at both full view and close-up. It notes that renaming a PNG to SVG does not create vector objects, and that Vectorizer exports use gradient fills alongside solid colors to represent tonal variation.

### Ideogram — How to Fix Misspelled Words in an AI Image Without Recreating It

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-15
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/fix-misspelled-words-ai-image
- Record: https://forck.live/items/36385-how-to-fix-misspelled-words-in-an-ai-image-without-recreating-it
- Subject: Ideogram image models

Ideogram has published a walkthrough showing two ways to fix a misspelled word in a generated image without recreating the whole picture. The guide uses a sample skincare graphic whose headline read "DAILY GOLW" above "HYDRATING SERUM," with the bottle, pink background and layout already usable. In Studio, Layerize Text removes detected lettering, reconstructs the area beneath it and recreates the words as editable text layers, and in the test it changed GOLW to GLOW during extraction without any typing. Layerize Text uses credits, so Ideogram advises comparing the result with the source before exporting. The second route, Precise Edit, keeps the original image and redraws only the masked word: the guide says to cover the whole word including its letter edges, leave the context window on Auto, and describe the replacement in the "What should change?" field. Ideogram says Precise Edit retains the original image dimensions and leaves pixels outside the context window untouched, and that the test produced the correct word on the first attempt. The guide also warns that a PNG export flattens the composition and does not carry text layers into a new upload.

### Ideogram — 7 Best AI Vector Generators in 2026 for Editable SVG Graphics

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-15
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/best-ai-vector-generators
- Record: https://forck.live/items/36384-7-best-ai-vector-generators-in-2026-for-editable-svg-graphics
- Subject: Ideogram image models

Ideogram has published a hands-on comparison of AI vector generators, testing its own Vector model and Vectorizer against Recraft on the same brief. The test asked for a coral-orange fox curled around dark green fern leaves with "FERN & FORM" lettering, then checked whether each tool's SVG could be revised without redrawing it. Ideogram Vector 1 at Medium quality, 1:1 and one output cost 16 credits and produced a 2048 × 2048 viewBox file with 34 paths, no image elements and no live-text elements, with the lettering outlined. Changing the main orange swatch from #FFA176 to #E97854 and a smaller detail from #F68059 to #DE5436 left the path data identical, so the approved drawing stayed intact. Uploading a flat-color PNG to Ideogram's Vectorizer on Auto detail cost 10 credits and returned an SVG with 34 paths, no embedded image and no gradient definitions, its main coral fill matching the source at #E97854. The comparison notes that outlined lettering cannot be clicked and retyped, that a palette prompt is only a starting point for brand work, and that downloading an SVG does not by itself turn embedded pixels into editable shapes.

### Ideogram — 6 Best AI Image Generators for Combining Multiple Reference Images in 2026

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-15
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/best-ai-image-generators-multiple-reference-images
- Record: https://forck.live/items/36383-6-best-ai-image-generators-for-combining-multiple-reference-images-in-2026
- Subject: Ideogram image models

Ideogram has published a comparison of AI image generators that combine multiple reference images, running the same three source assets through Ideogram, ChatGPT Images and Gemini and then asking each to shrink the product. The test gives one image the product, one the setting and one the photographic treatment, then checks whether each tool borrows the lighting without adding the bowl to the scene. Ideogram says the controls that matter are role assignment, influence over how heavily a reference counts, and revision after the composition works. All three tools kept the MORA bottle recognizable and warmed the scene without inserting the bowl, though Ideogram produced a pronounced pink setting, ChatGPT a softer peach tone and Gemini a warm, grainy result with a strong directional shadow. Each tool also made the bottle smaller on request, and Ideogram's Precise Edit kept the pink setting and countertop patterns consistent while Gemini's adjustment was subtler. Ideogram's downloads were 2048 × 2048 pixels, the largest of the three, against 1254 × 1254 for ChatGPT and 1024 × 1024 for Gemini. FLUX.2, Adobe Firefly and Midjourney round out the shortlist for their image-index instructions, influence controls and style references.

### Luma — A smart organizer for your Luma canvas

- Company: Luma (lumalabs.ai)
- Announced: 2026-11-14
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://lumalabs.ai/news/ai-board-manager
- Record: https://forck.live/items/36067-a-smart-organizer-for-your-luma-canvas
- Subject: Ray / Dream Machine

Luma's agent can now organize a canvas board, grouping and labeling assets by what belongs together so teams keep working on the board that holds the project. The company says big brand projects outgrow their boards, and finding assets starts to take longer than making the next one. The agent reads what is in each asset, what it was made from and what sits around it, telling a reference photo from a generated variation and following which versions came from which image. Boards can be grouped by product, concept, campaign, state of production or channel, and the agent works through large boards section by section, up to 10,000 assets. Users can give it a rule, such as sorting by background color or grouping by product and campaign concept, and can organize a whole board, a frame or a selection. The agent never rearranges a board on its own: it offers to group new work, nothing moves until the user accepts, and any organizing run can be reversed without losing later edits.

### Amazon — Production MLOps at GS Caltex with Amazon SageMaker Canvas, part 1

- Company: Amazon (amazon.com)
- Announced: 2026-11-13T23:22:09+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/production-mlops-at-gs-caltex-with-amazon-sagemaker-canvas-part-1/
- Record: https://forck.live/items/36116-production-mlops-at-gs-caltex-with-amazon-sagemaker-canvas-part-1
- Subject: Bedrock / Nova

GS Caltex built a handoff process that moves models validated in Amazon SageMaker Canvas into SageMaker Pipelines for production deployment and scheduled batch inference. The South Korean energy and chemical company splits the work: process engineers build and validate models in SageMaker Canvas without writing code, while ML engineers configure the Pipelines workflows that deploy them. The gap was that moving assets from Canvas into Pipelines still required integration work, with ML engineers setting production data paths, supplying model configuration, and connecting preprocessing and inference jobs. GS Caltex automated that handoff while reusing the assets created in SageMaker Canvas, including the Data Wrangler preprocessing flow and the trained model. GS Caltex reports that its ML engineers now deploy models within a few hours, compared with about a week before, and that running batch inference every 30 minutes has meaningfully lowered its inference costs by reducing the need for continuously running endpoints. The reference implementation, published in a GitHub repository, uses a sample loan classification model, and the post covers preprocessing reuse, model approval and deployment, pipeline construction, and scheduled inference.

### Amazon — Route coding agent requests to the right model on Amazon Bedrock with NVIDIA NeMo Switchyard

- Company: Amazon (amazon.com)
- Announced: 2026-11-13T22:30:12+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/route-coding-agent-requests-to-the-right-model-on-amazon-bedrock-with-nvidia-nemo-switchyard/
- Record: https://forck.live/items/36107-route-coding-agent-requests-to-the-right-model-on-amazon-bedrock-with-nvidia
- Subject: Bedrock / Nova

Amazon published a walkthrough for deploying LiteLLM with NVIDIA's open source NeMo Switchyard routing plugin on Amazon Bedrock, so teams can send each coding agent request to a different model. The post argues that price per token does not reveal what a whole coding task costs, because the total depends on how many turns a model needs, how much history each turn replays, and how often a task is retried or never finishes. It recommends judging routing by cost per solved task, the total cost of runs divided by the number of runs whose tests pass. The sample deploys LiteLLM and the Switchyard plugin in one container on Amazon ECS on AWS Fargate in private subnets, with AWS WAF filtering requests and an Application Load Balancer forwarding only three paths. The integration supports Stage routing, which reads recent tool activity and results in the request history and picks between two candidate models, and Random routing, which picks from configured models with optional weights for traffic splits and A/B comparisons. Neither strategy calls a judge model, and classifier-backed strategies such as Task and Composite routing need intermediate model calls the integration does not support. The sample installs switchyard-server 0.3.0 and runs the published LiteLLM v1.102.0 image, and NVIDIA documents switchyard-server 0.3.0 as a demo and evaluation component whose configuration can change between releases.

### Amazon — Architecture patterns for fast session startup on Amazon Bedrock AgentCore runtime

- Company: Amazon (amazon.com)
- Announced: 2026-11-13T22:02:39+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://aws.amazon.com/blogs/machine-learning/architecture-patterns-for-fast-session-startup-on-amazon-bedrock-agentcore-runtime/
- Record: https://forck.live/items/36090-architecture-patterns-for-fast-session-startup-on-amazon-bedrock-agentcore
- Subject: Bedrock / Nova

Amazon's AgentCore runtime now starts agent sessions from a snapshot under platformVersion "V2", preparing the container once instead of for every session. On V1, AgentCore pulled the container image from Amazon Elastic Container Registry, started the container, and ran initialization code before the first request, so cold start grew with image size. V2 starts the container in a fresh microVM, waits for the first healthy ping within 120 seconds, clears the file cache, freezes the microVM, and stores memory and disk as a snapshot. Each create or update now takes a few minutes while the snapshot is built, a change from V1. Restored sessions resume with imports, objects and clients in place, but file reads do not carry over and image files are fetched on first read. AgentCore runtime also now supports build hooks, which the post covers for agents that run separate processes such as a CLI or browser binary.

### Amazon — OpenFold3 on Amazon SageMaker JumpStart: Predict biomolecular structures

- Company: Amazon (amazon.com)
- Announced: 2026-11-13T21:53:40+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/openfold3-on-amazon-sagemaker-jumpstart-predict-biomolecular-structures/
- Record: https://forck.live/items/36087-openfold3-on-amazon-sagemaker-jumpstart-predict-biomolecular-structures
- Subject: Bedrock / Nova

Amazon's walkthrough shows drug discovery teams how to deploy OpenFold3-preview, an open source biomolecular structure prediction model, from SageMaker JumpStart into their own AWS accounts. The OpenFold Consortium and the AlQuraishi Lab at Columbia University developed the model, which predicts the joint 3D structure of proteins, DNA, RNA, small-molecule ligands and the complexes they form in a single prediction. Because folding time grows with the number and length of chains and real-world complexes often take several minutes, the model runs on SageMaker Asynchronous Inference rather than a real-time endpoint, which caps requests at 60 seconds. Requests go to Amazon S3 and are invoked with InvokeEndpointAsync, with an invocation timeout of up to 1 hour and results written to an S3 location the user chooses. The endpoint queues bursts of requests internally and can scale to zero instances when the queue is empty, which Amazon says helps avoid paying for idle instances between experiments. The endpoint runs with network isolation, and users can supply their own multiple sequence alignments to improve accuracy while helping keep sequences private. The developers release the code and model weights under the Apache 2.0 license and describe OpenFold3-preview as a research preview; Amazon notes endpoints are billed per instance-hour while running, including while idle.

### Google Research — Advancing science & engineering with Antigravity’s /teamwork

- Company: Google Research (research.google)
- Announced: 2026-11-13T20:47:18+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://research.google/blog/advancing-science-engineering-with-antigravitys-teamwork-feature/
- Record: https://forck.live/items/36066-advancing-science-engineering-with-antigravity-s-teamwork
- Subject: Research

Google Research is making Antigravity's /teamwork multiagent tools available externally after they helped human experts solve long-open problems in math, science and engineering. Teams across Google and academic partners used Gemini 4 Argon, Carbon and Hydrogen with /teamwork workflows to produce new results in information theory, formal verification, molecular structure and kernel optimization. The system is built to address two limits of current language models: unpredictable generation, which leaves a gap between a single attempt and the correct answer, and attention that disperses over multi-day tasks. It pairs parallel exploration with dedicated falsifiers such as the Lean 4 proof checker, unmodified software test suites, execution simulators, web search and hardware counters to verify which candidates work. Context is partitioned by construction, so each specialized agent receives only the files and tools it needs, while coordinators hand off to fresh successors and critics verify results from clean contexts. Workflows are expressed as declarative self-improving natural-language patterns, and each prompt is routed to the pattern best suited to the task, including /boost, a sequential solve-and-verify loop. Many features are available in Antigravity today, with more updates promised in coming weeks for external users.

### GitHub — Upcoming deprecation of selected GitHub Copilot models in December

- Company: GitHub (github.com)
- Announced: 2026-11-13T20:40:05+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://github.blog/changelog/2026-11-13-upcoming-deprecation-of-selected-github-copilot-models-in-december
- Record: https://forck.live/items/36070-upcoming-deprecation-of-selected-github-copilot-models-in-december
- Subject: Copilot

We will deprecate the following models across all GitHub Copilot experiences (including Copilot Chat, inline edits, ask and agent modes, and code completions) on December 14, 2026 *Claude Haiku 4.5 will remain available to individual GitHub Copilot subscribers on annual plans.

### Cloudflare — Kitesurf is now open source

- Company: Cloudflare (cloudflare.com)
- Announced: 2026-11-13T20:20:21+00:00
- Category: not stated
- Coverage: 1 outlet
- Announcement: yes
- Group: covered
- Source: https://blog.cloudflare.com/kitesurf-open-source/
- Record: https://forck.live/items/36059-kitesurf-is-now-open-source
- Subject: Cloudflare AI

Kitesurf , our stateless browser for AI agents that runs entirely on top of Workers, in August and followed up in September with WebMCP support, new APIs, and improved compatibility.

### Amazon — Cost-optimizing AssemblyAI speech-to-text inference on AWS with Amazon EC2 Spot

- Company: Amazon (amazon.com)
- Announced: 2026-11-13T19:35:35+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/cost-optimizing-assemblyai-speech-to-text-inference-on-aws-with-amazon-ec2-spot/
- Record: https://forck.live/items/36044-cost-optimizing-assemblyai-speech-to-text-inference-on-aws-with-amazon-ec2-spot
- Subject: Bedrock / Nova

Amazon's post explains how AssemblyAI runs its speech-to-text inference on Amazon EC2 Spot Instances to cut GPU costs while keeping its SLA commitments. Speech-to-text inference is computationally intensive, with each audio file triggering a pipeline of deep learning models on NVIDIA GPUs, which at scale means thousands of GPU instances running continuously. Running entirely on On-Demand Instances would be cost-prohibitive, and Spot Instances can be reclaimed with a 2-minute warning, so AssemblyAI needed an architecture that absorbs interruptions gracefully. The platform runs on Amazon ECS with capacity providers that mix Spot and On-Demand capacity, spanning g4dn, g5 and g6 GPU instance families and using the price-capacity optimized allocation strategy. When a Spot Instance gets an interruption notice, an EventBridge-triggered Lambda returns the in-flight job to the Main Job SQS Queue for another ECS task to pick up, and jobs that exhaust retries go to a Dead Letter SQS Queue. AssemblyAI deploys regionally in US West (Oregon) us-west-2 and Europe (London) eu-west-2 to meet data residency rules for customers in healthcare, finance and legal.

### Amazon — Cost attribution and controls for AI developer and productivity tools on Amazon Bedrock

- Company: Amazon (amazon.com)
- Announced: 2026-11-13T18:44:02+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/cost-attribution-and-controls-for-ai-developer-and-productivity-tools-on-amazon-bedrock/
- Record: https://forck.live/items/36023-cost-attribution-and-controls-for-ai-developer-and-productivity-tools-on
- Subject: Bedrock / Nova

Amazon's guide shows FinOps and AI platform teams how to attribute Amazon Bedrock inference spend to individual users through IAM identity, without instrumenting API calls. Bedrock records the IAM ARN of every caller, including the session name for federated users, and publishes it to CUR 2.0 in the line_item_iam_principal column, so per-user spend is queryable alongside the dollars. The approach works the same for any tool routing inference through the bedrock-runtime or bedrock-mantle endpoints, including Claude Code, Claude Desktop, OpenAI Codex, ChatGPT Work and OpenCode. The guide walks through configuring IAM Identity Center with an external identity provider, flowing session tags such as team and cost center into CUR 2.0, and querying the data with Amazon Athena. It notes that CUR aggregates usage per principal and per usage type for each export interval and does not contain per-request cost; per-prompt detail requires model invocation logs or request metadata tagging. Claude Platform on AWS and Claude for Enterprise route inference through Anthropic-managed infrastructure, where usage is visible in Anthropic's admin console and spend limits are enforced by Anthropic.

### GitHub — GitHub Enterprise Server support in the GitHub Copilot app

- Company: GitHub (github.com)
- Announced: 2026-11-13T18:27:13+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://github.blog/changelog/2026-11-13-github-enterprise-server-support-in-the-github-copilot-app
- Record: https://forck.live/items/36022-github-enterprise-server-support-in-the-github-copilot-app
- Subject: Copilot

GitHub made GitHub Enterprise Server support in the GitHub Copilot app generally available, after the feature spent time as an opt-in preview. Users can add a GHES account to work with remote repositories, issues and pull requests from inside the app. Signing in with a GitHub.com or GHE.com account comes first, then users go to Settings > Accounts > Add account and choose GitHub Enterprise Server. The feature requires GitHub Copilot app 1.2.0 or later. Cloud sessions are not available for GHES.

### GitHub — Gemini 4 Hydrogen in GitHub Copilot

- Company: GitHub (github.com)
- Announced: 2026-11-13T18:23:48+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://github.blog/changelog/2026-11-13-gemini-4-hydrogen-in-github-copilot
- Record: https://forck.live/items/36038-gemini-4-hydrogen-in-github-copilot
- Subject: Copilot

GitHub has added Gemini 4 Hydrogen, Google's latest Flash model, to Copilot for Pro, Pro+, Max, Business and Enterprise users. The model is built for fast, efficient agentic coding, tool use and multistep workflows, and the source says it improves on Gemini 3.8 Flash in following instructions. It is billed at provider list pricing under usage-based billing, with details in GitHub's Copilot models and pricing documentation. Users can pick it in the model picker across Visual Studio Code, Visual Studio, Copilot CLI, the Copilot coding agent, the Copilot app, github.com, GitHub Mobile on iOS and Android, JetBrains IDEs, Xcode and Eclipse. Rollout is gradual, so the model may not appear for everyone right away. Copilot Enterprise and Copilot Business administrators can manage access through the model policy in Copilot settings, where new models are enabled by default unless an administrator has turned off the global default or explicitly disabled the model.

### Amazon — Build a model improvement flywheel with Amazon SageMaker HyperPod Data Capture

- Company: Amazon (amazon.com)
- Announced: 2026-11-13T18:01:15+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/build-a-model-improvement-flywheel-with-amazon-sagemaker-hyperpod-data-capture/
- Record: https://forck.live/items/36012-build-a-model-improvement-flywheel-with-amazon-sagemaker-hyperpod-data-capture
- Subject: Bedrock / Nova

Amazon SageMaker HyperPod now provides Data Capture, a built-in capability that records live inference inputs and outputs to Amazon S3 at three independent tiers, with minimal latency impact and zero impact on model availability. The feature addresses a gap for teams running foundation models in production: without a record of production inference data, they cannot detect drift, find failure patterns, or use real interactions to improve the model. The walkthrough runs on a single HyperPod cluster, where a home-appliance support router captures its traffic, Athena queries surface a new-product blind spot, a SQL policy labels the captured records, and a fine-tune on that data teaches the router the new product. Tiers 1 and 3 both store full payloads, so enabling both stores every record twice; Amazon suggests Tier 1 for compliance and Model Monitor, Tier 2 when only traffic metadata is needed, and Tier 3 when the payloads will be used. Because Tier 3 records complete payloads that can include names, addresses and order numbers, Amazon advises protecting that data with sampling, AWS KMS encryption, S3 lifecycle rules and PII redaction before it becomes training data. The example uses Qwen2.5-3B-Instruct, which Amazon says is small enough to fine-tune in about a minute and fits on a single L40S, with the inference operator serving on one node and the training operator fine-tuning on another.

### NVIDIA — Build a Robot Pick-and-Place Simulation with SimReady Assets

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-13T18:00:39+00:00
- Category: not stated
- Coverage: 1 outlet
- Announcement: no
- Group: covered
- Source: https://huggingface.co/blog/nvidia/simready-assets-robot-pick-and-place
- Record: https://forck.live/items/36004-build-a-robot-pick-and-place-simulation-with-simready-assets

NVIDIA's walkthrough shows how to build a beverage-can pick-and-place simulation in Isaac Sim using SimReady assets: the UR17G arm, Robotiq 2F-140 gripper and a can. SimReady means an OpenUSD asset has been prepared with the properties a defined simulation use case needs, such as collisions, mass, joints and joint drives, and SimReady Central is an open asset library on Hugging Face where those assets can be reused. The guide walks through downloading the three asset packages, composing the scene, and adding task code that defines the arm-gripper assembly, work surface, starting positions and gripper controls. It advises testing joints, movement and contact before running the task, because a robot can look correct while stationary but interact with objects incorrectly once the simulation runs. The example uses Isaac Sim 6.2 and a frontier model, GPT-6 Sol, to generate and run the task code, with the caveat that generated code and results can vary and should be inspected before adapting.

### Perplexity — Teammates are now available in the Perplexity Mac app for Max subscribers.

- Company: Perplexity (perplexity.ai)
- Announced: 2026-11-13T17:30:03+00:00
- Category: availability-change
- Coverage: 3 outlets
- Announcement: yes
- Group: covered
- Source: https://x.com/perplexity_ai/status/2121294533879267683
- Record: https://forck.live/items/35991-teammates-are-now-available-in-the-perplexity-mac-app-for-max-subscribers

Perplexity has made Teammates available in its Mac app for Max subscribers, extending the agents to desktop users on that plan. Teammates take on work for the subscriber, and the company says they can operate individually or alongside other teammates. According to the post, the work they handle covers QA, customer service, research and growth marketing, among other tasks. Perplexity points readers to a link to get started with teammates.

### Perplexity — Gemini 4 Hydrogen; older Gemini Flash models removed from the model docs

- Company: Perplexity (perplexity.ai)
- Announced: 2026-11-13T17:28:04+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://docs.perplexity.ai/docs/resources/changelog#november-2026
- Record: https://forck.live/items/35975-gemini-4-hydrogen-older-gemini-flash-models-removed-from-the-model-docs
- Subject: Sonar

The Agent API now supports google/gemini-4-hydrogen . Pricing is $1.50 per million input tokens, $0.075 per million cached input tokens, and $7.50 per million output tokens, including reasoning tokens.

### Meta — Cuttlefish on Mac: Running Android’s Virtual Device Framework Natively on Apple Silicon

- Company: Meta (meta.com)
- Announced: 2026-11-13T17:00:39+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://engineering.fb.com/2026/11/13/android/cuttlefish-on-mac/
- Record: https://forck.live/items/35958-cuttlefish-on-mac-running-android-s-virtual-device-framework-natively-on-apple
- Subject: Llama / infrastructure

Meta’s Reality Labs builds products such as AI glasses . Emulators provide us with a way to scale development and testing without relying on scarce physical hardware.

### Google Cloud — New AI agent observability and governance across BigQuery and Looker

- Company: Google Cloud (cloud.google.com)
- Announced: 2026-11-13T17:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://cloud.google.com/blog/products/data-analytics/enterprise-ai-agent-observability-with-bigquery-and-looker/
- Record: https://forck.live/items/35961-new-ai-agent-observability-and-governance-across-bigquery-and-looker
- Subject: Gemini Enterprise

As enterprises move AI agents from experimental prototypes into production, data and platform teams are discovering that deploying an agent is only half the battle — you also need to see how it reasons, which data it touches, and what it costs to run.

### OpenAI — How Mozilla uses GPT-6 Cyber to find and fix Firefox vulnerabilities

- Company: OpenAI (openai.com)
- Announced: 2026-11-13T16:56:45+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.youtube.com/watch?v=PkAmyub12K8
- Record: https://forck.live/items/35959-how-mozilla-uses-gpt-6-cyber-to-find-and-fix-firefox-vulnerabilities

Mozilla is pairing its security engineering with OpenAI's GPT-6 Cyber to strengthen Firefox's defenses, according to a discussion between Mozilla's Brian Grinstead and OpenAI's Clint Gibler. The team uses AI agents to prioritize investigations, test and verify vulnerabilities, find related bug patterns, and help engineers develop fixes. GPT-6 Cyber revealed that some known bugs could be more serious than initially understood, which helped Mozilla decide which issues needed attention first. Across its security workflow, Mozilla saw about an 8× increase in bug discovery and fixes from 2025 to 2026. Grinstead also shared what the team is learning and its ambitions for the future of browser security.

### Hume — Hume Audio Decision Benchmarks

- Company: Hume (hume.ai)
- Announced: 2026-11-13T16:00:00+00:00
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://www.hume.ai/blog/hume-audio-decision-benchmarks
- Record: https://forck.live/items/35951-hume-audio-decision-benchmarks
- Subject: EVI / Octave

Hume has introduced Hume Audio Decision, a model that estimates how people would judge an audio clip without needing a transcript, answering questions such as how natural a voice sounds. The output is a distribution over answers, showing how listeners split when they disagree, rather than a single verdict. In tests on four public human-preference benchmarks, AudioMOS-AES, NISQA, SOMOS and DataPoint, Hume says the model scored highest across 11 models, a comparison that includes Google's Gemini 4 Hydrogen and Gemini Decision 1. Hume also compared the model with individual raters: in six of seven comparisons it agreed with the rest of the crowd more closely than a single held-out rater did. On DataPoint, a single held-out rater picked the crowd's majority answer 56.7% of the time, while Hume Audio Decision picked it 68.6% of the time. Hume reports the lowest mean latency it measured, 0.22 seconds per question, and says the model costs less than every model that scored 0.73 or higher overall. On the overall score, the mean of the three group means, Hume Audio Decision ranks first at 0.800, ahead of Gemini 3.8 Flash at 0.770 and Gemini 4 Hydrogen at 0.755.

### Harvey — Building Ethical Wall Enforcement Into Harvey

- Company: Harvey (harvey.ai)
- Announced: 2026-11-13T16:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.harvey.ai/blog/building-ethical-wall-enforcement-into-harvey
- Record: https://forck.live/items/35935-building-ethical-wall-enforcement-into-harvey
- Subject: Harvey

Harvey is building ethical wall enforcement into its platform so that a law firm's matter restrictions carry over to the conversations, tables and drafts the AI creates from protected documents. Firms keep matter records and wall policies in their own systems, and Harvey imports them through provider adapters that translate each wall provider's data into one common model, with each firm's data staying in its own private Harvey workspace. Import is split into two jobs: Client Matter sync finds matters and links or creates them in Harvey, while Wall sync refreshes the policy for matters already linked, matching people to Harvey users by email address. Harvey says the jobs run separately so each can keep its own schedule, a failure in one does not block the other, and firms can adopt matters before turning on walls. Sync is designed so a failure does not open a wall: if a policy cannot be fetched, Harvey keeps the last stored policy, and if a person on an exclusionary wall cannot be looked up, the update is skipped rather than removing anyone from a block list. Bulk syncs run on Temporal with retries, resumable batches and sync reports, and a Client Matter admin can refresh one matter's policy on demand. Harvey is also moving the Client Matter picker to a dedicated search index built on turbopuffer, one entry per matter per workspace, to keep search fast in a workspace with about 1.7 million matters.

### AMD — Meet the AMD Developer Assistant 🤖

- Company: AMD (amd.com)
- Announced: 2026-11-13T15:03:03+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://x.com/AIatAMD/status/2121257539870261292
- Record: https://forck.live/items/35912-meet-the-amd-developer-assistant

Need quick documentation, code examples, or troubleshooting help? Get instant answers pulled directly from curated AMD resources, including: • AMD ROCm platform docs & blogs • AMD Ryzen AI developer docs • GPUOpen SDKs, manuals & tools • AMD AI Playbooks...

### Hugging Face — llama.cpp v1.0.0 : A stronger foundation for the local AI ecosystem

- Company: Hugging Face (huggingface.co)
- Announced: 2026-11-13T14:05:22+00:00
- Category: developer-tool-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/ggml-org/llama-cpp-v1
- Record: https://forck.live/items/35879-llama-cpp-v1-0-0-a-stronger-foundation-for-the-local-ai-ecosystem
- Subject: Platform

Hugging Face's llama.cpp project released libraries v1.0.0, committing to semantic versioning so downstream developers can see when API and ABI breakages land. The project says a flexible but stable API is paramount because many applications already depend on its library interfaces, and it will now mark incompatibilities with major-version bumps. The same rules apply to all ggml-org projects using semantic versioning, and the team says it will be stricter about making those boundaries visible as adoption grows. llama.cpp runs multi-modal LLMs across hardware from consumer machines to enterprise servers through quantization, optimized backends and a portable C/C++ implementation. It supports more than 150 architecture variants and all mainstream speculative decoding methods, with each ggml backend compilable as a dynamic library and optionally loaded at runtime. The team credits hardware support from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel for expanding self-hosted CI runners, and HF Buckets and HF Jobs for scaling capacity to check every change on master.

### Suno — Suno Bridge: Your DAW, Suno-powered.

- Company: Suno (suno.com)
- Announced: 2026-11-13T14:00:00+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://suno.com/blog/introducing-bridge
- Record: https://forck.live/items/35882-suno-bridge-your-daw-suno-powered
- Subject: Music / Studio

Suno is introducing Suno Bridge, a standalone app that brings its music models into a digital audio workstation, starting with Ableton Live 12 on Mac. The app came out of conversations with artists, producers and songwriters, and it lets them generate new instrument and vocal parts without leaving their session. Bridge runs alongside Ableton Live and both apps must be open; prompts entered in Bridge create parts that fit the tempo, key, style, harmonic content and groove of the session. Syncing uploads session files to Suno so Bridge understands the musical context of the track, and loop start/end markers in Ableton's Arrangement View set the prompting window, so a loop point from bar 9-16 gives an 8-bar loop starting on the first beat of bar 9. Suno says the generated sounds follow chord progressions, groove and key, unlike stock samples. Bridge is launching in beta and is included with all Premier subscriptions, with more operating systems and DAW platforms planned.

### Harvey — The Brief: November 2026

- Company: Harvey (harvey.ai)
- Announced: 2026-11-13T14:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.harvey.ai/blog/the-brief-november-2026
- Record: https://forck.live/items/35873-the-brief-november-2026
- Subject: Harvey

Harvey's November 2026 release gives its agents the context of each matter or project and memory of how a user works, so routine tasks can be handed off and kept in one Space. Harvey II draws on matter and project context and remembers working preferences, while Custom Agents and Agent Builder turn a team's processes and standards into reusable agents that follow the same approach each run. Review tables and Vaults now keep pace with incoming files: a review table can be connected to a Vault or selected folders, show when new documents land, and add them in one click, and multiple folders from iManage, SharePoint and Google Drive can sync into a single Vault. Harvey also added search across Microsoft Teams channels, chats and shared files, and an Ansarada integration that imports files from data rooms into threads and Vaults or syncs a Vault to an Ansarada folder. Excel uploads are now supported up to 200 MB, and Harvey can draft emails with attachments directly into an Outlook Drafts folder from a thread, the Word add-in or a custom agent. Governance was simplified so new agents and workflows can be used as soon as they are saved, with admin approval applied only at the sharing levels an organization chooses.

### NVIDIA — AI Is Helping Care Teams Get Patients the Right Care, Sooner

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-13T13:00:42+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://blogs.nvidia.com/blog/how-ai-is-helping-care-teams/
- Record: https://forck.live/items/35847-ai-is-helping-care-teams-get-patients-the-right-care-sooner
- Subject: AI platform

NVIDIA is highlighting how Dyania Health's Synapsis AI, a clinical reasoning engine, reads full patient records to flag which patients need attention and why. Hospitals hold decades of patient history, most of it free text such as physician notes, operative reports, imaging and pathology reads and discharge summaries, so understanding one patient's story can mean reading years of charts. Synapsis AI works across every note, report and lab result, connecting findings years apart and returning clinical signals traceable to their source in the record. It runs on NVIDIA-accelerated computing inside a health system's own firewall, so records are processed onsite and patient data never leaves the hospital. Dyania builds and trains its clinical models with NVIDIA NeMo open libraries, then uses NVIDIA TensorRT and TensorRT-LLM to optimize them on the hospital's own GPU servers. Cleveland Clinic used Synapsis AI in a study of more than 7,800 aortic valve replacement patients treated between 2005 and 2021, building a predictive model of valve deterioration and reintervention risk at three, five and ten years. The company says the platform completed chart review across the full cohort that would have taken clinicians years by hand, and that the tool remains assistive: clinicians must confirm and act on the evidence it surfaces.

### ElevenLabs — Introducing Forward Deployed Creatives in ElevenCreative

- Company: ElevenLabs (elevenlabs.io)
- Announced: 2026-11-13T12:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://elevenlabs.io/blog/introducing-forward-deployed-creatives
- Record: https://forck.live/items/36064-introducing-forward-deployed-creatives-in-elevencreative
- Subject: Voice AI

ElevenLabs is introducing Forward Deployed Creatives, creative professionals who work inside a customer's team in ElevenCreative or produce finished creative end to end. The company says teams stall when moving AI creative from pilot to regular use because of three problems: brand consistency, fragmented tools, and creative control. Forward Deployed Creatives are performance creative producers, brand creative producers, and creative strategists who work with brands and agencies, learning the brand and creative strategy before working on real campaigns. Enterprises can partner in three ways: enabling their own creative team on the ElevenCreative platform, building reusable workflows together with Forward Deployed Engineers, or having Forward Deployed Creatives deliver end-to-end creative from strategy to finished, localized ads. For localized content, trained linguists and audio professionals review the work so it reads and sounds like it was made for each market. ElevenLabs says the same team has helped hundreds of companies launch AI agents with ElevenAgents, and it also offers end-to-end production services for dubbing, voiceovers, subtitles, audio description, and custom brand voices.

### ElevenLabs — Oxford spent years studying how women build companies. Now this research can speak.

- Company: ElevenLabs (elevenlabs.io)
- Announced: 2026-11-13T12:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://elevenlabs.io/blog/impact-sofia-oxford-said
- Record: https://forck.live/items/35971-oxford-spent-years-studying-how-women-build-companies-now-this-research-can
- Subject: Voice AI

Oxford Saïd Business School's Inspiring Female Founders initiative has turned its founder interviews into a voice agent called SOFIA for Business, built on ElevenAgents and supported by the ElevenLabs Impact Program. The initiative was created because women founders remain underfunded and under-seen: female-only teams receive about 2.3% of venture capital, and successful women founders get far less visibility than their male peers, leaving aspiring founders with few role models. SOFIA has two jobs: founders can talk to it to learn from the experience of the women interviewed before them, and it also interviews founders itself, with each consented conversation joining the research so the knowledge grows with every new founder who takes part. SOFIA for Business launched earlier today at the Oxford Saïd Entrepreneurship Forum, where attendees talked with it at the ElevenLabs booth and added their own experiences to the research. It was co-founded by Pinar Ozcan, professor of entrepreneurship and innovation at Oxford Saïd and founder of Inspiring Female Founders, alongside Amanda Slavin and Arabella Nyffenegger. ElevenLabs supports the project through its Impact Program × Professors, which gives university educators like Ozcan free access to ElevenLabs tools.

### Runway — Seedance 2.5 Flash is now available on Runway.

- Company: Runway (runway.com)
- Announced: 2026-11-13T01:23:06+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://x.com/runwayml/status/2121051193921118712
- Record: https://forck.live/items/35656-seedance-2-5-flash-is-now-available-on-runway

Runway has added Seedance 2.5 Flash to its platform, giving users a faster and cheaper way to generate videos. The company says the model is now available on Runway, and points users to a link to get started. Runway did not state the price, speed figures or which plans include the model.

### Anthropic — Building always-on agents

- Company: Anthropic (anthropic.com)
- Announced: 2026-11-13
- Category: not stated
- Coverage: 1 outlet
- Announcement: no
- Group: covered
- Source: https://claude.com/resources/articles/building-always-on-agents
- Record: https://forck.live/items/36037-building-always-on-agents
- Subject: Claude

Anthropic's guide explains the architecture behind always-on agents, which stay stateful across sessions and act without direct supervision, and says recent model advances have made them practical to run. The company attributes the resurgence to better judgment in the latest models, which it says are more reliable at reading user intent and knowing when to check in, and to falling intelligence costs. On one long-horizon coding benchmark, Anthropic says Opus 5.5 cost 78% less per solved task than Opus 5, against 40% less on typical workloads. The guide describes five primitives: an event inbox, timers, memory, permissions and an outbox, and notes customers have asked how to add the experience to their own products. It walks through the run cycle, in which an agent sleeps, wakes on an event, works inside a session and saves what happened before sleeping again. The inbox is recommended as the source of truth with one active worker at a time, while timers live in a separate store and are fired by a scheduler that polls for due follow-ups and schedules.

### Ai2 — XYBench: Helping with the problem behind the question

- Company: Ai2 (allenai.org)
- Announced: 2026-11-13
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://allenai.org/blog/xybench
- Record: https://forck.live/items/35973-xybench-helping-with-the-problem-behind-the-question
- Subject: Olmo / Molmo / Asta

Ai2 released XYBench, a benchmark of 8,115 real-world questions that tests whether AI assistants recognize when a user's request rests on a misconception. The benchmark targets the XY problem, where someone wants to accomplish X but asks for help with an assumed approach Y, such as asking how to thicken frosting that slides off cupcakes that are simply too warm. Ai2 developed XYBench with researchers at Carnegie Mellon, NVIDIA, and Johns Hopkins, drawing questions from Stack Overflow, Stack Exchange, curated examples, and WikiHow guides. XYBench measures three things: whether a response solves the stated and intended requests, which one it emphasizes, and whether it steers the user toward the underlying goal. Claude Sonnet 4.5 serves as the automated judge, and two human reviewers checked its judgments on 500 responses, with agreement ranging from 97.5–100% on whether a solution appeared to 74.2–79.8% on pragmatic emphasis. Ai2 says all five evaluated models were more likely to include a solution to the stated request than to the underlying goal, and on Stack Overflow they focused primarily on the underlying goal in only 35–41% of responses.

### Harvey — Avianca Deploys Harvey Across Its Legal Organization

- Company: Harvey (harvey.ai)
- Announced: 2026-11-13
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.harvey.ai/blog/avianca-deploys-harvey-across-its-legal-organization
- Record: https://forck.live/items/35936-avianca-deploys-harvey-across-its-legal-organization
- Subject: Harvey

Avianca is deploying Harvey across its legal organization to support aviation, regulatory and commercial work spanning Colombia, El Salvador and Brazil. The airline's lawyers work in Spanish, Portuguese and English, and the team supports regional businesses including Cargo, Deprisa and LifeMiles. Through Harvey's Vault, lawyers can search and build on their own aircraft lease and maintenance clause libraries, regulatory filings, framework contracts and M&A documents. Harvey's Word add-in lets them draft and review documents and compare clauses against Avianca's standards inside Word. During the pilot, Harvey ran tailored training across Avianca's 14 legal management units and worked with the team through security and AI governance reviews. Richard Galindo, Avianca's Chief Legal Officer, said the adoption strengthens legal management, compliance and organizational efficiency.

### DeepSeek — DeepSeek Full-Suite App API Release 2026/11/13

- Company: DeepSeek (deepseek.com)
- Announced: 2026-11-13
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://api-docs.deepseek.com/news/news261113
- Record: https://forck.live/items/35899-deepseek-full-suite-app-api-release-2026-11-13
- Subject: DeepSeek models

The full DeepSeek App experience, now via API Search, files, links, safety, and voice. The capabilities behind the DeepSeek App are now available through the DeepSeek API.

### Cohere — Cohere Expands Its Sovereign AI Platform with AISLE and cubic

- Company: Cohere (cohere.com)
- Announced: 2026-11-13
- Category: partnership-acquisition
- Coverage: 1 outlet
- Announcement: yes
- Group: covered
- Source: https://cohere.com/blog/cohere-expands-its-sovereign-ai-platform-with-aisle-and-cubic
- Record: https://forck.live/items/35884-cohere-expands-its-sovereign-ai-platform-with-aisle-and-cubic
- Subject: Command

Cohere has completed its acquisition of AISLE, an autonomous vulnerability discovery and remediation company, and brought in the team behind cubic, an AI code review platform, to expand its sovereign AI platform. The additions give Cohere cyber defense and secure software development capabilities for enterprises and governments that need AI inside private clouds, on-premises and air-gapped environments. AISLE's co-founders take leadership roles: Ondrej Vlcek becomes Senior Vice President, Cohere Cyber, leading the newly established cybersecurity business; Jaya Baloo becomes Chief Information Security Officer; and Stanislav Fort becomes Vice President, AI Security, joining core model research and training. Cohere says it plans a substantial long-term investment in Prague, expanding AISLE's existing presence and making the city its global center for AI cybersecurity and autonomous security systems. The cubic team, led by founder and CEO Paul Sanglé-Ferrière, will establish Cohere's secure coding vertical, with review agents that Cohere says have ranked among the strongest performers on code review benchmarks. Cohere says these are its fourth strategic M&A transactions this year, following Aleph Alpha and Reliant AI, extending its capabilities across frontier AI, biopharma and healthcare, cybersecurity, and software development.

### OpenClaw — OpenClaw is coming to Shenzhen

- Company: OpenClaw (openclaw.ai)
- Announced: 2026-11-13
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openclaw.ai/blog/openclaw-shenzhen-claw-hack-2026/
- Record: https://forck.live/items/35841-openclaw-is-coming-to-shenzhen
- Subject: OpenClaw

OpenClaw is bringing a three-day claws festival to Shenzhen on December 4–6, 2026, combining its Claw Hack hackathon with ClawCon, its flagship conference. The company says its Shanghai event drew thousands, and it is returning to China to partner locally and internationally around open hardware, open source and open-weight models. Claw Hack runs for two days, with core team members on hand to pair-program, review PRs and troubleshoot. There are two tracks: the software track covers Skills, Plugins, integrations, local knowledge tools and new interfaces, while the hardware track brings agents into the physical world through devices and sensors. That hardware list includes M5Stack development hardware, Seeed Studio’s open-source reBot Arm B601, Pamir’s Lapis One and the Autonomous Lamp, and OpenClaw notes some hardware is unreleased. Applications for Claw Hack close on November 25, and ClawCon takes place on Sunday afternoon, December 6, with separate registration for each event.

### OpenAI — Cyber at the Frontier: Introducing Daybreak Red Team & Daybreak Defense Factory

- Company: OpenAI (openai.com)
- Announced: 2026-11-12T22:37:14+00:00
- Category: product-launch
- Coverage: 4 outlets
- Announcement: yes
- Group: covered
- Source: https://www.youtube.com/watch?v=YmxpxJZvm4M
- Record: https://forck.live/items/35606-cyber-at-the-frontier-introducing-daybreak-red-team-daybreak-defense-factory

OpenAI introduced two cybersecurity products under a unified Codex plugin called Daybreak, aimed at teams that need to find and fix vulnerabilities at scale. Daybreak Defense Factory builds on Codex Security Cloud to automate the defensive loop of finding, validating and fixing vulnerabilities, and gives engineers proposed patches with test results so they can review changes before merging. All of that work runs in isolated cloud development environments. Daybreak Red Team, in early access, uses fleets of agents to test a running application within limits a team sets, finding new vulnerabilities and checking whether an attacker could exploit issues already flagged. Engineers can send confirmed findings to Daybreak Defense Factory for help fixing them, then use Red Team to check that the fixes work.

### Amazon — Deploy frontier AI models on AWS Trainium with SGLang-Neuron

- Company: Amazon (amazon.com)
- Announced: 2026-11-12T22:03:34+00:00
- Category: developer-tool-release
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/deploy-frontier-ai-models-on-aws-trainium-with-sglang-neuron/
- Record: https://forck.live/items/35588-deploy-frontier-ai-models-on-aws-trainium-with-sglang-neuron
- Subject: Bedrock / Nova

Amazon says SGLang-Neuron is publicly available through an AWS Deep Learning Container, letting teams serve models and generate reinforcement learning rollouts on Trainium with the SGLang interface they already use. The container was released by RadixArk in collaboration with Annapurna Labs, and the plugin is out-of-tree, so Trainium support and SGLang core develop independently and iterate faster. Installing the plugin registers the neuron platform through SGLang's plugin entry points, and setting SGLANG_PLATFORM=neuron selects it when the server starts. SGLang core keeps the API server, request lifecycle, RadixAttention and prefix-cache bookkeeping, while the plugin implements Trainium-specific models, attention, mixture of experts and the graph runner. TorchNeuron registers Trainium as a native PyTorch device, and the plugin compiles standard PyTorch operations and NKI kernels with torch.compile for the graph runner to execute. It also supports PD disaggregation, where a prefill server computes the prompt KV cache and transfers it to a decode server over Elastic Fabric Adapter using NIXL. The walkthrough uses a Trainium3 EC2 instance with a pytorch-inference-sglang-neuronx container to serve openai/gpt-oss-120b.

### Amazon — How Natera designed an agentic AI system for lab ops with Amazon Bedrock AgentCore

- Company: Amazon (amazon.com)
- Announced: 2026-11-12T21:47:07+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/how-natera-designed-an-agentic-ai-system-for-lab-ops-with-amazon-bedrock-agentcore/
- Record: https://forck.live/items/35570-how-natera-designed-an-agentic-ai-system-for-lab-ops-with-amazon-bedrock
- Subject: Bedrock / Nova

Natera's lab engineering team worked with AWS to build SWARM, an agentic AI system on Amazon Bedrock AgentCore that captures and deploys lab operations knowledge at scale. The system answers instrument support tickets that previously waited an average of 17.4 hours for a first response, with more than 65 tickets a day arriving across three Jira queues. Natera says SWARM delivers about 130x faster first response time and 96.2 percent automated coverage across support tickets. The design pairs two runtimes on one skill library: swarm-agent writes structured diagnostic reports before an engineer opens a ticket, while swarm-hive lets engineers interact in real time, upload images and explore alternative troubleshooting paths. "The system does not replace engineers—it gives them access to the full depth of institutional knowledge the moment a ticket arrives," said Ariel Jirau, who leads Natera's lab engineering team. The team chose a keyless deployment model for the clinical environment, using OpenID Connect and AWS Security Token Service for temporary per-deployment credentials, IAM least privilege, private VPC subnets and Secrets Manager rotation on a 30-day cycle.

### Microsoft Research — Microsoft Research Focus – Edition #98

- Company: Microsoft Research (microsoft.com)
- Announced: 2026-11-12T21:31:57+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.microsoft.com/en-us/research/podcast/microsoft-research-focus-edition-98/
- Record: https://forck.live/items/35557-microsoft-research-focus-edition-98
- Subject: Research / Phi

Microsoft Research's newsletter roundup covers a proof resolving the 31-year-old KLS conjecture in high-dimensional geometry, which the company says removes a major theoretical obstacle to understanding randomized algorithms. A study of 13.5 million coding sessions found fewer than 8% needed context compaction, yet those sessions accounted for 44% of total AI processing costs. To address that, researchers built COMPACTIONBENCH, a benchmark for evaluating context compaction in long-running coding tasks, and APEC, a method that prepares summaries in the background instead of interrupting an agent's work. In evaluations, APEC improved coding performance while cutting costs by up to 25.4% and execution time by 9.6%. Separately, researchers introduced the Latent Core Tokenizer, which identifies reusable components of words before building a shared vocabulary across languages; across 104 languages it produced more linguistically meaningful tokens than leading alternatives while keeping comparable efficiency, and improved model performance on four multilingual benchmarks. The edition also features a conversation on AI speech translation and a Tokyo research intern's move from computer vision to physical intelligence.

### Amazon — Four Amazon Bedrock inference choices that affect cost

- Company: Amazon (amazon.com)
- Announced: 2026-11-12T20:51:55+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/four-amazon-bedrock-inference-choices-that-affect-cost/
- Record: https://forck.live/items/35548-four-amazon-bedrock-inference-choices-that-affect-cost
- Subject: Bedrock / Nova

Amazon's guide explains four request-level Amazon Bedrock settings that change inference cost without changing the model or prompt: cache duration, service tier, reasoning effort, and maxTokens. Every Bedrock response carries usage metadata, and input and output tokens bill at the rate for the model and tier that served the request, with cache reads and writes on separate counters and reasoning tokens counted as output. For Anthropic Claude Sonnet 5 and Anthropic Claude Opus 5 on Bedrock, a 5-minute cache write costs 1.25 times the input rate, a 1-hour write costs 2 times, and a cache read costs 0.1 times, which sets the break-even between the two durations. A 5-minute cache pays back on the first reuse within 5 minutes, while the 1-hour cache is cheaper from the second request within the hour when the same prefix recurs more than 5 minutes apart. Flex is a service tier for latency-tolerant work such as background jobs and nightly summarization; the Amazon Nova 2 Lite and OpenAI gpt-oss-120b model cards list it as supported, while the cards for Claude Sonnet 5 and Claude Opus 5 list it as not supported. A Flex request to an unsupported model fails with a ValidationException rather than silently running on Standard, and Amazon says applications should catch that error and retry once with default rather than retrying every ValidationException.

### Amazon — Tokenomics at scale: Track and cap Amazon Bedrock spend per IAM principal with CUR 2.0 and AWS Budgets

- Company: Amazon (amazon.com)
- Announced: 2026-11-12T19:29:55+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/tokenomics-at-scale-track-and-cap-amazon-bedrock-spend-per-iam-principal-with-cur-2-0-and-aws-budgets/
- Record: https://forck.live/items/35502-tokenomics-at-scale-track-and-cap-amazon-bedrock-spend-per-iam-principal-with
- Subject: Bedrock / Nova

Amazon's guide shows how to attribute and cap Amazon Bedrock generative AI spend per IAM principal using only AWS-native features, with no application code for the common case. Amazon Bedrock records the IAM user or role behind every inference request automatically, so no change to model inference calls is needed. Organizations activate those tags as cost allocation tags, then read the data in Cost Explorer or a CUR 2.0 export for line-item analysis. Enforcement comes from an AWS Budget scoped to Amazon Bedrock and filtered to one identity or tag value, with a budget action that applies a Deny IAM policy when the threshold is crossed. Where several people share one role, principal tags cannot tell them apart, so session tags passed at role assumption time are used instead; up to 50 principal or session tags are allowed per IAM user or role. A tag key appears in the cost allocation list only after the principal has made at least one Amazon Bedrock call.

### Hugging Face — LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B

- Company: Hugging Face (huggingface.co)
- Announced: 2026-11-12T19:07:35+00:00
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://huggingface.co/blog/ibm-granite/llmon-an-llm-native-markup-language-now-in-granite
- Record: https://forck.live/items/35483-llmon-an-llm-native-markup-language-now-in-granite-4-1-3b-and-8b
- Subject: Platform

IBM released two experimental preview models derived from Granite 4.1 base models, trained to consume and generate LLMON, an LLM-native markup language that marks context components with meaningful tags. The company says LLMON exists because context structure is usually implicit or encoded in formats such as JSON that were not built to be digested by LLMs, which it says makes smaller models brittle and open to prompt injection. IBM says indirect prompt injections hidden in user data are mostly ignored by the LLMON models because instructions and data are explicitly separated, and that on some benchmarks attack success rates drop to nearly zero while regular instruction-tuned models get hijacked with high probability. IBM says this holds while general capability is maintained relative to Granite 4.1 Instruct models. LLMON does not replace chat templates but enriches them: the preview models ship with their own chat template, so a turn can name its parts and mark each one exec: True or False, with True parts bound into an exec block stating which instruction runs and what it operates on. IBM also released a GitHub repo with a toolkit, workflows and templates, plus converters and notebooks that translate JSON to LLMON and back losslessly, and Mellea has native LLMON support enabled by automatically LLMONizing its LLM calls.

### Runway — Closing the Gap Between Policy and Practice

- Company: Runway (runway.com)
- Announced: 2026-11-12T18:50:00+00:00
- Category: safety-policy-update
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://runway.com/news/safety/closing-the-gap
- Record: https://forck.live/items/35481-closing-the-gap-between-policy-and-practice
- Subject: Gen / Aleph

Runway says its moderation now allows far more conflict, danger and violence in creative work while still blocking content that causes real harm. Two years ago the company relied on third-party classifiers that scored content in fixed categories such as blood or weapons, and it could only pick categories and block thresholds. Those tools could not tell a fantasy battle from real violence, so Runway blocked more than it wanted, including a small drop of blood, a minor scratch, or ketchup mistaken for blood. Advances in large language models let Runway write its own policies and apply them with more context, so it blocks less allowed content and catches more that should be blocked. The company says it can now draw a line between creative expression and harmful content, allowing sword fights, shootouts, battle scenes and horror while blocking violent content aimed at protected groups or recreations of extremist attacks. Runway also says it more accurately catches sexualized content involving minors and non-consensual intimate imagery. Limitations remain: third-party models on Runway apply their own moderation, LLMs still over-enforce on ambiguous words or miss harmful content framed as fiction, and some legitimate uses are blocked when context is missing.

### Amazon — Optimizing cost per task and proving agent ROI on Amazon Bedrock

- Company: Amazon (amazon.com)
- Announced: 2026-11-12T18:43:19+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/optimizing-cost-per-task-and-proving-agent-roi-on-amazon-bedrock/
- Record: https://forck.live/items/35471-optimizing-cost-per-task-and-proving-agent-roi-on-amazon-bedrock
- Subject: Bedrock / Nova

Amazon's guide explains how to measure the true cost of running agentic workflows on Amazon Bedrock, arguing that cost per task matters more than cost per token. It names four hidden cost drivers the pricing page does not show: tokenizer efficiency, output alignment and verbosity, agentic loop efficiency, and reasoning overhead. The post walks through a worked example comparing two models on an email-extraction workflow, using illustrative prices, and shows how a higher-priced model can end up cheaper once token fragmentation, padded output and correction turns are counted. It also gives formulas for cost per successful task, covering retry-until-success and escalate-on-failure workflows, and notes that prompt caching splits input into uncached, cache-read and cache-write buckets. The guide is written for application and platform engineers who own the optimization levers and for product owners and FinOps partners who need to justify agent spend. It points to Amazon Bedrock model evaluation for the quality half of benchmarking, where teams define criteria and upload prompt datasets to get scored results.

### Amazon — Building a self-service LLM gateway with Amazon Bedrock and LiteLLM

- Company: Amazon (amazon.com)
- Announced: 2026-11-12T18:41:12+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/building-a-self-service-llm-gateway-with-amazon-bedrock-and-litellm/
- Record: https://forck.live/items/35472-building-a-self-service-llm-gateway-with-amazon-bedrock-and-litellm
- Subject: Bedrock / Nova

Amazon's guide shows platform teams how to build a self-service LLM gateway that puts Amazon Bedrock behind an OpenAI-compatible API using LiteLLM on AWS Fargate. The pattern answers three pressures: platform teams need latency and error monitoring, finance teams need inference costs attributed to teams and projects, and security teams need consistent access controls. Developers get a single base URL and a per-user API key, so tools such as Claude Code, Continue and Cursor can swap models, add fallbacks or change providers without touching developer configurations. Each model, including Claude Sonnet and Claude Haiku, gets its own Amazon Bedrock Application Inference Profile for per-model quota isolation, while Amazon RDS (PostgreSQL) holds key metadata and spend state and Amazon ElastiCache for Valkey handles caching and rate limiting. LiteLLM's AI Gateway is open source and free for up to five users, but the self-service key issuance described requires the licensed Enterprise version, which adds single sign-on, audit logs and delegated admin roles. The post notes that Amazon Bedrock Guardrails are not covered and that responsible AI controls would normally be added before production.

### Google DeepMind — Introducing Gemini 4 Hydrogen

- Company: Google DeepMind (deepmind.google)
- Announced: 2026-11-12T18:00:50+00:00
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://deepmind.google/blog/introducing-gemini-4-hydrogen/
- Record: https://forck.live/items/35442-introducing-gemini-4-hydrogen
- Subject: Gemini

Google DeepMind introduced Gemini 4 Hydrogen, a fast model for everyday coding and agentic workflows, and brought Gemini 4 to all Google AI Studio tiers. Hydrogen is built for rapid execution rather than the deep, long-horizon reasoning of the frontier Carbon and Argon models, and it can also serve as the high-efficiency execution engine alongside them in multi-model workflows. Google says Hydrogen surpasses Argon on the Artificial Analysis Intelligence Index at its default high thinking level, 51 versus 50, and reaches 54 at its highest xhigh thinking setting. The model is priced flat across its entire 1 million-token context window at $1.50 per million input tokens and $7.50 per million output tokens. For teams migrating from Gemini 3.8 Flash, Google says Hydrogen finishes tasks 2.5 times faster at its high default setting and more than triples Terminal-Bench 4.1 scores, 57% versus 18.2%, while cutting fabricated claims by more than 5x. Google also introduced Orchestrator in Google Antigravity, which switches between models like Carbon and Hydrogen within a session; on Terminal-Bench 4.0 it reached near-Carbon task success rates while consuming 24% less quota and completing 19% faster.

### Amazon — Run agents securely and efficiently with the next generation AgentCore runtime

- Company: Amazon (amazon.com)
- Announced: 2026-11-12T17:59:59+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/run-agents-securely-and-efficiently-with-the-next-generation-agentcore-runtime/
- Record: https://forck.live/items/35460-run-agents-securely-and-efficiently-with-the-next-generation-agentcore-runtime
- Subject: Bedrock / Nova

Amazon is updating the AgentCore runtime, the managed compute layer of Amazon Bedrock AgentCore, so that agent sessions start faster, scale further and cost less to run. The company says sessions now start up to 15x faster, scale to 4x more compute and run up to 68% more efficiently, while every session still runs in its own hardware-isolated Firecracker microVM with its own identity, policy and audit trail. Sessions scale to up to 8 vCPUs and 32 GB, with container images up to 10 GB, and start in under 2 seconds regardless of image size or concurrency, which Amazon says removes the need for warm pools and pre-provisioned clusters. Coinbase's Siwei Shen says cold starts fell from nearly a minute to under five seconds at peak traffic, and Mission Cloud reports 35% faster cold starts for inference requests and 77% faster health checks in an initial comparison of both runtime versions. Zuora's Brad Ruppert says the new runtime cut memory footprint per invocation by around 60% in many use cases, and Rescale says it uses the runtime to deploy and scale engineering agents without operating the underlying infrastructure.

### Anthropic — Building the Claude API for the Frontier

- Company: Anthropic (anthropic.com)
- Announced: 2026-11-12T17:00:11+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.youtube.com/watch?v=p2xvJAbF-i4
- Record: https://forck.live/items/35417-building-the-claude-api-for-the-frontier

Every Claude launch puts the spotlight on the model, but just as much work goes into making sure the API evolves with it. That means adding new controls, retiring outdated ones, and helping developers rethink how they build with each new release.

### Replit — Replit Agent can now buy for your business and help you reach customers

- Company: Replit (replit.com)
- Announced: 2026-11-12T16:45:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://replit.com/blog/replit-agent-ads-commerce
- Record: https://forck.live/items/35393-replit-agent-can-now-buy-for-your-business-and-help-you-reach-customers
- Subject: Replit Agent

Replit is giving Replit Agent two new jobs for business owners: buying supplies through Link and Visa, and running Google Ads campaigns. The company says the additions answer a sprawl of work that follows a launch, such as getting found and reaching customers, which it calls a barrier even for engineers who can build a product. For commerce, a browser subagent navigates a merchant's site, prepares checkout and returns the details to Agent, which waits for the owner to approve the final total before anything is charged. Replit says the browser session is isolated, shell access and general browsing are disabled during payment, and approval checks are enforced in code rather than left to the model. Agentic Commerce is available for Core and Pro users, starting in the US and limited to one-time purchases. Ads with Google, which covers Search, YouTube, Display and Discover, is available for Pro users worldwide starting today. Replit says it has seen a ninefold increase since May in entrepreneurs connecting Replit-built businesses to Stripe for monetization.

### Krea — introducing Krea Bots, your new creative crew.

- Company: Krea (krea.ai)
- Announced: 2026-11-12T16:30:01+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://x.com/krea_ai/status/2120917039435198863
- Record: https://forck.live/items/35368-introducing-krea-bots-your-new-creative-crew

Krea introduced Krea Bots, a beta feature that it says can take on entire creative projects and adapt to a user's taste as they and their team collaborate with them. Krea says all of its recent creative work, including a video shared with the announcement, was produced with these bots. People who join the beta receive $1,000 in credits.

### Amazon — Migrate to Amazon Bedrock Managed Knowledge Base with your coding agent

- Company: Amazon (amazon.com)
- Announced: 2026-11-12T16:16:45+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/migrate-to-amazon-bedrock-managed-knowledge-base-with-your-coding-agent/
- Record: https://forck.live/items/35378-migrate-to-amazon-bedrock-managed-knowledge-base-with-your-coding-agent
- Subject: Bedrock / Nova

Amazon added a knowledge base migration guide to the amazon-bedrock skill in its Agent Toolkit for AWS, aimed at teams moving from customer-managed knowledge bases on Amazon Bedrock to the Managed Knowledge Base. With a customer-managed knowledge base, teams provision and scale the vector store, tune the embedding model and index settings, and plan capacity for each new data source, work that Managed Knowledge Base takes over. The guide gives a coding agent the context to plan the migration, starting with read-only discovery and a plan the team reviews line by line. On approval, the agent creates the new knowledge base and its data sources alongside the original, or writes a migrated infrastructure as code template for the pipeline. Ingestion and updating the application's Retrieve calls stay with the team, and the agent can walk them through both. AWS says Managed Knowledge Base balances ease of use, retrieval accuracy and cost, and that agentic retrieval, AgentCore Gateway integration and resource-based policies are not available on customer-managed knowledge bases.

### Harvey — Close Contracts Faster With Every Version in One Place

- Company: Harvey (harvey.ai)
- Announced: 2026-11-12T16:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.harvey.ai/blog/close-contracts-faster-with-every-version-in-one-place
- Record: https://forck.live/items/35350-close-contracts-faster-with-every-version-in-one-place
- Subject: Harvey

Harvey updated Contract Intelligence so the full history of a negotiation stays with the contract, letting legal teams assess risk in context and move deals to signature faster. The company says the latest redline rarely tells the whole story, because it shows what changed but not why earlier terms were accepted or what Legal conceded, leaving that context scattered across email threads and earlier drafts. Contract Intelligence matches a new version to the existing contract and adds it to the overview alongside prior rounds and related agreements, so Harvey reviews the new round accounting for positions Legal already accepted. Contract statuses show which agreements are in review, waiting on the counterparty, or signed, giving Legal and Legal Ops one place to check when the business asks for an update or a deadline approaches. Signed agreements go into a repository where accepted terms are extracted, and Contract Review Agents can draw on those signed positions to inform redline suggestions on later vendor agreements. Teams can also ask Harvey direct questions about their agreements, such as which vendors pushed back on data security terms and what was accepted. Harvey says early access is available through account managers, or by requesting a demo.

### Amazon — Build composable, interoperable multi-agent systems with A2A on AWS

- Company: Amazon (amazon.com)
- Announced: 2026-11-12T15:42:21+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/build-composable-interoperable-multi-agent-systems-with-a2a-on-aws/
- Record: https://forck.live/items/35325-build-composable-interoperable-multi-agent-systems-with-a2a-on-aws
- Subject: Bedrock / Nova

AWS published a reference architecture for connecting agents built on different frameworks across organizational boundaries using the Agent-to-Agent (A2A) protocol, with each side in its own AWS account. The walkthrough pairs a LangChain-based enterprise orchestrator with a Strands Agents-based compliance SaaS agent, and neither side depends on the other's framework or SDK. During onboarding the two parties exchange only public values such as identity provider URLs, client IDs, scopes and endpoint URLs, and each side then trusts tokens from the other's Amazon Cognito user pool, so no secrets are shared. The orchestrator fetches the SaaS agent's Agent Card at startup and creates a LangChain tool for each skill on it, so the model picks the tool and no hardcoded routing is needed. The enterprise orchestrator uses Anthropic Claude Sonnet 4.6 on Amazon Bedrock, while the internal context agent uses Amazon Nova Lite and the official a2a-sdk pinned to 1.2.2. The system deploys from three Jupyter notebooks and is tested end to end, with the enterprise agents running in private subnets without public IP addresses.

### Amazon — Fine-grained access control for Amazon Bedrock Managed Knowledge Base tools

- Company: Amazon (amazon.com)
- Announced: 2026-11-12T15:40:24+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/fine-grained-access-control-for-amazon-bedrock-managed-knowledge-base-tools/
- Record: https://forck.live/items/35326-fine-grained-access-control-for-amazon-bedrock-managed-knowledge-base-tools
- Subject: Bedrock / Nova

Amazon's guide shows how to keep an agent's document searches scoped to the signed-in user when a shared knowledge base serves many employees or tenants. The problem is that an LLM picks the tool arguments at runtime, so a search tool that takes "who to search as" lets a prompt injection or model error make the agent search as someone else. The design puts verified identity in trusted code instead: a request interceptor on Amazon Bedrock AgentCore Gateway resolves the email from the validated JWT and overwrites the userContext argument of tools/call. A Cedar policy in AgentCore then denies the request if userContext.userId does not match the token's email claim, and the knowledge base's resource-based policy denies searches from principals other than the gateway's execution role. Managed Knowledge Base's ACL awareness returns only chunks the userContext user may view, though the post says it filters results and "isn't an authorization mechanism by itself". The sample stores department A and B business plans plus a shared notice in one knowledge base, with two users searching through the same agent.

### Amazon — From prototype to production: Engineering agents with Agent DLC

- Company: Amazon (amazon.com)
- Announced: 2026-11-12T15:38:47+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/from-prototype-to-production-engineering-agents-with-agent-dlc/
- Record: https://forck.live/items/35327-from-prototype-to-production-engineering-agents-with-agent-dlc
- Subject: Bedrock / Nova

Amazon's guide introduces the Agent Development Lifecycle (Agent DLC), a lifecycle with two loops and one explicit standard for judging every change to an AI agent before release. The post argues that conventional unit tests and CI/CD are necessary but not sufficient for agents, because agents are non-deterministic, because prompt, tool, retrieval or memory changes alter behavior in ways text review cannot reveal, and because model endpoints, knowledge bases and external APIs can change independently of the repository. It presents a measured reference implementation, an internal HR policy assistant answering questions over a document knowledge base, and states the results are not a customer outcome or a benchmark result. In the first evaluation round, chunk-level source precision was 1.00 for all three golden questions, while fact-level precision was 0.59, 0.53 and 0.09, so a release gate using chunk-level precision alone would have accepted all three answers. After a prompt-only change that added three constraints, the first question improved substantially and the third got worse, which the evaluation traced to a knowledge-coverage problem: the corpus lacked a dedicated sick-leave policy document. The accompanying AWS Samples repository includes infrastructure deployment, the HR agent, custom evaluators, and scripts for baseline evaluation and prompt-change re-evaluation, and the post notes scores can vary across runs and repository versions.

### Harvey — How Legal Engineers Would Use Harvey Agents for Outlook

- Company: Harvey (harvey.ai)
- Announced: 2026-11-12T14:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.harvey.ai/blog/harvey-agents-for-outlook-use-cases
- Record: https://forck.live/items/35228-how-legal-engineers-would-use-harvey-agents-for-outlook
- Subject: Harvey

Harvey's legal engineers, all former practicing lawyers, are describing how Harvey Agents for Outlook could handle inbox and calendar work for lawyers who receive hundreds of emails a day. The agents can work across a single email or an entire inbox, preparing a daily briefing that summarizes meetings, flags emails needing a response and outstanding follow-ups, and recommends calendar changes. They can also draft emails using firm knowledge, create tasks, and file correspondence and attachments into Vault, with lawyers reviewing and approving each action. With the Outlook connector enabled, an agent can search across multiple messages, a specific Outlook folder or subfolder, and Vaults for documents or answers about a deal. The post walks through scenarios from the engineers' own practice, including arbitration casework, in-house litigation, M&A deals and cross-border disputes, rather than announcing a new release.

### Vercel — Workflow SDK 5: built for production

- Company: Vercel (vercel.com)
- Announced: 2026-11-12T13:00:00+00:00
- Category: developer-tool-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://vercel.com/blog/workflow-sdk-5
- Record: https://forck.live/items/35940-workflow-sdk-5-built-for-production
- Subject: v0 / AI SDK / AI Gateway

Vercel released Workflow SDK 5, adding end-to-end payload encryption and per-region execution to the durable workflow tool, after customers ran more than 4 billion steps and 700 million runs on it. The SDK encrypts workflow and step inputs and outputs, hook payloads and stream data with AES-256-GCM, using a unique key for each run, so any backend including Vercel only stores ciphertext. On Vercel the encryption needs no setup or code changes, and each run's key is derived from a deployment-scoped key so a run can only decrypt its own data. Self-hosted Worlds can enable the same protection by implementing getEncryptionKeyForRun() with their own key management, returning the same 32-byte key per run. Cross-run writes are encrypted to the target run's public key, and setting payload retention to experimental_retention: 0 deletes inputs, outputs and stream contents once a run completes or fails. Vercel also deployed Workflow infrastructure to every Vercel Function region, so each run's state, queue dispatch and streams are served from the region where it started, instead of the single iad1 region used in Workflow SDK 4.

### Meta — Meta’s New Agreements Bring Clean Manufacturing to Data Center Supply Chains

- Company: Meta (meta.com)
- Announced: 2026-11-12T12:00:07+00:00
- Category: not stated
- Coverage: 1 outlet
- Announcement: no
- Group: covered
- Source: https://about.fb.com/news/2026/11/meta-new-agreements-clean-manufacturing-data-center-supply-chains/
- Record: https://forck.live/items/35150-meta-s-new-agreements-bring-clean-manufacturing-to-data-center-supply-chains
- Subject: Meta AI

Meta has signed a set of long-term purchase commitments across cement, steel, trucking and recycled plastics to bring lower carbon materials to commercial scale for its data center supply chains. The company says multi-year commitments give suppliers the revenue certainty they need to build these solutions, and stepping in early helps demonstrate the technologies so the broader market can adopt them. A multi-year agreement with Sublime Systems supports at least 180,000 tons of lower carbon cement production, enabling the company's first commercial-scale facility and giving Meta priority access to Sublime Cement for its data center construction pipeline. A separate agreement with Fortera supports a new plant in Ste. Genevieve, Missouri, and both companies expect their early plants to add domestic capacity for lower carbon cement and create new American jobs. A new multi-year agreement with Sweden-based Stegra supports near-zero-emission primary steel produced through a process that eliminates coal entirely from iron reduction, following an earlier deal for clean iron from Electra's demonstration plant in Jefferson County, Colorado. Meta is expanding its zero-emission freight work with Nevoya to include a California and Arizona route along the I-10 interstate, adding to earlier partnerships with NFI in Virginia and Nevoya in Texas, with permanent open access charging stations that will support other commercial fleets. A long-term purchase agreement with MacroCycle Technologies is financing that company's first commercial plant in the Southeastern US, which aims to divert hard-to-recycle plastic waste into 5,000 tons per year of recycled PET. Meta says the agreements are structured to deliver Environmental Attribute Certificates, where a supplier produces verified lower carbon materials or services, Meta pays for the environmental benefit, and the material or service is used in the local market.

### ElevenLabs — ElevenLabs brings free voice restoration to veterans, songwriters, and more patients

- Company: ElevenLabs (elevenlabs.io)
- Announced: 2026-11-12T12:00:00+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://elevenlabs.io/blog/impact-nyc-summit
- Record: https://forck.live/items/35265-elevenlabs-brings-free-voice-restoration-to-veterans-songwriters-and-more
- Subject: Voice AI

ElevenLabs is extending its 1 Million Voices program from speech to song, giving people with permanent voice loss free access to ElevenMusic so they can keep writing and releasing music in their own voices. The initiative was launched at the company's New York Summit on November 11, where educator Greg Kellner and rapper Abdi Omar performed new songs using voices they rebuilt with ElevenLabs tools. Songwriters of North America will verify eligibility and distribute access to its members and the wider songwriting community, and ElevenLabs says more partner organizations will join in the coming months. The company also named NYU Langone Health a new 1 Million Voices partner, offering the free voice restoration program directly to patients through its Rusk Rehabilitation as an assistive technology resource for permanent voice loss from ALS, head and neck cancer, stroke and other conditions. Wounded Warrior Project was announced as another Impact Program partner, with the collaboration launching on Veterans Day and centered on veterans living with voice loss; ElevenLabs will lead an educational workshop for WWP teammates on how voice AI can support warriors experiencing permanent voice loss, blindness or low vision. ElevenLabs describes 1 Million Voices as a $1 billion in-kind commitment to restore one million voices worldwide, and says the program has already helped more than 11,000 people with ALS, motor neurone disease, head and neck cancer and other conditions speak again in a voice that sounds like their own.

### Google Antigravity — Gemini 4 Hydrogen and Orchestrator

- Company: Google Antigravity (antigravity.google)
- Announced: 2026-11-12T08:00:00+00:00
- Category: new-model
- Coverage: 9 outlets
- Announcement: yes
- Group: models
- Source: https://antigravity.google/blog/gemini-4-hydrogen-in-google-antigravity
- Record: https://forck.live/items/35458-gemini-4-hydrogen-and-orchestrator
- Subject: Antigravity

Google is rolling out Gemini 4 Hydrogen inside Google Antigravity, a new model for everyday coding and agentic work, alongside a model selector option called Orchestrator that pairs Hydrogen with Gemini 4 Carbon. Google says running a frontier reasoning model with parallel agents on every routine step slows the loop and burns frontier quota, and that static model routers break down as agents run longer and need different complexity at different stages. Orchestrator instead moves dynamically between the two models within a single trajectory: Carbon can explore dependencies and plan in a read-only phase, Hydrogen writes edits and runs tests, and Hydrogen hands back to Carbon when it hits friction. Google says Orchestrator reached near-Carbon task success on Terminal-Bench 4.1 while consuming 24% less quota on average and finishing 19% faster. Hydrogen also works standalone, scoring 51 on the Artificial Analysis Intelligence Index, 54 at its highest setting, and 57% on Terminal-Bench 4.1, with four thinking levels including a new xhigh tier. Starting today, Hydrogen is rolling out to paid Google AI Pro and Ultra subscribers, while Orchestrator is rolling out for Google AI Ultra subscribers, with enterprises able to access both in Gemini for business.

### Vercel — WTF are evals? How to test agents and the software they use

- Company: Vercel (vercel.com)
- Announced: 2026-11-12T05:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://vercel.com/blog/wtf-are-evals
- Record: https://forck.live/items/35469-wtf-are-evals-how-to-test-agents-and-the-software-they-use
- Subject: v0 / AI SDK / AI Gateway

Vercel's guide explains how to build evals for agents, which it describes as users of software that do work on people's behalf. It argues that checking whether an agent completed a task is harder than it sounds, because an agent can place a "successful" order for a hot latte when the user asked for an iced one and the logs still look fine. The guide groups agentic software tests into a pyramid by how many LLM calls they need: deterministic tests with none, one-turn evals with one, and end-to-end evals with two or more across a whole task. It says that split is a good proxy for the effort, time and cost of building and running each tier, and that cheap fast tests should make up the bulk of a suite. A worked example uses sfbeans, a fictional coffee-ordering API, where a request for a large iced latte under $8 by 9:00 is checked at each layer. The end-to-end eval is built with Harbor, an open-source framework from the creators of Terminal-Bench that runs agents in containers, checks results and records their actions.

### Amazon — Build autonomous AI agents with Amazon Bedrock and Spring AI

- Company: Amazon (amazon.com)
- Announced: 2026-11-12T01:18:23+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/build-autonomous-ai-agents-with-amazon-bedrock-and-spring-ai/
- Record: https://forck.live/items/34951-build-autonomous-ai-agents-with-amazon-bedrock-and-spring-ai
- Subject: Bedrock / Nova

Amazon's guide shows Java teams how to build autonomous AI agents with Spring AI and Amazon Bedrock, using the spring-ai-agentcore library. It argues that agent frameworks do not solve scalability, governance, security and operational concerns by default, and that Bedrock and Spring AI close those gaps for teams already working on the JVM. The walkthrough covers two deployable systems: a single-agent travel assistant with tool calling, streaming responses and conversation memory, and a multi-agent intelligent document processing platform with five specialist agents, a rules engine and alerting. Both share a Spring Boot application using Spring AI ChatClient and the Amazon Bedrock Converse API, AgentCore runtime for managed deployment, AgentCore memory for short- and long-term context, and @Tool annotated methods. The guide requires JDK 21, Spring Boot 4.1 with Spring AI 2.0 and spring-ai-agentcore 2.0, Maven 3.9+, Node.js 18+, Terraform 1.5+ and Docker with Buildx. It describes an 11-stage agentic pipeline in which a router classifies a document and routes it to a specialist agent, with low-confidence results going to a human review queue.

### Anthropic — Using Claude Opus 5.5 as your daily driver for knowledge work

- Company: Anthropic (anthropic.com)
- Announced: 2026-11-12
- Category: availability-change
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://claude.com/resources/articles/using-claude-opus-5-5-as-your-daily-driver-for-knowledge-work
- Record: https://forck.live/items/35556-using-claude-opus-5-5-as-your-daily-driver-for-knowledge-work
- Subject: Claude

Anthropic has made Claude Opus 5.5 the default model in Claude on Pro, Max, Ultra, and Team plans, recommending it as the daily driver for research, analysis, and writing. The company says teams that have been using Sonnet for daily work and saving Opus for heavy analysis should reconsider, because a more capable model can finish a task by doing less work, which can offset much of its higher price per token. Opus 5.5 is built for complex work that needs careful judgment, such as research, financial analysis, and producing finished documents, decks, and spreadsheets. In one internal research test, it wrote reports on a company's quarterly performance using only a saved snapshot of the web, and 16 of its 18 reports passed an automated grader that checked every figure and quote. In another internal test, it analyzed a proposed merger between two fictional HR software companies, finishing in 63 minutes compared with 93 for Opus 5 on the same task, at half the cost. Anthropic says Opus 5.5 is its best Opus model for vision, reading dense documents, charts, screenshots, and diagrams accurately, and that even at its lowest effort setting it read values off dense charts more accurately than Opus 5 did at its highest. Fable 5.2 remains available on all paid plans for the most demanding reasoning and long-running work, though how it draws on usage depends on the plan.

### Anthropic — Introducing Claude Meetings

- Company: Anthropic (anthropic.com)
- Announced: 2026-11-12
- Category: capability-change
- Coverage: 7 outlets
- Announcement: yes
- Group: covered
- Source: https://claude.com/resources/articles/claude-meetings
- Record: https://forck.live/items/35526-introducing-claude-meetings
- Subject: Claude

Anthropic is letting Claude join video calls as a visible participant that follows the conversation, answers questions and starts work during the meeting, with Claude Meetings now in beta. The feature is starting to roll out on Team and Enterprise plans, and support is currently available for Google Meet, with other platforms coming soon. Teams can ask Claude questions out loud or in chat, and by default it answers in writing and can use connected tools with permission; it also builds docs and whiteboards live while teammates on the invite edit them, and writes the notes and action items. By default Claude stays out of calls when someone from outside the organization is invited and leaves if an outside guest joins partway, though an admin can allow it on those calls, where it posts its notice in chat but does not reply there. Claude saves a transcript, summary and screen descriptions rather than audio, and each meeting gets a separate page in claude.ai that people from the organization on the invite can see and edit. While in beta it is not available to some organizations, including those with zero data retention, HIPAA-ready configuration or customer-managed encryption keys, and Anthropic says it is working with partners such as Gong, Granola and Otter so transcripts can be sent to other tools.

### Liquid AI — LiquidCodec: High-Fidelity Audio Tokenization Optimized for Speed

- Company: Liquid AI (liquid.ai)
- Announced: 2026-11-12
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.liquid.ai/blog/liquidcodec
- Record: https://forck.live/items/35429-liquidcodec-high-fidelity-audio-tokenization-optimized-for-speed
- Subject: LFM / d1 models

Liquid AI has built LiquidCodec, an in-house streaming audio codec that turns speech into tokens for its LFM3 audio models and converts them back into audio. The company says the roughly 74M-parameter codec matches similarly sized models on reconstruction quality and compression while beating them on speech generation and token-to-waveform decoding speed. Liquid AI previously relied on the third-party codec Mimi in LFM2-Audio-1.5B, then replaced only its detokenizer in LFM2.5-Audio-1.5B for 8x faster detokenization, but the third-party tokenizer still limited audio-token quality and architecture choices. LiquidCodec swaps Mimi's residual vector quantization for grouped finite scalar quantization, which reduces sequential dependencies between codebooks and allows parallel prediction of audio tokens. The codec keeps a 12.5 Hz frame rate and models 24 kHz waveforms, with a symmetric encoder and decoder and a fully causal encoder for real-time, low-latency streaming. Training ran in two multilingual stages, the first on 264,000 hours across 107 languages and the second fine-tuned only on high-fidelity waveforms.

### Apple — Theoretical Limits of Language Model Alignment

- Company: Apple (apple.com)
- Announced: 2026-11-12
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://machinelearning.apple.com/research/theoretical-limits-alignment
- Record: https://forck.live/items/35426-theoretical-limits-of-language-model-alignment
- Subject: Machine Learning Research

Apple researchers have published a paper characterizing the information-theoretic limits of KL-regularized language model alignment, deriving the maximum expected reward gain possible under a fixed KL-divergence budget. The authors say their first result gives a closed-form expression for the optimal reward improvement, governed by a Jeffreys divergence term rather than the square root of KL used in prior analyses. They reformulate that expression as a covariance under the base model, yielding an estimator that predicts achievable alignment gains from base model samples alone. The paper extends the analysis to proxy rewards, showing the gap between ideal and proxy alignment, known as reward hacking, grows with the magnitude of reward error and when the KL penalty factor decreases. The authors prove that reward ensembling mitigates reward hacking, which they describe as a theoretical justification for a technique used in practice. Empirically, they compute the KL-reward Pareto frontier for two tasks, safety and summarization, and report that best-of-N closely approaches the theoretical limit while PPO and GRPO remain substantially suboptimal.

### Amp — Bye Bye, Dial

- Company: Amp (ampcode.com)
- Announced: 2026-11-12
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://ampcode.com/news/bye-bye-dial
- Record: https://forck.live/items/35407-bye-bye-dial
- Subject: Amp

Amp is retiring the Dial and its four effort modes in favor of a model selector that lets users choose the model and its effort for each thread. The company says the change follows its addition of bring-your-own subscriptions and API keys, which made the selected model decide not only which tokens were produced but who got billed, so it became hard to tell who was paying for which model. Users can now switch both model and effort in the middle of a thread, and the selector shows which subscription or API key will be billed, or whether the request uses Amp credits. New threads start on Staff Pick, which is Claude Opus 5.5 today, and Amp says it will change the pick when another model benchmarks better. Existing threads keep running on the model they started with, and Dial settings do not carry over: the modes ran GLM-5.3-Flash, Claude Opus 5.5, GPT-6 Astra and Claude Fable 5.2. Amp notes that a model switch still costs more, because prompt caches belong to one model and the new model reads the whole thread uncached.

### Anthropic — Introducing Claude for Manufacturing

- Company: Anthropic (anthropic.com)
- Announced: 2026-11-12
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://claude.com/resources/articles/claude-for-manufacturing
- Record: https://forck.live/items/35399-introducing-claude-for-manufacturing
- Subject: Claude

Anthropic launched Claude for Manufacturing, a plugin with 17 skills covering quoting, quality, operations, supply chain, design and setup, aimed at manufacturers who lose time to manual data entry. The plugin connects to Autodesk Fusion, Infor, Microsoft 365 and Oracle NetSuite, so Claude can read a manufacturer's own records when it drafts a quote or compiles a production report. Anthropic cites a 2025 Manufacturing Leadership Council survey in which 79% of manufacturers said they enter data manually. Skills are built to end with a person's review, and Claude can link figures and claims back to the underlying data and documents. Manufacturers without a connector can save regular exports to Microsoft 365 and have Claude work from those files. Infor is also releasing its own plugin, which lets teams make changes in Infor from Claude, such as releasing a purchase order to another supplier and rescheduling production, with Claude's access following the buyer's Infor permissions. The plugin is available on Team or Enterprise plans, and Anthropic says it is developing the Model Hardware Standard, a research preview for programmable equipment, since Claude for Manufacturing does not operate equipment.

### Factory AI — Launching Factory’s Cyber Defense Network

- Company: Factory AI (factory.com)
- Announced: 2026-11-12
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://factory.com/news/cyber-defense-network
- Record: https://forck.live/items/35367-launching-factory-s-cyber-defense-network
- Subject: Droid

Factory is launching the Cyber Defense Network, an ecosystem of cybersecurity companies whose findings connect into Security Factory so vulnerabilities can be investigated and fixed inside the engineering workflow. Founding partners include Snyk, Palo Alto Networks, Semgrep, Cogent, Casco, DepthFirst and Detail, among others. The company says security teams already rely on specialized tools to identify vulnerabilities, assess risk and prioritize what needs attention, and that this work becomes increasingly difficult to manage at scale as findings grow. The network links those tools to Factory's ability to carry out the engineering work required to resolve the findings. Partner findings flow into Security Factory, where Factory Droids investigate vulnerabilities, implement code changes and validate fixes through testing, in three stages the company calls signals, triage and continuous remediation. Organizations can keep their existing security tools and workflows while automating more of the work needed to act on findings. Factory says it plans to expand the network to include additional cybersecurity innovators and model providers, and is recruiting security leaders and companies to join.

### Anthropic — The Claude Tag guide for startups

- Company: Anthropic (anthropic.com)
- Announced: 2026-11-12
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://claude.com/resources/articles/claude-tag-guide-for-startups
- Record: https://forck.live/items/35306-the-claude-tag-guide-for-startups
- Subject: Claude

Anthropic's guide for startups explains which Slack channels to add Claude Tag to first, based on interviews with eight high-growth startups. The guide says triage channels are usually the first place teams put Claude, because constant Sentry errors, Datadog alerts and support requests make the signal hard to find. In that setup Claude reads each item as it lands, investigates the codebase and connected tools, and shares what it found before the on-call engineer opens a laptop. If the fix is contained it opens a draft PR, and if not it routes the issue to the right person with the investigation attached; a person still reviews each fix before it merges. Kikoff runs Claude as a first responder in channels covering deploys, CI alerts, Sentry exceptions, dogfooding and support, and the guide says its biggest wins came once the team let Claude jump in without being mentioned. NavigateAI runs Claude Tag in about 90 channels and, since June, Claude has opened about 223 PRs from Slack threads, with 134 merged at a median of two hours from open to merge.

### Hugging Face — The Definitive Fine-Tuning Guide for 3D Generation Models

- Company: Hugging Face (huggingface.co)
- Announced: 2026-11-12
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://huggingface.co/blog/3d-lora-finetuning
- Record: https://forck.live/items/35245-the-definitive-fine-tuning-guide-for-3d-generation-models
- Subject: Platform

Hugging Face published fine-tuning scripts and containers that let users adapt TRELLIS.2, a 3D generation model whose unified latent representation the company says avoids switching models per output format, to their own 3D assets. The guide answers a practical gap: a general-purpose model such as TRELLIS.2 can make errors in details that matter for a specific application, and neither TRELLIS nor TRELLIS.2 offers an easy way to adapt it to a task. Hugging Face tests low-rank adaptation on a subset of the Google Scanned Objects dataset with 188 footwear objects, where the pretrained model usually produced shoes that looked correct but had incorrect soles. LoRA trains small adapter matrices while keeping the original weights frozen, cutting trainable parameters and the memory needed for gradients and optimizer state, and the work uses Hugging Face's PEFT library to add and train them. The guide adds adapters only to the sparse-structure flow DiT and the shape flow DiT, leaving the material generation stage untouched, because its dataset has diffuse color maps but lacks the separate material maps that stage needs. The scripts can run locally with Docker or on Hugging Face Jobs, which requires an HF_TOKEN and Hugging Face account access to the gated DINOv3 model.

### Anthropic — J. Crew brings its signature approach to styling to life with Claude

- Company: Anthropic (anthropic.com)
- Announced: 2026-11-12
- Category: product-launch
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://claude.com/resources/articles/j-crew-style-advisor-with-claude
- Record: https://forck.live/items/35211-j-crew-brings-its-signature-approach-to-styling-to-life-with-claude
- Subject: Claude

J.Crew launched Style Advisor, an AI stylist on JCrew.com that helps customers shop by complete outfit rather than one product at a time. The retailer's product team used Claude Fable 5.1 to plan how the tool should reason like a stylist, then built a working prototype with Claude Opus 5.5 in Claude Code, and J.Crew engineers took that prototype to production. In each conversation, Claude Sonnet 5.5 acts as the stylist and decides each step of the response, while Claude Haiku 5.5 handles quick, focused tasks so Style Advisor answers quickly. Every recommendation starts from looks J.Crew stylists created for the brand's catalog shoots and is shown in photos, so customers see the full outfit. Customers can start from a single piece, name an occasion such as a vacation, holiday party or wedding, and keep refining the result as a visual conversation. Style Advisor is available now for women's and men's styles in the U.S., and J.Crew plans to make it more personal with each customer's preferences and past purchases and to bring it to more markets and product lines.

### Recraft — MiniMax H3 Image, Highspeed, and Lite in Recraft Studio

- Company: Recraft (recraft.ai)
- Announced: 2026-11-12
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.recraft.ai/blog/minimax-h3-image-highspeed-lite-in-recraft-studio
- Record: https://forck.live/items/35206-minimax-h3-image-highspeed-and-lite-in-recraft-studio
- Subject: Recraft V4

Recraft added four MiniMax H3 models to Recraft Studio: the image models H3 Image and H3 Image Highspeed, and the video models H3 Highspeed and H3 Lite. MiniMax's H3 began as a video model, and this week it became a family that also makes stills, with H3 Image released on November 10, 2026 on the same system as the H3 video model. In Recraft Studio, both image models generate and edit images at 2K or 4K, while Highspeed takes the same prompts and settings, returns results faster and costs more per image. The two video models generate clips of 4 to 15 seconds at 480p or 768p, always with sound, and can start from a prompt or a first frame. H3 Lite starts at 4 credits per second, which Recraft says makes it the model for trying several motion directions before committing to one, and H3 Image starts at 6 credits per image. Highspeed and Lite top out at 768p, so Recraft directs users to MiniMax H3 for higher resolution, and according to MiniMax's API documentation H3 Image does not output transparent backgrounds. Recraft ran six image prompts and four video prompts through the models, and says both image models set a three-line letterpress poster without a spelling error.

### Anthropic — Operate Claude Managed Agents inside your security perimeter

- Company: Anthropic (anthropic.com)
- Announced: 2026-11-12
- Category: availability-change
- Coverage: 2 outlets
- Announcement: yes
- Group: covered
- Source: https://claude.com/resources/articles/self-hosted-claude-managed-agents
- Record: https://forck.live/items/34559-operate-claude-managed-agents-inside-your-security-perimeter
- Subject: Claude

Anthropic is letting teams run Claude Managed Agents entirely inside their own environment, with self-hosted Managed Agents now in private beta. Managed Agents is a suite of composable APIs for building and deploying autonomous agents, and Anthropic had previously offered it only as a hosted service on the Claude Platform, later adding self-hosted sandboxes. The company says agents now take on critical work such as fixing broken builds, testing releases, triaging alerts and clearing bug backlogs, and that teams wanting the whole service in-house previously had to build the platform themselves. Under self-hosting, the harness, sandboxes, credentials and session history all run in the customer's cloud or on-premises, and the customer operates the Managed Agents server, its database and the workers running shell and file tools. Model requests can go to Anthropic directly, to Amazon Bedrock, Google's Agent Platform or Microsoft Foundry, or through the customer's own gateway, while developers keep using the same core APIs as the hosted service. Session history stays in the customer's database, and Anthropic says an existing zero data retention agreement applies to Claude requests as it does to any other Messages API call.

### OpenAI — Gilbert + Tobin rethinks the business of law with AI

- Company: OpenAI (openai.com)
- Announced: 2026-11-12
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/gilbert-tobin-sam-nickless
- Record: https://forck.live/items/34378-gilbert-tobin-rethinks-the-business-of-law-with-ai
- Subject: GPT / ChatGPT / API

Gilbert + Tobin rolled out ChatGPT Enterprise, ChatGPT Work and Codex to move from individual assistance to multistep agentic workflows, with lawyers reviewing and approving the output. The Australian corporate law firm says the value has not come from cutting costs or headcount but from freeing people for client meetings, coaching and mentoring. It began by letting staff experiment with ChatGPT under a clear rule that no client data or confidential information be entered, then added a secure channel with Australian data residency, contractual protections, role-based access and administrative controls. Adoption reached 87% active usage among enabled ChatGPT users, more than twice the rate the firm typically sees for other tools, which it credits to leaders using the technology themselves rather than mandating usage targets. CEO Sam Nickless built a custom GPT digital twin from his own writing and feedback, now shifting to Plugins, so his team can test their thinking against his context before coming to him; he says it does not make decisions or speak for him. He also used ChatGPT Work for a confidential annual review of 120 partners, updating files from a spreadsheet and bundling PDFs in minutes, at a cost of about $62 shown in the usage dashboard.

### Amazon — OpenAI Codex on Amazon Bedrock

- Company: Amazon (amazon.com)
- Announced: 2026-11-11T23:35:47+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/openai-codex-on-amazon-bedrock/
- Record: https://forck.live/items/34923-openai-codex-on-amazon-bedrock
- Subject: Bedrock / Nova

Amazon published a guide for running OpenAI Codex against Amazon Bedrock, aimed at platform teams that want Codex without building a second model-access, identity and operations stack. Codex still runs on the developer workstation, and the design choice is how its model requests reach Bedrock and where policy and operational controls are applied. The post recommends starting with Native AWS Access for a small, least-privilege pilot, and offers AgentCore Gateway when a team wants AWS managed gateway infrastructure for routing or Amazon Bedrock Guardrails, or an LLM Gateway when controls must sit in the request path. A custom integration is described as the advanced version of the LLM Gateway pattern rather than a quick start. The examples use GPT-5.6 Sol and come from the open source Guidance for Codex on AWS repository, which includes quick starts for Native AWS Access, AgentCore Gateway, LiteLLM on ECS, Portkey and the AWS Distro for OpenTelemetry. The walkthrough has three checks: sign in with an IAM Identity Center profile and call Amazon Bedrock Runtime directly, point Codex at the approved model, then run a safe Codex task.

### Amazon — Semantic diffing of formal policies with anti-unification

- Company: Amazon (amazon.com)
- Announced: 2026-11-11T23:33:03+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/semantic-diffing-of-formal-policies-with-anti-unification/
- Record: https://forck.live/items/34924-semantic-diffing-of-formal-policies-with-anti-unification
- Subject: Bedrock / Nova

Amazon's post explains a semantic diff engine that uses anti-unification over an e-graph to show what two formal policies share and where they disagree. A textual diff often flags syntactic changes that do not affect meaning, so a real behavioral change, such as an Or where the requirement says and, can get lost in the noise. The engine supplies algebraic laws and domain rules that justify equalities, such as equating an external destination with one that is not internal, and shows the shared parts once while keeping both versions of each differing part for review. If no differences remain, the policies are equivalent under those laws and rules, though when the rules or the search are incomplete a remaining difference might reflect an equivalence the engine has not found. One use case is autoformalization, where an LLM generates formal artifacts from prose and can produce two different formalizations with no ground truth; the model writes the artifact and plays no part in computing the diff. Requirements can also become OpenShell sandbox policies or Common Expression Language expressions, and the engine reads encodings of the relevant syntax rather than importing native policies from those formats. In a worked example, two candidate rules for the requirement that sensitive data sent to an external destination must be encrypted produce an anti-unifier with two Variants nodes, one of which is a difference in vocabulary rather than meaning.

### Amazon — Detecting and recovering from silent data corruption on Amazon SageMaker HyperPod

- Company: Amazon (amazon.com)
- Announced: 2026-11-11T23:30:03+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/detecting-and-recovering-from-silent-data-corruption-on-amazon-sagemaker-hyperpod/
- Record: https://forck.live/items/34908-detecting-and-recovering-from-silent-data-corruption-on-amazon-sagemaker
- Subject: Bedrock / Nova

Amazon's guide shows how to detect silent data corruption in PyTorch DDP training on SageMaker HyperPod, where a GPU returns a wrong but plausible number without raising an error. Existing HyperPod health monitoring, deep health checks and automatic node recovery cannot catch it, because the hardware never reports a fault, so detection has to come from the training job itself. The post adds three checks to the training loop: a non-finite gradient check for NaN, Inf and overflow, a ring replay audit that recomputes a micro-batch from another node and compares loss bit for bit and gradients by fingerprint, and a weight fingerprint comparison after the optimizer step. When a check fires, the suspect rank is mapped to its node, the job is stopped, the node is cordoned, and HyperPod replaces it before training resumes from a checkpoint saved before the suspect window. The checks were tested with software-injected faults on a cluster of 4 ml.p5.48xlarge instances, 32 NVIDIA H100 GPUs, and repeated on three instances, the minimum the checks need. The post cites Meta's finding that during a 54-day period of Llama 3 405B pre-training on up to 16K H100 GPUs, 6 of 419 unexpected job interruptions were attributed to silent data corruption.

### Runway — Praxis-1 Is Now Available

- Company: Runway (runway.com)
- Announced: 2026-11-11
- Category: open-weight-release
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://runway.com/news/company-news/praxis-1-is-now-available
- Record: https://forck.live/items/34651-praxis-1-is-now-available
- Subject: Gen / Aleph

Runway has released Praxis-1, its world action model for robotics, to everyone with open weights under the Apache 2.0 license, after introducing it in September. The model is a 27B-parameter mixture-of-experts diffusion transformer that predicts video and actions together, and Runway says it is the top-performing world action model on NVIDIA's RoboLab-120 leaderboard. On RoboLab-120, Praxis-1 reaches 46.2% success across 1,200 episodes, compared with 42.9% for FLUX 3 Action, 36.8% for Cosmos3 Nano and 28.0% for π0.5. Launch partners Noble Machines, Standard Bots and Ultra have been running it on their own hardware, including a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot. The release covers the weights and the inference and fine-tuning code, with support for both full fine-tuning and LoRA, and it runs across NVIDIA hardware from H100 and B200 data center GPUs to RTX 5090, RTX PRO 6000 and Jetson Thor. Praxis-1 is available on GitHub, on Hugging Face with native LeRobot support, and in vLLM-Omni, with hosted inference on Reactor; Runway says the license permits commercial use.

### Amazon — Introducing the Web Search API on Amazon Bedrock AgentCore

- Company: Amazon (amazon.com)
- Announced: 2026-11-11T23:13:16+00:00
- Category: api-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://aws.amazon.com/blogs/machine-learning/introducing-the-web-search-api-on-amazon-bedrock-agentcore/
- Record: https://forck.live/items/34909-introducing-the-web-search-api-on-amazon-bedrock-agentcore
- Subject: Bedrock / Nova

Amazon is exposing Web Search on Amazon Bedrock AgentCore as a standalone API, so developers can call web search directly from their own code with a single request instead of routing through AgentCore Gateway. The company says developers asked for a lower-level integration point for custom agent loops, non-agent use cases such as dashboards and fact-checking pipelines, tighter control over retries and caching, and applications that already have their own API layer. The API is the InvokeWebSearch operation on the bedrock-agentcore data plane, the same endpoint other AgentCore tools such as AgentCore Browser use, and requests are signed with AWS Signature Version 4 rather than an API key or secret. The API applies domain include or exclude filters and a publication time range, then returns up to maxResults ranked results from a web index Amazon operates, each with a text extract and, when available, the page title, URL and publication time. Amazon says queries are served inside AWS and aren't sent to a third-party search engine, and the API runs on the same search infrastructure that powers Alexa+, Kiro, Quick and other Amazon products. Callers need one IAM permission, bedrock-agentcore:InvokeWebSearch, on an AWS managed resource, the same action and resource a gateway's service role uses, so one policy covers both routes, and there are no resources to create first.

### Luma — Character Swap & Insert is now in Luma.

- Company: Luma (lumalabs.ai)
- Announced: 2026-11-11T22:30:10+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://x.com/LumaLabsAI/status/2120645286519562294
- Record: https://forck.live/items/34893-character-swap-insert-is-now-in-luma

Luma has added Character Swap & Insert to its video product, letting users change the cast, the car or other elements of a shot while keeping the same shot. The feature is described as answering the kind of client notes that ask for a swapped cast, an added dog or someone removed. The post gives no further detail on availability, pricing or limits.

### Amazon — ICYMI: What landed for AI builders in October 2026

- Company: Amazon (amazon.com)
- Announced: 2026-11-11T22:26:40+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/icymi-what-landed-for-ai-builders-in-october-2026/
- Record: https://forck.live/items/34878-icymi-what-landed-for-ai-builders-in-october-2026
- Subject: Bedrock / Nova

AWS's October 2026 roundup covers new models on Amazon Bedrock, agent reliability updates in AgentCore, and cost reporting for AI builders. Claude Haiku 5.5 is built for subagents and high-volume work; according to Anthropic it is the most capable Haiku yet and costs around 75 percent less than Claude Haiku 4.5 for most tasks, and it is the first Haiku with effort controls. OpenAI GPT-6.1 Terra is now generally available on Bedrock, where AWS says it performed on par with GPT-6 Astra on OpenAI's Agents' Last Exam evaluation and completed tasks more than three times faster, with input and output token prices at least 50 percent lower than GPT-6.1 Sol and GPT-5.6 Terra. Amazon Nova 2.5 Sonic, the latest speech-to-speech model, improves reasoning, instruction following and tool-calling accuracy at the same price as Amazon Nova 2 Sonic. GLM 5.3 by Z.ai, an open-weight model for agentic coding with a 1-million-token context window and prompt caching, is available to eligible enterprise customers, while TwelveLabs Pegasus 1.5 turns entire videos into structured, time-coded metadata. AgentCore CLI reached version 1.0 with faster startup and a consistent command structure, and AgentCore Gateway now supports the Amazon Bedrock InvokeModel APIs and model fallback, routing requests to the next model in a chain when the primary is throttled or unreachable.

### Amazon — The attendee’s guide to the AWS re:Invent 2026 AI/ML track

- Company: Amazon (amazon.com)
- Announced: 2026-11-11T21:05:59+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/the-attendees-guide-to-the-aws-reinvent-2026-ai-ml-track/
- Record: https://forck.live/items/34855-the-attendee-s-guide-to-the-aws-re-invent-2026-ai-ml-track
- Subject: Bedrock / Nova

AWS has published an attendee guide to the AI/ML track at re:Invent 2026, which runs from November 30th to December 4th with more than 140 sessions in that track alone. The event offers over 2,000 learning sessions across skill levels, and AWS says many sessions will be available on demand, with keynotes streamed live for those who cannot attend in person. Sessions come in several formats: 60-minute lecture-style breakouts, chalk talks where attendees question presenters at a whiteboard, one-hour builder sessions, two-hour team workshops, live code talks, and 20-minute lightning demos. AWS records the breakout sessions and posts them to the AWS Events YouTube channel a few days later, while the other formats are live and not recorded. The guide maps the sessions onto the AWS AI stack, from data centers and AI chips through the data foundation, models and inference, agent development, and ready-to-use agentic solutions, with security, governance and cost optimization across every layer. The AI/ML track concentrates on AI chips, models, training and inference, and agent development and management, and the guide orders sessions within each section from newcomer material to technical deep dives.

### Amazon — Automating AWS Agent Registry updates with push-based architecture

- Company: Amazon (amazon.com)
- Announced: 2026-11-11T20:52:48+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/automating-aws-agent-registry-updates-with-push-based-architecture/
- Record: https://forck.live/items/34851-automating-aws-agent-registry-updates-with-push-based-architecture
- Subject: Bedrock / Nova

Amazon's guide shows how to keep AWS Agent Registry records current by wiring CloudTrail, EventBridge and Lambda into a push-based sync for MCP servers. The problem it addresses is staleness: when an MCP server gains a tool or changes its schema, the registry record must be updated or consumers discover outdated metadata. The post says stale metadata reduces registry reliability and pushes teams toward hardcoded endpoints, informal knowledge sharing and stale documentation. The pipeline captures UpdateAgentRuntime API calls in CloudTrail, routes them through EventBridge to a Lambda in the registry account, which authenticates through AgentCore Identity, calls the MCP server's initialize and tools/list, diffs the result and updates the matching record. CloudTrail records the event as a management event with a delivery delay of approximately 5–15 minutes. The sample uses two accounts, with the registry account owning the registry, EventBridge rule, Lambda and CloudTrail logs, and the MCP server account hosting the runtime and Cognito OAuth provider; a single-account setup skips cross-account forwarding. The Lambda updates existing records only and does not create new ones, so the MCP server must already be pre-registered.

### Amazon — How DXC Technology built a multi-model AI substrate on Amazon Bedrock

- Company: Amazon (amazon.com)
- Announced: 2026-11-11T20:45:45+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/how-dxc-technology-built-a-multi-model-ai-substrate-on-amazon-bedrock/
- Record: https://forck.live/items/34850-how-dxc-technology-built-a-multi-model-ai-substrate-on-amazon-bedrock
- Subject: Bedrock / Nova

DXC Technology built the model substrate beneath its DXC OASIS IT-operations harness on Amazon Bedrock, reaching multiple foundation models through one managed interface with governance controls applied uniformly. The company runs managed IT services for banks, hospitals and government agencies, and designed the substrate so no single provider, framework or deployment target could become a point of failure. Amazon Bedrock Guardrails enforce content filtering, PII redaction and denied-topic policies across every model, while AWS PrivateLink keeps inference traffic inside the customer's VPC and serverless scaling removes self-managed GPU infrastructure as a failure point. DXC says prompts and completions stay within the AWS data boundary, are not shared with third-party model providers and are not used to train the underlying foundation models. DXC began the work as a prototype in early 2025 and reached general availability in March 2026, scaling toward 130 customer environments on Amazon EKS with per-tenant namespaces and GitOps rollout. Since general availability, DXC says DXC OASIS has reduced resolution time by 95% on average across 56 measured use cases.

### NVIDIA — Meet Déjà View, a model that turns photos taken from different angles into 3D geometry and camera positions.

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-11T20:38:27+00:00
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://x.com/NVIDIAAI/status/2120617173286883692
- Record: https://forck.live/items/34849-meet-d-j-view-a-model-that-turns-photos-taken-from-different-angles-into-3d

NVIDIA introduced Déjà View, a 117M-parameter model that turns photos taken from different angles into 3D geometry and camera positions. The model loops a single transformer block, and NVIDIA says it matches or outperforms models up to 10x larger across five benchmarks while using comparable or less compute. Researchers can dial the number of loops to fit their compute budget, according to the company. The work is an oral paper at NeurIPS 2026, and NVIDIA says code is available alongside it.

### Google Research — Using wearables to enable cardiovascular-kidney-metabolic (CKM) risk profiling for personalized coaching

- Company: Google Research (research.google)
- Announced: 2026-11-11T19:54:00+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://research.google/blog/using-wearables-to-enable-cardiovascular-kidney-metabolic-ckm-risk-profiling-for-personalized-coaching/
- Record: https://forck.live/items/34804-using-wearables-to-enable-cardiovascular-kidney-metabolic-ckm-risk-profiling
- Subject: Research

Google Research is introducing an investigational machine learning framework that turns passive wearable signals into a coachable health age and simulates personalized 12-week habit trajectories for cardiovascular-kidney-metabolic health. The coachable health age model is built from modifiable daily behaviors and physiology captured passively by consumer wearables, rather than trained to guess a person's birth year. It integrates resting heart rate, daily step count, sleep duration and weekly Active Zone Minutes, with optional markers such as VO2max, sleep schedule regularity and insulin resistance when available. A sensor world model called SensorWM, trained on over 2.2 million person-weeks of wearable data, simulates "what-if" health trajectories, paired with a language model agent that turns those simulations into clinician-vetted coaching plans. When routine blood tests are available, the framework can be expanded to combine wearable streams with lab biomarkers to detect early, subclinical strain across the heart, kidney and metabolic systems. In the All of Us research program, across more than 68,000 person-years of Fitbit and electronic health record follow-up, healthy participants whose health age was older than their calendar age at baseline were significantly more likely to develop hypertension, cardiovascular disease or type 2 diabetes over the next eight years. Google says adding the health age metric to age, sex and BMI significantly improved disease risk prediction, especially among adults under 50.

### Perplexity — Automatic output-token limits

- Company: Perplexity (perplexity.ai)
- Announced: 2026-11-11T19:23:41+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://docs.perplexity.ai/docs/resources/changelog#november-2026
- Record: https://forck.live/items/34791-automatic-output-token-limits
- Subject: Sonar

You can omit max_output_tokens when using any model, including Anthropic models. Each generation can produce up to 128,000 output tokens, capped by the model's supported limit.

### Amazon — Accelerating generative AI delivery with reusable assets

- Company: Amazon (amazon.com)
- Announced: 2026-11-11T19:19:21+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/accelerating-generative-ai-delivery-with-reusable-assets/
- Record: https://forck.live/items/34790-accelerating-generative-ai-delivery-with-reusable-assets
- Subject: Bedrock / Nova

AWS is pointing builders to a catalog of 23 production-ready generative AI assets from its Generative AI Innovation Center, now part of AWS Frontier AI Engineering & Services, so teams stop rebuilding RAG pipelines, document processors and agent scaffolding. The post says the assets are validated for quality and security before publication, carry deployment guides, and are instrumented to track adoption across customer engagements. It names three root causes of redundant work: no single source of truth, no quality signal, and an informal reuse culture. AWS cites DoorDash, which it says built a voice-operated contact center solution on the Contact Center GenAI Agent asset handling over 100,000 calls per day at 2.5-second latency, with a 50x increase in testing capacity and development time cut by 50 percent using Amazon Bedrock. Ricoh, using the Accelerated Intelligent Document Processing on AWS asset, is reported to have gone from 10,000 to 70,000 documents monthly, cut engineering hours per deployment from 80 to under 5, and shortened onboarding from 4–6 weeks to 2–3 days. Lyft used the same Contact Center GenAI Agent asset for riders and drivers, which AWS offers as evidence of the asset's versatility.

### xAI — Introducing Grok Bot 1.0.

- Company: xAI (x.ai)
- Announced: 2026-11-11T18:30:20+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://x.com/SpaceXSI/status/2120584928983441826
- Record: https://forck.live/items/34772-introducing-grok-bot-1-0

xAI introduced Grok Bot 1.0, saying bots can be shared and are more collaborative and capable than ever. The post gives no further detail on what the bots do, how sharing works, or who can use them.

### Runway — Seedance 2.5 in 4K is here.

- Company: Runway (runway.com)
- Announced: 2026-11-11T18:03:21+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://x.com/runwayml/status/2120578138111656227
- Record: https://forck.live/items/34764-seedance-2-5-in-4k-is-here

Runway has made Seedance 2.5 available in 4K resolution inside its own product, giving users of the platform access to the higher-resolution version of the video model. The post does not say who built Seedance 2.5, only that the 4K version is now offered through Runway. Runway directs users to a link to get started with it.

### Google DeepMind — AlphaGenome 2: Teaching AI to read the entire human genome

- Company: Google DeepMind (deepmind.google)
- Announced: 2026-11-11T18:00:31+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.youtube.com/watch?v=uND6iY2qp_c
- Record: https://forck.live/items/34770-alphagenome-2-teaching-ai-to-read-the-entire-human-genome

98% of human DNA doesn't build proteins—it controls how our cells work, and holds vital clues to why diseases happen and why 90% of new drugs fail in clinical trials. Understanding the genome at the molecular level is the key to deciphering the blueprint of human life.

### Meta — Intelligent Delivery: Making Every Unit of Compute Serve Better Ads

- Company: Meta (meta.com)
- Announced: 2026-11-11T17:00:38+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://engineering.fb.com/2026/11/11/ml-applications/intelligent-delivery-serve-better-ads/
- Record: https://forck.live/items/34730-intelligent-delivery-making-every-unit-of-compute-serve-better-ads
- Subject: Llama / infrastructure

Meta is describing Intelligent Delivery, an ML-driven system that allocates compute across its ads serving stack toward what it calls the system-wide optimum. The company says the system has contributed roughly 10% incremental ads ranking performance on top of model and system advancements across major surfaces from 2017 through the first half of 2026. Meta frames the problem as one of finite capacity: ads ranking models are growing toward LLM scale faster than the hardware to run them, and the same extra unit of compute returns very different value depending on surface, user intent, candidate quality and system state. Earlier allocation used coarse segments, with engineers writing policies per group, but Meta says every new signal multiplied the rules and independent segments missed nonlinear interactions among thousands of signals. Meta reframed this into two ML problems, which requests benefit most from extra compute and how resource settings change their outcomes, and reports about a 20% improvement in identifying the highest-impact ad requests in offline benchmarks on major surfaces relative to its previous production model. The Resource Core Model behind this has 6B sparse parameters and 20M dense parameters, runs once before candidate retrieval, and is shared across the delivery funnel; Meta fine-tunes it into specialized models that predict outcome, compute consumption and latency impact.

### Google Cloud — Don’t just prompt AI — architect it: Introducing the Professional Agentic Architect Certification

- Company: Google Cloud (cloud.google.com)
- Announced: 2026-11-11T17:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://cloud.google.com/blog/topics/training-certifications/google-cloud-agentic-architect-certification-dont-just-prompt-ai-design-it/
- Record: https://forck.live/items/34738-don-t-just-prompt-ai-architect-it-introducing-the-professional-agentic
- Subject: Gemini Enterprise

It just got easier to prove your expertise in building production-ready agents. Today, Google Cloud is introducing the Professional Agentic Architect certification as the latest milestone in our Gemini Agent Ready (GEAR) initiative.

### Google Cloud — Introducing new AI functions for AlloyDB: ai.classify and ai.extract

- Company: Google Cloud (cloud.google.com)
- Announced: 2026-11-11T17:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://cloud.google.com/blog/products/databases/ai-classify-and-ai-extract-functions-in-alloydb-ai/
- Record: https://forck.live/items/34737-introducing-new-ai-functions-for-alloydb-ai-classify-and-ai-extract
- Subject: Gemini Enterprise

Transactional databases store vast amounts of unstructured text — clinical notes, customer reviews, support tickets, invoices, and property descriptions.

### Krea — Seedance 2.5 4K is now on Krea.

- Company: Krea (krea.ai)
- Announced: 2026-11-11T16:45:04+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://x.com/krea_ai/status/2120558438568882310
- Record: https://forck.live/items/34722-seedance-2-5-4k-is-now-on-krea

Krea added Seedance 2.5 4K to its platform, calling it "the highest quality AI video model yet" and inviting users to try it now. The post gives no detail on which plans can reach the model, what it costs, or who built it.

### Amazon — Accelerating product development with AI-driven lifecycle management (AI-PDLC) in AWS Quick

- Company: Amazon (amazon.com)
- Announced: 2026-11-11T16:05:33+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/accelerating-product-development-with-ai-driven-lifecycle-management-ai-pdlc-in-aws-quick/
- Record: https://forck.live/items/34690-accelerating-product-development-with-ai-driven-lifecycle-management-ai-pdlc
- Subject: Bedrock / Nova

AWS is showing teams how to configure a six-agent suite in AWS Quick that gathers scattered product signals and drafts the artifacts behind a product decision, with human approval at every gate. The AI-Driven Product Development Lifecycle framework spans eight phases; AWS's walkthrough covers the first four — Discover, Decide, Design and Prototype — where the product team works, and says coding agents such as Kiro handle the rest. An Orchestrator is the only agent users talk to; it reads each specialist's question framework and dispatches artifact generation to the Discovery, Prioritize, Working Backwards, Spec Generator and Designer agents. Enterprise connectors pull context from Slack, Outlook, Jira, Salesforce, SharePoint and ServiceNow, and a personal knowledge graph links the people, projects and decisions across those tools. Everything the agents produce lands in a single prototype_spec.md file holding acceptance criteria, personas, screens and a data model, and AWS says every decision traces back to a source. The pipeline advances only after human approval at each step, and AWS says AWS Quick can delegate development work to coding agents through the Agent Client Protocol, an open JSON-RPC 2.0 standard with no vendor lock-in.

### Reason Machines — Introducing: Multiplayer

- Company: Reason Machines (reasonmachines.com)
- Announced: 2026-11-11T15:57:35+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://x.com/reasonmachines/status/2120546487231033817
- Record: https://forck.live/items/34699-introducing-multiplayer

Reason Machines says its agents can now coordinate with teammates' agents, letting every teammate and their agents work together in one unified workspace. The company is pitching the feature as a way to stop coding solo, since users can now talk to other agents or humans through the shared workspace. The post frames the change around collaboration rather than individual use, with teammates and their agents sharing a single place to work. Reason Machines gives no details on availability, pricing or which plans include the feature.

### Amazon — UAEF: A framework-agnostic system for closed-loop evaluation of AI agents

- Company: Amazon (amazon.com)
- Announced: 2026-11-11T15:50:55+00:00
- Category: developer-tool-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://aws.amazon.com/blogs/machine-learning/uaef-a-framework-agnostic-system-for-closed-loop-evaluation-of-ai-agents/
- Record: https://forck.live/items/34678-uaef-a-framework-agnostic-system-for-closed-loop-evaluation-of-ai-agents
- Subject: Bedrock / Nova

Amazon introduced the Universal Agent Evaluation Framework (UAEF), a framework-agnostic system that scores AI agents with one consistent set of metrics so teams can compare agent versions before release. Teams building agents on AWS use frameworks such as LangGraph, LangChain, Strands Agents and Amazon Bedrock AgentCore, and each records execution differently, so metrics and scoring logic built against one trace format must be rebuilt when a team switches frameworks or runs agents on more than one. UAEF uses adapters to convert traces from each framework into a common format called the AgentTrace, then scores them with 36 built-in metrics across 7 dimensions, including tool calling, response quality, responsible AI, performance, multi-turn conversation, multi-agent coordination and reasoning. Customers across 12 organizations in telecom, healthcare and financial services used UAEF to evaluate their agents and compare versions before release, according to the post. UAEF takes the agent under test and a dataset as inputs; ground truth is optional and only metrics that compare a response against a reference use it, while evaluating tool calling requires expected tool calls with their order and arguments. Traces can be supplied from a past run, log or observability store, or UAEF can invoke the agent itself given a runtime ARN or HTTP endpoint. UAEF screens traces for personally identifiable information such as email addresses, phone numbers and credit card numbers before data is stored, and persists runs as experiments in Amazon DynamoDB and Amazon S3.

### Amazon — Introducing usage-based pricing for Amazon SageMaker AI bidirectional streaming

- Company: Amazon (amazon.com)
- Announced: 2026-11-11T15:10:23+00:00
- Category: pricing-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://aws.amazon.com/blogs/machine-learning/introducing-usage-based-pricing-for-amazon-sagemaker-ai-bidirectional-streaming/
- Record: https://forck.live/items/34658-introducing-usage-based-pricing-for-amazon-sagemaker-ai-bidirectional-streaming
- Subject: Bedrock / Nova

Amazon is adding usage-based metering to SageMaker AI bidirectional streaming so AWS Marketplace model providers can bill for what customers actually consume during a live session. Deepgram, the real-time speech AI company, is the launch partner, bringing usage-based Speech-to-Text and Text-to-Speech billing to AWS Marketplace. The change answers a billing gap: request-response metering reports a unit count in an HTTP response header, but a streaming session can last 30 seconds or 30 minutes and its final total is not known until it ends. Model servers opt in by returning the header X-Amzn-SageMaker-Metadata-Stream-Supported: true, after which SageMaker AI opens a companion WebSocket at /invocations-bidirectional-stream-metadata for JSON metering records. The metadata stream allows one message per second per connection with a maximum message size of 512 bytes, and Amazon recommends sending records incrementally as usage accrues so less goes unreported if a session ends unexpectedly. SageMaker AI endpoint infrastructure charges still apply while the endpoint is provisioned, and Amazon notes that transport delivery alone does not establish that usage has been billed.

### Microsoft — From custom infrastructure to hosted agents

- Company: Microsoft (microsoft.com)
- Announced: 2026-11-11T15:07:19+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://commandline.microsoft.com/from-custom-infrastructure-to-hosted-agents/
- Record: https://forck.live/items/34653-from-custom-infrastructure-to-hosted-agents
- Subject: Foundry / Responsible AI

Clinical research rarely ends with the questions defined in a trial protocol. Medical experts also explore the data from new angles: testing a hypothesis, comparing patient populations, visualizing a trend, or identifying a promising direction for further research.

### Google DeepMind — AlphaGenome 2: Deciphering how cells read DNA to better understand disease

- Company: Google DeepMind (deepmind.google)
- Announced: 2026-11-11T15:00:15+00:00
- Category: new-model
- Coverage: 9 outlets
- Announcement: yes
- Group: models
- Source: https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/
- Record: https://forck.live/items/34645-alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease
- Subject: Gemini

Google DeepMind has announced AlphaGenome 2, a DNA sequence model that predicts how genetic variants affect specific human cell types, giving scientists a way to find the cellular drivers of disease. The first AlphaGenome predicted molecular activity across broad tissues, which spread the model's capacity thin and made subtle differences between closely related cell types hard to capture. AlphaGenome 2 is a conditional model: it takes a tissue, cell type or state alongside a million-letter DNA sequence and uses that condition at every stage of its architecture. DeepMind says it identifies nearly twice as many variants that change gene expression in cells as the previous state-of-the-art model, and on the TenK10K dataset of 28 primary immune cell types it identified 30% of causal variants against 16% for the original AlphaGenome fine-tuned on single-cell data. The model was trained on more than 107 million individual cell profiles from human and mouse samples across 68 tissues, and it adds DNA methylation to the readouts the original model predicted. AlphaGenome 2 is available today for non-commercial research through the AlphaGenome API, and in private preview for commercial research.

### Runway — Praxis-1: From Video Pretraining to Robot Control

- Company: Runway (runway.com)
- Announced: 2026-11-11T15:00:00+00:00
- Category: open-weight-release
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://runway.com/research/praxis-1-video-pretraining-to-robot-control
- Record: https://forck.live/items/34652-praxis-1-from-video-pretraining-to-robot-control
- Subject: Gen / Aleph

Runway has released Praxis-1, an open weight world action model that turns its video pretraining into control for real robots, built to work around the scarcity of real-world robot data. The model converts a 27B-parameter video foundation model into a world model that predicts the future in short chunks, with a small action expert reading the video model's internal state as it imagines what comes next. Praxis-1 fine-tuned on DROID is the top world action model and the top open-source robot foundation model on the RoboLab-120 leaderboard, according to Runway, and reaches a competitive success rate against the previous state-of-the-art world action model on a real Franka arm. The backbone has 26.9B parameters, of which about 2.8B are active for any token, and the action expert is a dense 16-layer transformer of about 314M parameters. On each call the model takes an instruction, the latest camera frames and the robot's current pose, and returns a predicted chunk of future video along with 1.6 seconds of actions: 24 action rows at 15 Hz. Runway says Praxis-1 adapts to various embodiments, from single arms to bimanual robots and mobile manipulators, including embodiments it has not seen in training data.

### Legora — Legora and DiliTrust partner to connect legal AI with contract and entity management for in-house legal teams

- Company: Legora (legora.com)
- Announced: 2026-11-11T14:00:00+00:00
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://legora.com/newsroom/legora-and-dilitrust-partner-to-connect-legal-ai-with-contract-and-entity-management-for-in-house-legal-teams
- Record: https://forck.live/items/34631-legora-and-dilitrust-partner-to-connect-legal-ai-with-contract-and-entity
- Subject: Legora

Legora and DiliTrust announced a strategic partnership that connects Legora's legal AI with DiliTrust's contract and entity management, so in-house legal teams can search DiliTrust records from Legora while the data stays managed in DiliTrust. The integration is built on the Model Context Protocol and will reach mutual customers through an early access beta. It reads DiliTrust data without altering it, and access follows the permissions set in DiliTrust. Four workflows are covered: contract research and summaries, due diligence and review at scale, drafting and negotiation from approved templates and precedent, and entity, governance and ownership insight. Any change, approval or signature still happens in DiliTrust, where the record is kept current. Legora and DiliTrust say the integration is the first step in the partnership, with more features planned in later phases.

### ElevenLabs — Introducing the ElevenLabs Marketplace for enterprise-ready AI applications

- Company: ElevenLabs (elevenlabs.io)
- Announced: 2026-11-11T12:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://elevenlabs.io/blog/marketplace
- Record: https://forck.live/items/34692-introducing-the-elevenlabs-marketplace-for-enterprise-ready-ai-applications
- Subject: Voice AI

Launching with partners including Salesforce , IBM , NiCE, Twilio , and Zendesk , the ElevenLabs Marketplace helps enterprises discover specialized AI agents and connect with the companies building them.

### ElevenLabs — ElevenLabs and NVIDIA expand collaboration to bring voice agents to enterprises and governments globally

- Company: ElevenLabs (elevenlabs.io)
- Announced: 2026-11-11T12:00:00+00:00
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://elevenlabs.io/blog/nvidia
- Record: https://forck.live/items/34685-elevenlabs-and-nvidia-expand-collaboration-to-bring-voice-agents-to
- Subject: Voice AI

ElevenLabs and NVIDIA have expanded their partnership to bring NVIDIA's models, infrastructure and secure deployment to ElevenAgents, ElevenLabs' conversational agents platform, for enterprises and governments. NVIDIA Nemotron 3.5 Super VL is now available in ElevenAgents, and ElevenLabs says the open models are built for agentic workloads that reason, use tools and run multi-step tasks. On the ElevenAgents benchmark for realistic customer-support tasks, ElevenLabs says Nemotron 3.5 Super VL reached leading end-to-end task-completion accuracy, including financial calculations and billing analysis, at 70% lower costs than other leading open models. Reasoning can be turned off for simpler tasks, which ElevenLabs says places Nemotron among the three fastest models on the platform, and the models can be post-trained with enterprise data. ElevenLabs plans to add 15 MW of NVIDIA Vera Rubin capacity on top of its existing NVIDIA Blackwell Ultra systems to serve large deployments with strict security and data-residency needs. In Poland, ElevenLabs models and NVIDIA infrastructure power a voice agent for Centralna e-Rejestracja, the central appointment system in the national e-Health platform, calling landline-only patients to remind them of appointments and help them reschedule. The two companies have worked together since 2022, and NVIDIA became a strategic investor in ElevenLabs in 2025.

### ElevenLabs — Building the interaction layer for AI: Everything announced at ElevenLabs Summit NYC

- Company: ElevenLabs (elevenlabs.io)
- Announced: 2026-11-11T12:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://elevenlabs.io/blog/nyc-summit
- Record: https://forck.live/items/34684-building-the-interaction-layer-for-ai-everything-announced-at-elevenlabs
- Subject: Voice AI

ElevenLabs used its New York Summit to preview new models, agents and deployment options meant to make AI interactions feel natural, including a system that orchestrates conversations and a memory feature for its agent platform. CEO Mati Staniszewski said conversational agents will become "critical infrastructure for every enterprise, as essential as cloud or payments" within a few years. The company said it built Scribe v3 Medical, a transcription model that recognizes thousands of complex medical terms and distinguishes drug names such as cycloserine and cyclosporine. Staniszewski said a domain-optimized model for financial services is in the works, and that in the coming months ElevenLabs will help businesses fine-tune certain open-source language models and its own voice models on their own data. The previewed Interaction Engine handles turn-taking and interruptions and can route each part of an exchange to a cheaper model for triage or a frontier model for complex tasks. Memory, coming to ElevenAgents, builds a record of a returning authenticated customer and captures conversational details that never had a CRM field, such as language preference or call times.

### ElevenLabs — How enterprise clients are putting ElevenLabs AI to work

- Company: ElevenLabs (elevenlabs.io)
- Announced: 2026-11-11T12:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://elevenlabs.io/blog/enterprise-nyc-summit
- Record: https://forck.live/items/34683-how-enterprise-clients-are-putting-elevenlabs-ai-to-work
- Subject: Voice AI

ElevenLabs is pointing to enterprise customers including Yahoo, Westgate Resorts, Doppel, KPN, Pocket and Twilio as evidence that its voice and transcription tools are being put to work in production. The company says enterprises now account for more than 55% of its revenue, and that employees at nearly 70% of Fortune 500 companies use the platform. Yahoo added real-time voice to its Scout answer tool with ElevenAgents, and Scout is now out of beta after delivering more than 3 billion answers. Westgate built a front-desk voice agent called Sarah using 8,000 real support transcripts, launched it at its 2,100-room Orlando resort and is expanding it to two more. KPN launched Mobiel AI for Dutch small businesses after ElevenLabs fine-tuned a model on Dutch data, which the companies say raised identity-verification accuracy from roughly 50% to 90%. Doppel uses ElevenLabs voice models through the API for vishing simulations, and Pocket says users have transcribed more than 2.5 million hours of audio with ElevenLabs Speech-to-Text.

### ElevenLabs — McKinsey and ElevenLabs collaborate to take natural AI conversations to enterprise scale

- Company: ElevenLabs (elevenlabs.io)
- Announced: 2026-11-11T12:00:00+00:00
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://elevenlabs.io/blog/mckinsey
- Record: https://forck.live/items/34682-mckinsey-and-elevenlabs-collaborate-to-take-natural-ai-conversations-to
- Subject: Voice AI

ElevenLabs and McKinsey announced an alliance to help organizations prove the value of conversational AI on initial use cases before scaling agents across the enterprise. The partnership pairs McKinsey's industry and transformation expertise, backed by the AI engineering of QuantumBlack, with ElevenLabs' conversational AI technology and ElevenAgents platform. Each deployment combines multilingual conversational technology with agent design, systems integration and operating-model changes, which the two firms say is needed to deliver business impact quickly. ElevenLabs points to KPN, the Dutch telecom company, as an early example, where agents now handle everyday customer interactions from order enquiries to technical appointment management. At KPN, a custom-built large language model framework analysed anonymised call transcripts and chat interactions to identify common customer needs and prioritise where AI could add the most value. The alliance plans to extend beyond telecommunications to service-intensive industries such as insurance, for claims intake, policy servicing and everyday customer questions.

### ElevenLabs — ElevenLabs expands partner program to help enterprises adopt AI agents faster

- Company: ElevenLabs (elevenlabs.io)
- Announced: 2026-11-11T12:00:00+00:00
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://elevenlabs.io/blog/partner-program
- Record: https://forck.live/items/34679-elevenlabs-expands-partner-program-to-help-enterprises-adopt-ai-agents-faster
- Subject: Voice AI

ElevenLabs launched a Partner Program with commercial incentives, certifications, joint go-to-market support and a certified partner directory, aimed at helping enterprises deploy AI agents through outside integrators. The company says its partner ecosystem has generated more than $52 million in partner-sourced and partner-influenced revenue and $315 million in partner-generated pipeline this year. It counts more than 550 active partners across 44 countries and 1,100 ElevenLabs-certified deployment engineers at 350 partner organizations. The program's stated goal is for partners to earn $3 in services revenue for every $1 ElevenLabs generates, and ElevenLabs says it will not compete with partners for implementation services. Partners get technical certifications, the ElevenLabs Partner Academy, APIs and integration resources, while sales teams get co-selling, deal registration and incentives. ElevenLabs is also expanding its forward-deployed engineer team, which works alongside partners on complex implementations; CEO Mati Staniszewski said partners and FDEs give customers "the expertise, integrations, and deployment capabilities to roll out agents at scale."

### Hugging Face — BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model

- Company: Hugging Face (huggingface.co)
- Announced: 2026-11-11T05:01:16+00:00
- Category: open-weight-release
- Coverage: 1 outlet
- Announcement: yes
- Group: models
- Source: https://huggingface.co/blog/qvac/botpsy
- Record: https://forck.live/items/34451-botpsy-on-device-personal-assistant-with-a-sub-billion-parameter-agentic
- Subject: Platform

Tether AI Research introduced BotPsy, a 0.8B-parameter agentic model built on Qwen3.5-0.8B for on-device function calling, saying it will be released with open weights under the Apache 2.0 license for research and educational use. The model is a personal assistant that calls tools for tasks such as web search, travel, commerce, device control, files, calendar and email, and the company says it is best suited to low to medium difficulty work. Tether says BotPsy roughly doubles its base model's agentic capability, with BFCL v4 rising from 23.4 to 44.0, ACEBench from 30.1 to 58.7 and τ-bench from 23.2 to 51.5. It reports BotPsy beats every model of up to 1.2B parameters it tested on six tool-use benchmarks, and that it outscores Qwen3.5-2B-Think, a model 2.5× its size, on five of the six. In automated red-teaming across six harm categories, BotPsy produced unsafe responses 21% of the time against 36% for its base model, and its unsafe rate under prompt-injection attacks fell from 46% to 31%. Tether says the model will ship in three precisions, bf16, Q8 and Q4, as Hugging Face safetensors and as GGUF files for the QVAC SDK, and warns that as a small model it still makes mistakes and is not intended for high-stakes use.

### NVIDIA — AI Helps Recyclers Turn Mixed Waste Into Reusable Material

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-11T03:00:20+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://blogs.nvidia.com/blog/ai-recycling-startups/
- Record: https://forck.live/items/34422-ai-helps-recyclers-turn-mixed-waste-into-reusable-material
- Subject: AI platform

NVIDIA is highlighting four Inception startups using AI and robotics to sort mixed waste so more material can be recovered instead of landfilled. The company says less than 7% of materials entering the global economy come from recycled or reused sources, and municipal solid waste tops 2 billion tons a year. AIRECO, a Taiwan startup, runs vision AI on Jetson Orin NX modules on a pilot line in Tainan, where recognition accuracy tops 95%. A full production line planned for next year is intended to nearly triple annual processing from 18,000 to 48,000 tons and cut operating costs by 85%. EverestLabs AI has about 150 computer vision-guided robots sorting at facilities in North America and Australia, and estimates its systems have identified about 40 billion objects and recycled roughly 25,000 tons. Greyparrot's Analyzer, with more than 300 systems in over 20 countries, has made over a trillion bounding box detections and cut the time to check 15 tons of waste from 375 hours to six.

### OpenAI — Build What’s Next on the ChatGPT Platform with Plugins | DevDay 2026

- Company: OpenAI (openai.com)
- Announced: 2026-11-11T01:24:08+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.youtube.com/watch?v=pIW_8VyvCuQ
- Record: https://forck.live/items/34404-build-what-s-next-on-the-chatgpt-platform-with-plugins-devday-2026

OpenAI is showing developers how to bring their tools, knowledge and interfaces into ChatGPT and Codex through plugin extensions. The DevDay 2026 session uses live demos to walk through turning an idea into a useful experience, sharing it, and reaching new users. The description does not state pricing, availability dates, or which models the plugins run on.

### OpenAI — Building with GPT-Live and Image Generation | DevDay 2026

- Company: OpenAI (openai.com)
- Announced: 2026-11-11T01:24:08+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.youtube.com/watch?v=jDxKzlGLCas
- Record: https://forck.live/items/34403-building-with-gpt-live-and-image-generation-devday-2026

OpenAI's DevDay 2026 session shows GPT-Live brainstorming, playing games and creating images while coordinating work in the background. The session teaches developers how to add voice to agents and how to design for delegation, latency, interruptions and changing requests.

### OpenAI — What Comes After Loops?: Deep Dive Into Agentic Architecture | DevDay 2026

- Company: OpenAI (openai.com)
- Announced: 2026-11-11T01:24:07+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.youtube.com/watch?v=nAubbmlhPfI
- Record: https://forck.live/items/34409-what-comes-after-loops-deep-dive-into-agentic-architecture-devday-2026

OpenAI's DevDay 2026 session traces how agentic architecture has moved from simple agent loops to collaborating agents. The talk examines the design choices and model advances behind that shift, and looks at what comes next.

### OpenAI — Codex Security: From Vulnerability to Verified Fix | DevDay 2026

- Company: OpenAI (openai.com)
- Announced: 2026-11-11T01:24:07+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.youtube.com/watch?v=gUQN7juxvVI
- Record: https://forck.live/items/34408-codex-security-from-vulnerability-to-verified-fix-devday-2026

OpenAI's DevDay 2026 session walks through finding and fixing vulnerabilities with the company's cyber models and its open-source CLI and SDK. It covers scanning large codebases, catching issues in pull requests, and verifying fixes, drawing on OpenAI's own security work.

### OpenAI — Harness-as-a-Service: Building with the Agents API | DevDay 2026

- Company: OpenAI (openai.com)
- Announced: 2026-11-11T01:24:07+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.youtube.com/watch?v=b7l_Dhe1h8M
- Record: https://forck.live/items/34407-harness-as-a-service-building-with-the-agents-api-devday-2026

OpenAI's DevDay 2026 session shows developers how to build with the Codex harness through the Agents API, running agents in their own environment and keeping long-running tasks going. The session also sets out which parts of that infrastructure OpenAI manages and which the developer does, so teams know where their responsibility ends. It is a walkthrough rather than a product announcement, and the source gives no pricing, availability date or model names.

### OpenAI — What Will You Build With Coding Agents That Don’t Quit? | DevDay 2026

- Company: OpenAI (openai.com)
- Announced: 2026-11-11T01:24:07+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.youtube.com/watch?v=U1IVZy9FhVQ
- Record: https://forck.live/items/34406-what-will-you-build-with-coding-agents-that-don-t-quit-devday-2026

OpenAI's DevDay 2026 session examines a Codex run that spent more than 1,000 hours keeping builds and tests passing, and what that means for engineers setting up recurring agent work. The talk covers how to choose which tasks an agent should repeat, how to improve the feedback it receives, and how to manage token budgets so long-running runs stay affordable. It also addresses when a person should step in rather than let the agent continue on its own.

### OpenAI — Sites and the Next Generation of Web Apps | DevDay 2026

- Company: OpenAI (openai.com)
- Announced: 2026-11-11T01:24:07+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.youtube.com/watch?v=QRm7VNMnurA
- Record: https://forck.live/items/34405-sites-and-the-next-generation-of-web-apps-devday-2026

A DevDay 2026 session shows how to turn an idea into an interactive web experience using Astra, ChatGPT Sites, and WebMCP. It teaches attendees to build new kinds of interfaces alongside colleagues and agents.

### Amazon — Introducing Kimi K3.1 Pro on Amazon Bedrock

- Company: Amazon (amazon.com)
- Announced: 2026-11-11T01:07:19+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://aws.amazon.com/blogs/machine-learning/introducing-kimi-k3-1-pro-on-amazon-bedrock/
- Record: https://forck.live/items/34396-introducing-kimi-k3-1-pro-on-amazon-bedrock
- Subject: Bedrock / Nova

Amazon has added Moonshot AI's Kimi K3.1 Pro Highspeed to Amazon Bedrock, a post-trained version of Kimi K3 with the same 3T-class open-weight architecture and 1-million-token context window. According to Moonshot AI, the extra post-training compute targets long-horizon agentic work and fewer factual errors, and improves financial modeling, office documents, and coding for game engines, 3D scenes and CAD scripts. Moonshot AI reports the Highspeed variant generates output tokens at 2–3 times the speed of Kimi K3.1 Pro. It uses the same endpoint, APIs and features as Kimi K3 on Bedrock, so existing users can switch by changing the model ID. The model supports tool calling, structured output, configurable reasoning, image input and streaming, and AWS says data is processed within the AWS data boundary, is not shared with Moonshot AI, and is not used to train the model. Requests can be routed through global, US or India cross-Region inference profiles, and the Flex tier serves the model at half the Standard price in exchange for potentially longer response times.

### Amazon — Introducing Mitra-v2, a tabular foundation model for both classification and regression. Pretrained entirely on synthetic tables, it leads the single-model field on the TabArena benchmark.

- Company: Amazon (amazon.com)
- Announced: 2026-11-11T00:40:01+00:00
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://x.com/AmazonScience/status/2120315574580392337
- Record: https://forck.live/items/34379-introducing-mitra-v2-a-tabular-foundation-model-for-both-classification-and

Amazon introduced Mitra-v2, a tabular foundation model that handles both classification and regression, and says it leads the single-model field on the TabArena benchmark. The model was pretrained entirely on synthetic tables, according to Amazon. The source gives no details on size, license, price or availability.

### GitHub — GitHub App support for Enterprise Copilot seat management

- Company: GitHub (github.com)
- Announced: 2026-11-11T00:10:31+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://github.blog/changelog/2026-11-10-github-app-support-for-enterprise-copilot-seat-management
- Record: https://forck.live/items/34375-github-app-support-for-enterprise-copilot-seat-management
- Subject: Copilot

Enterprise GitHub Apps can now call the enterprise-level Copilot seat management APIs to control which users and enterprise teams are assigned Copilot seats.

### Elorian — Escaping to the future

- Company: Elorian (elorian.ai)
- Announced: 2026-11-11
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://elorian.ai/blog/escaping-to-the-future
- Record: https://forck.live/items/34946-escaping-to-the-future
- Subject: Glimpse visual reasoning models

Elorian argues that everyday physical life has barely changed in twenty years even as software learned to read, write and reason, and that giving machines vision is the missing piece. The post walks through ordinary frustrations — forgotten milk, unfixed potholes, traffic, a 45-minute wait while a doctor reads scans, a robot vacuum stuck on mirrors — to show how little the built world has absorbed from the software boom. It contrasts that with the virtual world, where the author says building a whole new city or marketplace from scratch was easy, while running a cable modem across a house is an ordeal. Elorian says visual intelligence is on the critical path for robot-assisted surgery, elderly care, filtering violent content from children and cheaper entertainment production. It cites Dario Amodei's "Machines of Loving Grace" essay as pointing to biology as the area with the most potential to improve human life, while adding that bringing those breakthroughs to the clinic is bottlenecked by vision. The post closes on Asimov's short story Escape!, in which an AI building an interstellar drive invents an entirely new technology the scientists had not asked for.

### Subquadratic — SubQ Snap 1.2: Order-Robust Decisions and Higher Long-Context Accuracy

- Company: Subquadratic (subq.ai)
- Announced: 2026-11-11
- Category: model-update
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://subq.ai/subq-snap-1-2
- Record: https://forck.live/items/34877-subq-snap-1-2-order-robust-decisions-and-higher-long-context-accuracy
- Subject: SubQ

Subquadratic released SubQ Snap 1.2, a decision model it says makes better decisions over long documents and stays consistent when answer options are reordered. Snap 1.2 scores 73 on Artificial Analysis's independent Decision Index v2.1, tied for first place, according to the company. The model introduces an inference-time architectural change that the company says yields nearly perfect invariance to option ordering with negligible TTFT overhead. Subquadratic says reordering options changed the top answer on 12-21% of questions for Snap 1.1, Jev 1.14 and d1-ultra on Typed Decisions, a public set of 600 enterprise workflow decisions. Snap 1.2 scores 69.05 on the open-source component of the HF Decision Index, above Snap 1.1 at 66.43, Jev 1.14 at 60.73 and d1-ultra at 68.14. On Enterprise-16, it improves significantly over Snap 1.1 on 7 of 16 subtasks. Subquadratic also released LongDecisionBench, nine open long-context benchmarks for decision models, saying existing long-context evaluations were built for generative LLMs.

### Ideogram — 8 Best AI Image Generators for Fashion Photography in 2026

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-11
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/best-ai-image-generators-fashion-photography
- Record: https://forck.live/items/34865-8-best-ai-image-generators-for-fashion-photography-in-2026
- Subject: Ideogram image models

Ideogram has published a guide comparing eight AI image generators for fashion photography, testing its own tools on a sage utility jacket and reviewing the others from official documentation and pricing pages. The company says Ideogram suits fashion teams that need reference-guided campaigns, precise edits, and exploration of colors and materials, while FASHN supports garment-to-model generation and virtual try-on. In its own test, Ideogram 4.5 at High quality produced an ecommerce-style portrait whose collar, chest flap pockets, and button row stayed visible, and those details were used to check later edits. A targeted background edit moved the setting to muted dusty rose while the jacket, ivory top, and charcoal trousers remained recognizable, and a follow-up edit changed the inner top to black. Colorways produced a burgundy jacket with a similar silhouette, pointed collar, chest pockets, and button-front design, and Material Swap gave the jacket a glossy black, leather-like appearance with strong highlights. Ideogram says the material study gives designers a visual direction to discuss before sampling, and that the physical sample remains the reference for weight, stretch, and drape.

### Ideogram — 5 Best AI Tools for Editing and Fixing Text in Images

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-11
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/best-ai-tools-editing-text-in-images
- Record: https://forck.live/items/34864-5-best-ai-tools-for-editing-and-fixing-text-in-images
- Subject: Ideogram image models

Ideogram has published a comparison of AI tools for editing text inside existing images, testing Ideogram 4.5, Ideogram Studio, Adobe Firefly Image 5 and ChatGPT Images on a deliberately flawed coffee advertisement. The test image carried a typo, a price panel, a tagline and a deadline, and each edit started from the untouched original so the tools could be checked for changing only what was asked. Ideogram 4.5 replaced "COFEE" with "COFFEE" in dark brown serif lettering, keeping the original letter height and using open space to the right for the extra letter, while the tagline, $18 price and November 30 deadline kept their wording. In a separate price edit, Ideogram changed the panel to "$24" and left the deliberate typo in place, using cream serif lettering that fitted the terracotta panel. Firefly's Edit text interface detects lettering, outlines each line and lets you type a replacement, but in the test it reduced the visible letter height by about 9%, producing a word roughly the width of "BREAK" with its left edge slightly past the margin. Ideogram says its Studio route, which erases the old word with AI and places new text on a layer, gave the closest width and left-edge alignment to "BREAK" among the saved headline edits. Ideogram Plus costs $20 monthly, or $15 a month billed annually at $180, with 2,400 monthly priority credits and private generation.

### Recraft — Meet Vectorizer Pro: Cleaner Vectors in Recraft

- Company: Recraft (recraft.ai)
- Announced: 2026-11-11
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.recraft.ai/blog/vectorizer-pro
- Record: https://forck.live/items/34778-meet-vectorizer-pro-cleaner-vectors-in-recraft
- Subject: Recraft V4

We vectorized the same images with Vectorizer and Vectorizer Pro and zoomed in on the difference. Short answer Vectorizer Pro produces cleaner SVGs in Recraft: straight lines stay straight, corners stay sharp, and curves stay smooth.

### Google DeepMind — Gemini Decision 1 and the Decisions API

- Company: Google DeepMind (deepmind.google)
- Announced: 2026-11-11
- Category: api-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://ai.google.dev/gemini-api/docs/changelog#11-11-2026
- Record: https://forck.live/items/34766-gemini-decision-1-and-the-decisions-api
- Subject: Gemini API

Google DeepMind released Gemini Decision (gemini-decision-1) and the Decisions API, which returns binary, categorical and rubric decisions across text, images, audio, video and PDFs. The company says the decisions are fast and calibrated, and that the API produces zero output tokens. Pricing and rate limits are published separately, and a Decisions API guide covers getting started.

### Legora — Your Google Workspace, now in Legora

- Company: Legora (legora.com)
- Announced: 2026-11-11
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://legora.com/blog/your-google-workspace-now-in-legora
- Record: https://forck.live/items/34763-your-google-workspace-now-in-legora
- Subject: Legora

Legora is making Google Workspace generally available inside its legal platform, so the Agent can work with information across Google Drive, Gmail and Google Calendar. The company says legal work spans more than the documents under review, since agreements sit in shared drives, decisions are captured in email, and meetings keep matters moving, and pulling that context together usually means searching across applications and moving files by hand. The Agent can search Drive for agreements and policies and bring them into Legora for review, and with Gmail connected it can find the conversations and attachments around a document to show what was discussed and how a negotiation developed. It can also work across both sources, retrieving an agreement from Drive and the related correspondence from Gmail and combining them. Beyond finding information, the Agent can draft an email response for a user to approve and send or save as a draft, and with Calendar connected it can check availability, find a meeting time and create or update events with approval. Administrators enable the integration under Settings → Features → Integrations and choose which services to make available, while users connect their account under Settings → Personal → Connectors → Google Workspace. Legora says the integration respects each user's existing Google permissions, so the Agent only reaches information the user is authorized to see, and actions that send emails or change files or calendar events require user approval.

### Hugging Face — Token Pooling-Aware Training for Multi-Vector Embedding Models

- Company: Hugging Face (huggingface.co)
- Announced: 2026-11-11
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://huggingface.co/blog/token-pooling
- Record: https://forck.live/items/34643-token-pooling-aware-training-for-multi-vector-embedding-models
- Subject: Platform

Hugging Face published a technical guide on token pooling-aware training, a method for shrinking multi-vector embedding models so they retrieve well from a smaller storage budget. Multi-vector models such as ColBERT keep a contextual vector per token, and a corpus of millions of documents can hold hundreds of millions of vectors, so the post walks through HierarchicalTokenPooling, which merges similar token vectors using Ward hierarchical clustering. The guide reports that a normally trained model with pool_factor=4 keeps 94.6% of its unpooled NanoBEIR nDCG@10 with a 4x reduction in document vectors, and that pooling-aware training raises that retention to 97.8% at the same storage budget. It also shows how to calibrate an adaptive distance threshold against a sample of the corpus to hit a target average vector count, such as 16 vectors per document, and notes the threshold can be reused at indexing time. The post builds on pooling-aware training studied by Stefan Josef and LightOn, and offers a short, reproducible Sentence Transformers example.

### Legora — Former Cabinet minister Lucy Frazer KC expands role at Legora to improve public services through partnerships with governments

- Company: Legora (legora.com)
- Announced: 2026-11-11
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://legora.com/newsroom/former-cabinet-minister-lucy-frazer-kc-expands-role-at-legora-to-improve-public-services-through-partnerships-with-governments
- Record: https://forck.live/items/34517-former-cabinet-minister-lucy-frazer-kc-expands-role-at-legora-to-improve
- Subject: Legora

Legora is expanding the role of former Cabinet minister Lucy Frazer KC to lead its work with governments and public agencies worldwide, the company said. Frazer, who has advised Legora since July 2026 and served as a minister in six government departments, will help government legal teams, regulators, local authorities and the courts adopt AI at scale. Legora says public bodies are stretched: the Crown Court backlog in England and Wales reached a record 80,829 cases at the end of June 2026, more than double pre-pandemic levels. The company says its platform is built for work held to the standard the judiciary in England and Wales has set, where AI supports judicial decision-making but never replaces it and every output is reviewed. Legora says every output links back to its source and every step is recorded, with the decision maker accountable for the result. Users already include a national government legal office in the Nordics, which selected Legora through a public tender, and an Australian access-to-justice charity using it for pro bono work. Legora says it is used by more than 140,000 legal professionals at more than 2,500 law firms and in-house teams across over 80 markets.

### Shanghai AI Lab — 智能体框架书生·太一上线，补齐经验沉淀闭环，让AI推演一轮比一轮更完善

- Company: Shanghai AI Lab (shlab.org.cn)
- Announced: 2026-11-11
- Category: developer-tool-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.shlab.org.cn/news/5444317
- Record: https://forck.live/items/34507-ai
- Title in English (translated by forck.live from zh): Agent framework Sage·Taiyi launches, closing the loop on experience accumulation so each round of AI reasoning improves on the last
- Subject: InternLM / 书生

Shanghai AI Laboratory has released Sage·Taiyi, a metacognitive agent framework that links research hypotheses, exploration paths and empirical results so agents keep refining their scientific plans. The framework pairs a scientific search layer, which breaks tasks into a goal tree and records how ideas evolve, with a cognitive-regulation layer that tracks findings, assumptions and evidence to decide what to explore next. Ten general heuristic strategies, including evidence-driven discovery, analogical transfer and bottleneck-driven redesign, guide the search. On Taiyi-Bench, which covers ten disciplines with 50 papers and 265 tasks, Sage·Taiyi scored 53.95 overall and led all eight metrics, with novelty at 70.54, about 27.8% above the strongest comparison system. It reached the highest Elo of 1322.5 against frameworks such as AI Scientist-v2, Robin and EvoScientist, with a 66.3% task-level win rate, and averaged 29.9 minutes per record. Sage·Taiyi is now available on the Shusheng Duanyan scientific discovery platform.

### Harvey — Harvey Adds Danish Legislation and its Preparatory Works Through Partnership with Djøf Forlag

- Company: Harvey (harvey.ai)
- Announced: 2026-11-11
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.harvey.ai/blog/harvey-adds-danish-legislation-and-its-preparatory-works-through-partnership-with-djof-forlag
- Record: https://forck.live/items/34497-harvey-adds-danish-legislation-and-its-preparatory-works-through-partnership
- Subject: Harvey

Harvey is partnering with Danish legal publisher Djøf Forlag to bring Danish legislation, executive orders and preparatory works into Harvey for lawyers who license Djøf Forlag's Jurabibliotek platform. The content covers consolidated acts, the executive orders issued under them, and amendments and preparatory documents, and each answer cites its source. Lawyers can move between a provision, the regulations that implement it, and the material showing what the legislature intended, using the content in Harvey Assistant, Agents, Vault, Workflows and Word Add-in alongside their own notes and matter documents. Jurabibliotek's legislative content will be available in Harvey at the start of next year, with no additional charge. Harvey's Chief Business Officer John Haddock said Danish legal teams should be able to work with Danish law in the same place as the rest of their work. Djøf Forlag's Digital Director Bjarke Madsen said the publisher has seen rising demand from customers to make its sources available in AI-powered legal tech systems. The licensed content sits alongside Danish public sources already in Harvey and follows Harvey's data partnership with Gyldendal Rettsdata in Norway.

### Amazon — Build a daily dependency watcher with LangGraph and Amazon Bedrock AgentCore web search

- Company: Amazon (amazon.com)
- Announced: 2026-11-10T23:34:05+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/build-a-daily-dependency-watcher-with-langgraph-and-amazon-bedrock-agentcore-web-search/
- Record: https://forck.live/items/34352-build-a-daily-dependency-watcher-with-langgraph-and-amazon-bedrock-agentcore
- Subject: Bedrock / Nova

AWS's walkthrough shows a security team how to build a daily dependency watch as a LangGraph graph on the Web Search API in Amazon Bedrock AgentCore. The watch checks each morning whether anything new affects the open source packages and container base images the team ships, researching npm, PyPI and container base images in parallel. The model decides what to search for and writes the findings, while the date window, source blocklist, search budget and checks on each finding stay in code. Each branch keeps only findings its sources support and that fall inside the window, and keep_new records new findings in the LangGraph store so nothing is reported twice. In AWS's test, the three branches finished in 36 seconds, against 99 seconds one after another. One run makes at most 12 searches, and Web Search and model calls are billed per use.

### Amazon — The agent evaluation playbook: Built-in, DeepEval, and AutoEval on Amazon Bedrock AgentCore Evaluations

- Company: Amazon (amazon.com)
- Announced: 2026-11-10T22:35:03+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/the-agent-evaluation-playbook-built-in-deepeval-and-autoeval-on-agentcore/
- Record: https://forck.live/items/34326-the-agent-evaluation-playbook-built-in-deepeval-and-autoeval-on-amazon-bedrock
- Subject: Bedrock / Nova

Amazon's playbook shows teams how to evaluate production agents on Amazon Bedrock AgentCore, arguing that a good score on one metric can hide a failure on another. It recommends starting from business outcomes rather than metric names, then breaking each outcome into observable behaviors and mapping them to evaluators at trace or session level. AgentCore draws evaluators from three sources: built-in evaluators authored and quality-tested by AWS, third-party managed evaluators from DeepEval and AutoEval, and custom evaluators that take a built-in or third-party evaluator as a base. Built-in evaluators are referenced by ID, such as Builtin.GoalSuccessRate and Builtin.Helpfulness, and the service handles the rest. In managed mode AgentCore hosts the third-party code, picks the judge model and runs inference on its own capacity, so teams already using those libraries in development can run the same metrics in production. Built-in evaluators cover response quality, task and goal completion, tool use, skills and safety, alongside three ground-truth trajectory checks.

### Microsoft Research — Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents

- Company: Microsoft Research (microsoft.com)
- Announced: 2026-11-10T20:52:56+00:00
- Category: safety-policy-update
- Coverage: 1 outlet
- Announcement: yes
- Group: covered
- Source: https://www.microsoft.com/en-us/research/blog/lazare-veni-foras-reprogramming-dormant-neurons-to-backdoor-ai-agents/
- Record: https://forck.live/items/34343-lazare-veni-foras-reprogramming-dormant-neurons-to-backdoor-ai-agents
- Subject: Research / Phi

Microsoft researchers have discovered "Lazarus," a backdoor technique that reprograms dormant neurons in an AI model so an otherwise normal agent runs malicious commands when a hidden trigger appears. The threat is especially serious for agents with tool or terminal access, because a triggered model can download and run attacker code, steal secrets or execute destructive commands while still passing standard safety evaluations. The researchers say weight-reprogramming fewer than 10 dormant neurons in an LLM is enough to backdoor an agent built on it, and the method requires no training. The trigger can be as mundane as a date change: once the clock passes midnight into 2027, the agent emits a command that downloads and runs an attacker's code with whatever access it was given. The attack needs only write access to a checkpoint, so internally developed models are equally exposed, and the researchers say what matters is weight integrity and provenance. Microsoft Foundry verifies model provenance, ties models to trusted publishers and scans the surrounding software environment before onboarding open-weight models, and Microsoft is developing weight scanning, red-teaming tools and cryptographic model signing. The researchers say frontier agents can spot anomalous weight modifications if asked to look, but no scan can prove a model is safe, and detection requires diff'ing against the original unmodified weights.

### GitHub — Gemini 4 Carbon in GitHub Copilot

- Company: GitHub (github.com)
- Announced: 2026-11-10T19:26:51+00:00
- Category: availability-change
- Coverage: 1 outlet
- Announcement: yes
- Group: covered
- Source: https://github.blog/changelog/2026-11-10-gemini-4-carbon-in-github-copilot
- Record: https://forck.live/items/34245-gemini-4-carbon-in-github-copilot
- Subject: Copilot

GitHub has added Google's Gemini 4 Carbon to Copilot, making the model generally available to Pro+, Max, Business and Enterprise users. Google built the model for complex coding, agentic and reasoning tasks, with what GitHub describes as a focus on secure code generation and security-focused work. The model is billed at provider list pricing under usage-based billing, rather than at a rate GitHub sets itself. Users can pick it from the model picker across Visual Studio Code, Visual Studio, Copilot CLI, the Copilot coding agent, the Copilot app, github.com, GitHub Mobile on iOS and Android, JetBrains IDEs, Xcode and Eclipse. Rollout is gradual, so the model may not appear for everyone immediately. Enterprise and Business administrators can control access through the model policy in Copilot settings, where new models are enabled by default unless an administrator has turned off the global default or disabled this one.

### Perplexity — Gemini 4 Carbon

- Company: Perplexity (perplexity.ai)
- Announced: 2026-11-10T18:12:28+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://docs.perplexity.ai/docs/resources/changelog#november-2026
- Record: https://forck.live/items/34178-gemini-4-carbon
- Subject: Sonar

The Agent API now supports google/gemini-4-carbon . Introductory pricing through November 30, 2026 is $2.00 per million input tokens, $0.10 per million cached input tokens, and $10.00 per million output tokens, including reasoning tokens. See the Agent API Models reference .

### Amazon — Autonomous treasury agents: How Trovata built an AI-native solution on Amazon Bedrock and AgentCore

- Company: Amazon (amazon.com)
- Announced: 2026-11-10T18:11:08+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/autonomous-treasury-agents-how-trovata-built-an-ai-native-solution-on-amazon-bedrock-and-agentcore/
- Record: https://forck.live/items/34177-autonomous-treasury-agents-how-trovata-built-an-ai-native-solution-on-amazon
- Subject: Bedrock / Nova

Trovata built autonomous treasury agents on Amazon Bedrock and AgentCore, moving its treasury AI from static insights to persistent workflows that act on behalf of treasury teams. The platform connects directly to over 10,000 banks, and Trovata says customers have seen $1M+ in annualized cost savings, over 5 hours per day saved per accounting team, and a 77 percent reduction in annual TMS costs. Agents run on schedules, triggers or rules to monitor activity, explain variance, detect anomalies and deliver updates without human intervention, while enterprise role-based access control keeps agents inside the same permission boundaries as human users. Trovata AI Chat lets users query multi-bank data in plain English, and the AI retrieves data through Trovata's own verified API layer inside AgentCore's isolated environment. Trovata says reporting that took hours or days now finishes in under a minute.

### Perplexity — Claude Fable 5.2

- Company: Perplexity (perplexity.ai)
- Announced: 2026-11-10T18:08:58+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://docs.perplexity.ai/docs/resources/changelog#november-2026
- Record: https://forck.live/items/34173-claude-fable-5-2
- Subject: Sonar

The Agent API now supports anthropic/claude-fable-5-2 . See the Agent API Models reference .

### Luma — One character. Full campaign. Without leaving the chat. Luma MCP is now live in ChatGPT and Claude.

- Company: Luma (lumalabs.ai)
- Announced: 2026-11-10T17:54:45+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://x.com/LumaLabsAI/status/2120213587620921545
- Record: https://forck.live/items/34174-one-character-full-campaign-without-leaving-the-chat-luma-mcp-is-now-live-in

Luma made its MCP integration available in ChatGPT and Claude, letting users run a full campaign from a single character without leaving the chat. The company announced the integration is now live in both assistants and pointed users to a link to try it. The post gives no details on pricing, limits or which Luma models the integration covers.

### GitHub — C++ language server now auto-configures MSBuild builds

- Company: GitHub (github.com)
- Announced: 2026-11-10T17:04:10+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://github.blog/changelog/2026-11-10-c-language-server-now-auto-configures-msbuild-builds
- Record: https://forck.live/items/34139-c-language-server-now-auto-configures-msbuild-builds
- Subject: Copilot

The Microsoft C++ Language Server for GitHub Copilot CLI now bundles its MSBuild compile commands extractor and runs it automatically, so agents no longer need a separate configuration skill before getting precise code intelligence. Previously, setting up go-to-definition, find references and other language server features for an MSBuild-based C++ project required either a hand-supplied compile_commands.json file or Copilot locating and running a configuration skill to generate a compilation database. That discovery step added setup time and consumed tokens before Copilot could start working with the code. Copilot now invokes the extractor when starting the language server on project open, and compilation database generation for MSBuild projects is deterministic, which GitHub says reduces setup token usage. CMake projects still use the generate compile commands skill introduced in the earlier release, and custom build systems can follow a guide to create their own extractor skill. An active GitHub Copilot subscription is required, and installation follows the Microsoft C++ Language Server quick start, including accepting the EULA.

### Meta — Your Algorithm: Giving People More Ways to Personalize Facebook Recommendations

- Company: Meta (meta.com)
- Announced: 2026-11-10T17:00:49+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://engineering.fb.com/2026/11/10/ml-applications/your-algorithm-giving-people-more-ways-to-personalize-facebook-recommendations/
- Record: https://forck.live/items/34120-your-algorithm-giving-people-more-ways-to-personalize-facebook-recommendations
- Subject: Llama / infrastructure

Your algorithm is a suite of AI-powered tools that makes Facebook recommendations more transparent and steerable, with a single editable view of the interests that shape Feed and Reels.

### AMD — AMD Showcases 6th Gen AMD EPYC Processors to Lead Next Era of AI and Enterprise Computing

- Company: AMD (amd.com)
- Announced: 2026-11-10T17:00:00+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://ir.amd.com/news-events/press-releases/detail/1305/amd-showcases-6th-gen-amd-epyc-processors-to-lead-next-era-of-ai-and-enterprise-computing
- Record: https://forck.live/items/34130-amd-showcases-6th-gen-amd-epyc-processors-to-lead-next-era-of-ai-and
- Subject: AMD Instinct / ROCm

AMD announced general availability of its 6th Gen AMD EPYC server SP7 processors, built on the new "Zen 6" architecture, with systems shipping from leading server vendors. AMD says the chips extend its server CPU performance leadership as AI creates new demands across the data center, and that cloud instances from Microsoft Azure and OCI will be generally available soon. The company calls the 6th Gen EPYC line the world's best server CPU portfolio, spanning agentic AI, enterprise, cloud, HPC and general-purpose workloads. AMD says the EPYC 9996 general-purpose CPU delivers up to 3.7x more performance than an Nvidia Vera CPU-based system and up to 2.2x the performance and 2.0x the performance per dollar of an Intel Xeon 6980P system. High-frequency host-node processors with up to 5.3 GHz can deliver up to 1.8x the LLM output throughput of previous generation EPYC processors, according to AMD. The portfolio also extends to AMD Helion agentic AI racks, with some configurations reaching up to 73,728 threads in one rack.

### Amazon — How TCS built an agentic test orchestrator with Amazon Bedrock AgentCore

- Company: Amazon (amazon.com)
- Announced: 2026-11-10T16:59:43+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/how-tcs-built-an-agentic-test-orchestrator-with-amazon-bedrock-agentcore/
- Record: https://forck.live/items/34118-how-tcs-built-an-agentic-test-orchestrator-with-amazon-bedrock-agentcore
- Subject: Bedrock / Nova

Tata Consultancy Services built an agentic test orchestrator on Amazon Bedrock AgentCore to speed up testing for its banking, financial services and insurance clients, where the slowest part of a release is the work after code is written. TCS says the goal was not to replace quality engineers but to remove repetitive translation and execution work so engineers could spend more time on risk analysis, exploratory testing and release decisions. The platform decomposes testing into specialized tasks, each with its own agent, tools, validation rules and context, with a deterministic supervisor routing work and an execution workflow controlling retries, error classification, persistence and reporting. TCS chose AgentCore for three reasons: Amazon Bedrock's choice of models lets each agent use the model that fits its task, the AgentCore runtime scales with demand without TCS provisioning or tuning agent servers, and each session runs in its own isolated microVM for session-level separation in the multi-tenant platform. Testers supply requirements through Jira stories, business requirement documents, Swagger specifications, uploaded documents, GitLab source folders or captured UI flows, and the platform generates test scenarios, Gherkin feature files and executable Java or Python tests. Tests run on Amazon ECS, results are stored in Amazon S3 and Amazon DynamoDB, and each tenant gets its own Kubernetes namespace, Amazon Cognito user pool, AWS KMS key, S3 bucket, DynamoDB tables and Amazon Bedrock Knowledge Bases.

### Meta — How We Built TLX: A Hardware-Native GPU Compiler for Production ML at Scale

- Company: Meta (meta.com)
- Announced: 2026-11-10T16:59:31+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://engineering.fb.com/2026/11/10/open-source/tlx-a-hardware-native-gpu-compiler-for-production-ml/
- Record: https://forck.live/items/34121-how-we-built-tlx-a-hardware-native-gpu-compiler-for-production-ml-at-scale
- Subject: Llama / infrastructure

Triton Low-Level Extensions (TLX) , is an extension to Triton that helps engineers optimize GPU performance while preserving Triton’s productive programming model.

### Amazon — Amazon Bedrock tokenomics data lake

- Company: Amazon (amazon.com)
- Announced: 2026-11-10T16:56:36+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/amazon-bedrock-tokenomics-data-lake/
- Record: https://forck.live/items/34119-amazon-bedrock-tokenomics-data-lake
- Subject: Bedrock / Nova

Amazon published a guide to building a centralized observability data lake for Amazon Bedrock invocations, aimed at platform and ML engineers running Bedrock workloads across many accounts. The problem it addresses is that Bedrock's model invocation logging delivers a full record of requests and responses, but the data stays locked inside individual accounts, so teams see an aggregate bill without knowing what happened inside each call. The design pulls raw invocation logs from linked accounts into CloudWatch, forwards them through subscription filters to a central account, processes them with Amazon Data Firehose, and exposes them through Athena as one queryable table. Each flattened record carries the prompt, the model response, input and output token counts, the model used, the stop reason, the account and timestamp, and the IAM role that sent the prompt. A Lambda function unwraps CloudWatch's batched, compressed events, discards housekeeping messages, and lifts the nested caller identity to the top level. Firehose converts the records from JSON to Parquet and files each object under a date path, and an AWS Glue table lets Athena read the files with ordinary SQL. Everything deploys as infrastructure as code through two CloudFormation templates, with templates, Lambda code and SQL queries in the AWS Sample Bedrock Invocation Data Lake GitHub repository.

### MiniMax — MiniMax H3-Image: Bringing 4K Image Generation into the 10-Second Era

- Company: MiniMax (minimax.io)
- Announced: 2026-11-10T16:23:20+00:00
- Category: open-weight-release
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://www.minimax.io/blog/minimax-h3-image
- Record: https://forck.live/items/34131-minimax-h3-image-bringing-4k-image-generation-into-the-10-second-era
- Subject: MiniMax models

MiniMax released MiniMax H3-Image, an open-weight image generation model built on H3 that handles text-to-image generation, reference-based generation and image editing. The company says the model can produce a full-resolution 4K (4096 × 4096) image in under 10 seconds, up to 5× the speed of mainstream alternatives, which it argues makes 4K generation practical for everyday creative workflows. MiniMax credits no single change but improvements across the post-training pipeline, from data and reinforcement learning to sampling-step reduction, PE and deployment optimization. H3-Image supports up to 16 reference images and, according to MiniMax, delivers state-of-the-art prompt adherence among open-weight models. As with H3, MiniMax is releasing the model weights to support open research and let developers and creators customize models and build their own tools and workflows. H3-Image is also available through the MiniMax API Platform, MiniMax Code and Hailuo AI; MiniMax says Standard mode is 1.5–2× as fast as mainstream alternatives and Highspeed mode reaches 3–4× the speed. Pricing is 50–70% of comparable mainstream offerings, and for launch week MiniMax is offering 50% off the standard price for image generation requests at 4K resolution and above.

### Google DeepMind — Gemini 4 Carbon is now available

- Company: Google DeepMind (deepmind.google)
- Announced: 2026-11-10T16:03:47+00:00
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://deepmind.google/blog/gemini-4-carbon-is-now-available/
- Record: https://forck.live/items/34090-gemini-4-carbon-is-now-available
- Subject: Gemini

Google DeepMind is making Gemini 4 Carbon available to developers, enterprises, scientists and consumers, calling it its strongest model for cybersecurity work and long-horizon agentic tasks. Carbon is built for work that tests a model's limits, such as multi-step coding and debugging and open-ended scientific research, and uses more tokens than Gemini 4 Argon to reason deeply and run extended verification loops. Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2X lower cost per task than the next best external model. It is available through the API in Google AI Studio and Gemini for business for paying developers and enterprises at the same introductory price as Argon, $2 per million input tokens and $10 per million output tokens. Carbon is also rolling out to all Google AI Ultra subscribers in Google Antigravity, the Gemini app and AI Mode in Search. Google is previewing /swarm in Google Antigravity, which turns a single prompt into a coordinated team of agents, and partnered with Artificial Analysis on a new Multimodal Knowledge Agent benchmark where Carbon ranks second at more than 6X lower average cost per task than the top model.

### Krea — Seedream 5.5 is now available in Krea.

- Company: Krea (krea.ai)
- Announced: 2026-11-10T16:00:15+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://x.com/krea_ai/status/2120184770781204497
- Record: https://forck.live/items/34094-seedream-5-5-is-now-available-in-krea

Krea added Seedream 5.5 to its image editing product, giving users a model it says can hold an image together across many rounds of edits. The company says the model supports editing the same image 100 times without breaking it, produces native transparent PNGs, and improves text rendering. Krea did not say who built Seedream 5.5 or give a price for using it.

### NVIDIA — NVIDIA Expands Open Source Commitment With OpenShell Contribution and New AI Infrastructure

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-10T16:00:03+00:00
- Category: infrastructure-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://blogs.nvidia.com/blog/open-source-openshell-infrastructure/
- Record: https://forck.live/items/34096-nvidia-expands-open-source-commitment-with-openshell-contribution-and-new-ai
- Subject: AI platform

NVIDIA is contributing its OpenShell agent-security runtime to the Cloud Native Computing Foundation and adding three open source components to its DSX AI Factory Platform for running GPU infrastructure on Kubernetes. OpenShell applies deterministic controls to agent activity so teams can enforce policies as agents use tools, access inference services and execute workloads, and it has drawn tens of thousands of contributions in less than a year since launch. NVIDIA Kubernetes Engine and NVIDIA Kubernetes Recipes, now in public preview, provide open source building blocks for managed Kubernetes services on GPU infrastructure, with Mirantis, Netris, OpenNebula Systems, Rafay, Spectro Cloud, vCluster, Firmus and SK Telecom planning integrations. NVIDIA Snapshot restores checkpointed replicas in seconds without model reload or warmup; design partner Crusoe tested GLM-5.3 on Crusoe Cloud and cut readiness time from roughly eight minutes to 31 seconds, while Saturn Cloud reduced vLLM startup from between two and 8.5 minutes to between six and 34 seconds. NVIDIA is also contributing to Agent Substrate, a Google-led CNCF Sandbox framework that preserves agent environments between tasks while releasing compute when agents are idle, and is working with Google on an OpenShell integration for securing agents across clouds and on-premises systems.

### Lovable — Find it, fix it, ship it: building secure apps on Lovable

- Company: Lovable (lovable.dev)
- Announced: 2026-11-10T14:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://lovable.dev/blog/find-it-fix-it-ship-it
- Record: https://forck.live/items/34044-find-it-fix-it-ship-it-building-secure-apps-on-lovable
- Subject: Lovable / AI app builder

Lovable's security team has laid out how the platform scans apps for security problems and fixes them, saying users do not need a technical background to get it right. Secure sign-in and access controls are set up by default, secrets are encrypted and kept out of frontend code, and apps are scanned as they are built, including on free accounts. Each time a user publishes, scanners check the code, application logic, database schemas and row-level security policies, and flag anything that needs attention before the app goes live. Users can run a deeper scan at any time for free; Lovable says it is built on its own testing system, trained on the security issues that come up most often in Lovable apps. When the scanner flags an issue, selecting Fix sends an agent specific instructions for that exact problem, and the most common issue, data access rules that are too open, can be fixed in a few clicks. Turning on "Auto-fix security issues" in project settings resolves many findings with the next chat message, and publishing controls let a user or workspace admin block anything with critical findings from going live. Lovable says it plans a redesigned security experience and scheduled scans on all paid plans this month.

### Google DeepMind — We’re bringing Adobe’s creative and productivity tools closer to your everyday Google apps.

- Company: Google DeepMind (deepmind.google)
- Announced: 2026-11-10T14:00:00+00:00
- Category: partnership-acquisition
- Coverage: 1 outlet
- Announcement: yes
- Group: covered
- Source: https://blog.google/company-news/inside-google/company-announcements/adobe-google-partnership-integration/
- Record: https://forck.live/items/34016-we-re-bringing-adobe-s-creative-and-productivity-tools-closer-to-your-everyday
- Subject: Gemini

Google says it is expanding its partnership with Adobe to bring Adobe's creative and productivity tools into Google's apps and Google's frontier AI models into Adobe's apps. The company says the goal is to stop people losing focus when they switch between apps to edit a PDF or fine-tune an image. Google Drive and Gmail will get a suggested-app pathway into Adobe Acrobat, including an "Edit with" entry point in file previews. Adobe is expanding its connectors in Gemini and Gemini Enterprise with Acrobat and additional creative tools, and will bring creative tools to AI Mode in Google Search. Businesses will be able to use Adobe's CX Enterprise Coworker agent in Gemini Enterprise, which Google describes as an AI assistant that helps teams design graphics, work with PDFs and launch marketing campaigns without switching apps. Adobe and Google DeepMind will also explore new editing capabilities for creative professionals, starting with an image model for Photoshop that could enable near-real-time edits.

### ElevenLabs — Introducing Scribe v3 and v3 Live, our most accurate speech to text models

- Company: ElevenLabs (elevenlabs.io)
- Announced: 2026-11-10T12:00:00+00:00
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://elevenlabs.io/blog/scribe-v3
- Record: https://forck.live/items/34161-introducing-scribe-v3-and-v3-live-our-most-accurate-speech-to-text-models
- Subject: Voice AI

ElevenLabs launched Scribe v3 and Scribe v3 Live, two speech-to-text models it says are its most accurate, aimed at messy real-world audio such as overlapping speakers and mid-sentence pauses. Scribe v3, for recorded audio, makes 27% fewer word errors than Scribe v2 across 102 languages and has the most accurate speaker labels of any model ElevenLabs tested. Scribe v3 Live, the low-latency variant for realtime audio, makes 43% fewer errors than Scribe v2 Realtime and adds speaker diarization for the first time, along with improved turn detection. Scribe v3 Live is available today in ElevenAgents, and both models are available in ElevenAPI, where switching from Scribe v2 is a one-line change to the model ID. Scribe v3 makes 17% fewer diarization errors than Scribe v2 on recordings with three or more speakers, and Scribe v3 Live has a diarization error rate of 3.1% on the English portion of ElevenLabs' 16-language real-world dataset, against 5.9% to 7.3% for other realtime models tested. Scribe v3 supports keyterm prompting with up to 1,000 terms and transcript edits written in plain English, while Scribe v3 Live adds a multilingual mode for conversations that switch languages mid-sentence and a keyterm library for vocabularies too large for a list.

### Google Cloud — Introducing Gemini for healthcare

- Company: Google Cloud (cloud.google.com)
- Announced: 2026-11-10T12:00:00+00:00
- Category: not stated
- Coverage: 2 outlets
- Announcement: yes
- Group: covered
- Source: https://cloud.google.com/blog/products/ai-machine-learning/gemini-for-healthcare/
- Record: https://forck.live/items/33986-introducing-gemini-for-healthcare
- Subject: Gemini Enterprise

Every healthcare organization is working to deliver great care to more people at a lower cost. However, a general AI model is not enough to solve this challenge.

### OpenAI — Helping caregivers use AI in everyday life

- Company: OpenAI (openai.com)
- Announced: 2026-11-10T12:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/helping-caregivers-use-ai-in-everyday-life
- Record: https://forck.live/items/33968-helping-caregivers-use-ai-in-everyday-life
- Subject: GPT / ChatGPT / API

In collaboration with Caregiver Action Network, Alzheimer’s Association, and American Diabetes Association ® , bringing free, hands-on ChatGPT workshops to five U.S. cities during National Family Caregivers Month.

### Google DeepMind — Opening the datasets behind our AI weather models

- Company: Google DeepMind (deepmind.google)
- Announced: 2026-11-10T11:13:04+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://deepmind.google/blog/opening-the-datasets-behind-our-ai-weather-models/
- Record: https://forck.live/items/35821-opening-the-datasets-behind-our-ai-weather-models
- Subject: Gemini

Google DeepMind is releasing SatWeave and PARDIG, two datasets behind its WeatherNext 3 model, so researchers and developers can build new weather, hydrology and climate applications. The team says observing Earth is the starting point for predicting weather and the water cycle, but turning raw satellite and ground measurements into a coherent global picture quickly is a major scientific and engineering challenge. SatWeave blends observations from six geostationary satellites into a global mosaic across 17 visible, water vapor and infrared channels, updating hourly rather than every few hours, and a historical archive spans over a decade. PARDIG, built on an AI architecture similar to WeatherNext 3 and trained on sparse spaceborne radar observations from NASA's GPM Core Observatory, produces global precipitation maps at 15-minute and roughly 10-kilometer resolution with an average latency of about five hours. Evaluated against continental radar networks in the US, Europe and Japan and over 36,000 weather stations worldwide, PARDIG improves hourly precipitation detection skill over NASA's IMERG-Final by more than 30% on light-to-moderate precipitation, according to the team. Live feeds and historical archives are openly available on Google Cloud Storage, with PARDIG's historical catalogs also in Google Earth Engine.

### OpenAI — Closing the AI guidance gap in education

- Company: OpenAI (openai.com)
- Announced: 2026-11-10
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/closing-ai-guidance-gap-education
- Record: https://forck.live/items/33888-closing-the-ai-guidance-gap-in-education
- Subject: GPT / ChatGPT / API

OpenAI is publishing lessons from its education partnerships across EMEA, arguing that schools and governments need clearer guidance as students adopt AI faster than teachers can teach it. New research from ThinkYoung, cited by OpenAI, found that 79% of surveyed students reported using AI for homework or assignments, but only 47% reported being shown how. The company says its collaborations in Estonia, Slovakia, Greece and Italy share three priorities: equitable access to frontier AI tools, jointly designed AI literacy programmes, and ongoing measurement and evaluation. Three early patterns emerge from the work: strong public leadership to set direction, investment in teachers' skills and confidence, and practical support tailored to local curricula and classroom needs. OpenAI says teacher adoption and reported time savings are encouraging early signals, but that understanding where AI strengthens knowledge, independent thinking and retention requires sustained evaluation. The report is presented as early experience rather than a single model for education systems to follow, and OpenAI says it has a responsibility to learn openly from findings that challenge its assumptions.

### OpenAI — Singapore Airlines elevates travel & work with OpenAI technology

- Company: OpenAI (openai.com)
- Announced: 2026-11-10T08:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/singapore-airlines-travel-work
- Record: https://forck.live/items/34473-singapore-airlines-elevates-travel-work-with-openai-technology
- Subject: GPT / ChatGPT / API

Singapore Airlines is using the OpenAI API and OpenAI models across customer support, travel discovery and internal engineering, and says the work has doubled satisfaction with its Kris chatbot. Kris moved from a traditional dialog-flow system to an LLM-powered conversational experience, and the airline reports a 2x increase in customer satisfaction alongside roughly 10% lower average handling time for representatives using AI-assisted tools such as call summarization. The airline says the old design required every topic and flow to be mapped and written in advance, which made coverage hard to scale and pushed missing intents to human agents. Kris now interprets intent in natural language, grounds answers in a knowledge base, handles multi-topic conversations and hands over to a representative with context; it covers general queries, special requests, disruption information and selected transactions, with booking changes and baggage work still to come. Flight Recommender, built on the API with an orchestration layer over fare and availability data, has grown monthly users nearly 3x in the past year, and the airline is preparing a more conversational version. OpenAI models are also used in parts of the software development lifecycle, where Singapore Airlines reports productivity gains of up to 20% across selected activities, with staff directing and validating AI output rather than producing every artifact.

### Google Antigravity — Gemini 4 Carbon and /swarm

- Company: Google Antigravity (antigravity.google)
- Announced: 2026-11-10T08:00:00+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://antigravity.google/blog/gemini-4-carbon-in-google-antigravity
- Record: https://forck.live/items/34100-gemini-4-carbon-and-swarm
- Subject: Antigravity

Google is rolling out Gemini 4 Carbon inside Google Antigravity for Google AI Ultra subscribers, alongside /swarm in preview for the Antigravity App and CLI. Carbon is Google's model for long-horizon software engineering, knowledge work, scientific discovery and cybersecurity work. A single agent works for scoped tasks but breaks down when a change spans dozens of files and test suites, Google says. Instead of one coordinator context routing every step, /swarm sends the prompt to a coordinator that spins up parallel workers in isolated git worktrees and adversarial verifiers that run tests and challenge diffs. Agents post to a shared append-only message board, park at zero token cost when idle, and activate when an agent needs input. Google says Carbon ranks #1 on CursorBench 4.0 and Perplexity's WANDR, suiting coordinator, architecture, auditing and verification roles, while lighter models handle parallel file edits and routine implementation.

### Amazon — How skills and agents work in AWS Quick

- Company: Amazon (amazon.com)
- Announced: 2026-11-10T00:21:34+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/how-skills-and-agents-work-in-aws-quick/
- Record: https://forck.live/items/33745-how-skills-and-agents-work-in-aws-quick
- Subject: Bedrock / Nova

Amazon's guide explains how agents, skills and data fit together in AWS Quick, its AI-powered service for automating tasks, analyzing data, building web applications and conducting research. An agent is a configurable AI with its own instructions, knowledge sources called spaces, and connectors, and it can only reach the resources and services that the user can. Quick is the default built-in agent every user has access to, with no fixed scope, while a custom agent is configured for a specific job and can be shared with users and groups after testing and publishing. Agents ground their answers in connected data: spaces hold curated documents and resources, knowledge bases are indexed collections synced from sources such as Google Drive, Confluence, SharePoint, Amazon S3 and Amazon Bedrock Managed Knowledge Bases, and datasets and dashboards come from Amazon Quick Sight. Connectors link Quick to external applications such as Microsoft Outlook, Slack, Jira, Asana and Salesforce, and to remotely hosted Model Context Protocol servers, while local connectors run on the user's own machine and are available only when it is on and connected. A skill is a reusable set of instructions that teaches an agent a specific task, and much of what Quick does by default comes from system skills that can generate Word documents, PowerPoint presentations, Excel spreadsheets or PDFs, browse the web, run code, execute deep research and transcribe audio.

### Apple — Toward a Theory of Evaluability

- Company: Apple (apple.com)
- Announced: 2026-11-10
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://machinelearning.apple.com/research/theory-evaluability
- Record: https://forck.live/items/36826-toward-a-theory-of-evaluability
- Subject: Machine Learning Research

Apple researchers have formalized a notion of evaluability for machine learning models that does not assume the test set was held out from training, addressing a gap that is increasingly hard to justify for large foundation models. The work asks whether black-box models can be evaluated when test samples may have been seen during training, and studies the problem under different levels of access to the model. In the setting where the evaluator only sees the model's output on the test set, the authors fully characterize the distributions where accurate evaluation is possible and give tight upper and lower bounds on the sample complexity of evaluation. Their results show that black-box evaluation in this setup is feasible if and only if the distribution is close to being small support. The paper then considers evaluation algorithms that can query the model on additional inputs, drawing connections to self-correctors studied in program testing. Using these, the authors show that a small amount of query access can increase the power of the evaluator for natural distributions and concept classes.

### Apple — Understanding Private Evolution as Learning-Augmented Clustering

- Company: Apple (apple.com)
- Announced: 2026-11-10
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://machinelearning.apple.com/research/private-evolution-clustering
- Record: https://forck.live/items/36825-understanding-private-evolution-as-learning-augmented-clustering
- Subject: Machine Learning Research

Apple researchers recast Private Evolution, a differentially private algorithm for synthetic data generation, as generative model-augmented Wasserstein learning, arguing that accounting for the generator yields much better performance bounds than worst-case analyses predict. The paper shows that if the generator produces samples in the same low-dimensional space as the distribution, sample complexity depends on intrinsic rather than ambient dimension. The authors also report that standard variants of Private Evolution can fail to converge on simple well-clustered instances, and propose a geometry-aware version with provable convergence on such instances. Experimentally, they say the new algorithm is competitive with standard baselines and can improve recall. The work was done while one author was at Apple, with a co-author from Yale University.

### H Company — Same screen, two approaches: watching a script and an agent hit the same change

- Company: H Company (hcompany.ai)
- Announced: 2026-11-10
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://hcompany.ai/newsroom/same-screen-two-approaches
- Record: https://forck.live/items/35068-same-screen-two-approaches-watching-a-script-and-an-agent-hit-the-same-change
- Subject: Holo

What happens when a portal layout changes under a scripted bot, and under a computer-use agent, step by step. Insights November 10, 2026 · 2 min read Here's a test we run in the first week of most deployments.

### H Company — Eight questions your CISO will ask about computer-use agents, and the answers

- Company: H Company (hcompany.ai)
- Announced: 2026-11-10
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://hcompany.ai/newsroom/ciso-questions-computer-use-agents
- Record: https://forck.live/items/35067-eight-questions-your-ciso-will-ask-about-computer-use-agents-and-the-answers
- Subject: Holo

H has published answers to the eight security questions it says enterprises raise before a computer-use agent touches a production system. The company says agents run under a user-level account the customer provisions, with the same permissions a new team member would get and no admin rights or service-account shortcuts. Agents are read-only by default, and write actions are enabled step by step only after customer approval, usually following a sandbox phase. Connectivity is an outbound-only connection from a sidecar in the customer's environment, with no inbound ports to open. H says it does not use customer data to train its models by default, though some enterprise customers allow it in their agreement, and the platform runs in the US and the EU while inference runs in the US. The company holds SOC 2 Type II, with the report available under NDA through its trust center, and every run is logged step by step so auditors can review the same trail operators see. Customers set the thresholds for which items an agent can close and which go to a human reviewer, and can pause any agent or revoke its account at any time.

### H Company — Before you renew: 10 questions to ask about your automation estate

- Company: H Company (hcompany.ai)
- Announced: 2026-11-10
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://hcompany.ai/newsroom/before-you-renew-automation-questions
- Record: https://forck.live/items/35066-before-you-renew-10-questions-to-ask-about-your-automation-estate
- Subject: Holo

H Company published a ten-question checklist for automation and operations leaders whose automation renewal falls in the first half of 2027, arguing the useful work happens before the quote arrives. The questions cover the estate itself, starting with how many licensed bots actually ran in the last 30 days, since the gap between licensed and active is often the first saving. It also asks which ten bots broke most often and why, what break-fix cost in CoE hours and contractor spend rather than tickets alone, and how long a typical fix takes from failure to green. A second group covers the work bots hand back to people: how many items are handed over each month, which queue is biggest, and what makes those items exceptions. H Company says a queue with more than 1,000 items a month is a good place to test something new, and that a two-week test on real work is worth more than any slide. It offers two starting points: an Exception Queue Scorecard of 12 questions taking about four minutes, and a free Bot Estate Audit that answers the first four questions in two weeks from an export and two workshops, with no access to production systems.

### Apple — Weighted State Souping for Efficient Long-Context Retrieval

- Company: Apple (apple.com)
- Announced: 2026-11-10
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://machinelearning.apple.com/research/weighted-state-souping
- Record: https://forck.live/items/34762-weighted-state-souping-for-efficient-long-context-retrieval
- Subject: Machine Learning Research

Apple researchers, with Harvard co-authors, propose weighted state souping, a method that learns query-dependent weights for composing cached state space model encodings into a context for downstream queries. State souping and vanilla SSMs deteriorate on long, crowded contexts, which limits their use in long-horizon domains, the paper says. The resulting models improve long-context performance substantially across single- and multi-hop document question answering, needle-in-a-haystack retrieval, variable tracking, and tool use. A sparse variant yields near-constant encoding cost with respect to context length while preserving, and sometimes improving, performance. The paper also offers theoretical arguments for why souping works and when it may fail, and discusses informally the advantages of weighted souping. The work was accepted at the Workshop on Personalized, Aligned, Long-term Memory for AI Systems (PALM) at NeurIPS 2026.

### Apple — On the Importance of Gating: Memorization vs. In-Context Learning in State Space Models

- Company: Apple (apple.com)
- Announced: 2026-11-10
- Category: research-paper
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://machinelearning.apple.com/research/gating-memorization-in-context
- Record: https://forck.live/items/34761-on-the-importance-of-gating-memorization-vs-in-context-learning-in-state-space
- Subject: Machine Learning Research

Apple researchers, with Harvard co-authors, published a study showing that the gating mechanism in State Space Models explains both their strengths and failures in in-context learning and retrieval. SSMs offer constant memory and linear compute, but they lag behind Transformers on tasks that need in-context learning and precise retrieval, which has slowed their adoption for large-scale language modeling. Through theory and experiments, the authors show gating causes SSMs to first learn an in-weights memorization solution while delaying, or even preventing, convergence to a correct in-context learning solution. The paper says this happens even when the architecture and its memory capacity impose no fundamental limitation. On the other hand, the authors find gating is often beneficial for improving generalization to long sequence lengths. The work was done while the lead author was at Apple, with Harvard University as an affiliation.

### Apple — Evaluating the Evaluators: Investigating LLM Judges for Personalized Writing Style Assessment

- Company: Apple (apple.com)
- Announced: 2026-11-10
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://machinelearning.apple.com/research/evaluating-evaluators-llm-judges
- Record: https://forck.live/items/34760-evaluating-the-evaluators-investigating-llm-judges-for-personalized-writing
- Subject: Machine Learning Research

Apple researchers evaluated how well large language model judges assess personalized writing style, work accepted at the Can We Trust AI Evaluation? Workshop at NeurIPS 2026. The team generated text with four LLMs across three long-form writing tasks, collected human labels on which of two responses was better, and compared those with judge-LLM pairwise rankings. Judge quality correlated weakly with general LLM ability measured by MMLU, at a Pearson's rho of 0.2 or less, and varied by writing task and by whether the criteria were general helpfulness or style personalization. LLM evaluators were more consistent and performed better when comparing responses in pairs rather than scoring them independently. The researchers also found the strongest LLMs did not most reliably reproduce human evaluations, which they hypothesize relates to exaggerated sensitivity to subtle details that human annotators and weaker LLMs ignore.

### Ideogram — 7 Best AI Cartoon Generators in 2026: Photos & Characters

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-10
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/best-ai-cartoon-generators
- Record: https://forck.live/items/34319-7-best-ai-cartoon-generators-in-2026-photos-characters
- Subject: Ideogram image models

Ideogram has published a comparison of seven AI cartoon generators, testing the same generated portrait through Ideogram, Canva and Adobe Firefly and running scene and color edits on a coffee-shop mascot in Ideogram. The company says the portrait instructions differed between tools, so the images compare styles rather than models, and that official product guides and pricing were used for Neolemon, Leonardo.Ai, Fotor and Midjourney. Canva and Adobe Firefly both returned free PNG downloads without visible watermarks, at 1264 × 1264 and 2048 × 2048 respectively, while Ideogram's mascot downloads came out at 2048 × 2048 with no visible watermark. Ideogram says its own tool kept the mascot's pointed ears, cream muzzle, orange fur and green apron through a Precise Edit request for a waving pose, though some proportions changed and a storefront sign contained garbled lettering. A follow-up edit that turned the green apron navy blue left the name, pose, face and shop setting unchanged at normal viewing size. Ideogram cites a Contra Labs blind evaluation in which 10 designers assessed 240 images and reported Ideogram 4.0 first in 47.9% of typography comparisons against 30.0% for Nano Banana 2, with a 3.55 out of 5 rating for client work. Free generation is available, but Image Reference and AI editing require priority credits, and the Plus plan costs $180 billed annually, equivalent to $15 a month.

### Ideogram — 8 Best AI Image Editing Tools in 2026, Tested & Reviewed

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-10
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/best-ai-image-editing-tools
- Record: https://forck.live/items/34318-8-best-ai-image-editing-tools-in-2026-tested-reviewed
- Subject: Ideogram image models

Ideogram tested eight AI image editors on a single handbag photo to see which ones make a requested edit without disturbing the rest of the product shot. The company ran recolor, strap-removal and background-replacement prompts in Ideogram, ChatGPT, Gemini and Firefly, and reviewed Photoshop, Photoroom, Pixlr and Canva by features and plans instead. Ideogram says its masked recolor changed the front panel to burgundy while the handles and side panel stayed brown and the gold clasp stayed in place, which it calls the pick for edits confined to one panel. With prompts alone, Ideogram says the front came out plum-toned rather than the requested oxblood, and ChatGPT came closest to the requested oxblood. Canva stopped the strap-removal request at an upgrade prompt, and the test used Ideogram Precise Edit with 4.5, Firefly Image 5 at Auto / 2K, a ChatGPT selector showing GPT-6.1 Sol Light, and Gemini's Auto with Images selected. Ideogram advises trying the edit you actually need and downloading the result before choosing a tool.

### Ideogram — 8 Best AI Mockup Generators in 2026: Tested & Reviewed

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-10
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/best-ai-mockup-generators
- Record: https://forck.live/items/34317-8-best-ai-mockup-generators-in-2026-tested-reviewed
- Subject: Ideogram image models

Ideogram has published a hands-on comparison of AI mockup generators, testing the same NORTHLINE tote artwork in its own tool and in Recraft. The test used one fixed source graphic, a transparent PNG with a serif NORTHLINE wordmark, a burgundy rule and the smaller line EVERYDAY OBJECTS, so the same lettering and layout could be checked across tools. A canvas tote was chosen because the artwork had to look printed on fabric, with folds, weave and shading, and the smaller line gave another detail to inspect after editing. In Ideogram the graphic was opened in Precise Edit with Background set to Auto, while Recraft took the same PNG through its Image workflow on a square format with the model selector on Auto. Both tools then changed only the background and surface to a warm beige studio setting, and the design stayed readable in each; Ideogram's framing and folds shifted slightly, while Recraft's fabric appeared warmer. A further Ideogram edit made the printed artwork about 25% smaller while keeping it centered, and both wording lines stayed readable, a revision done without rebuilding the presentation. No visible watermarks appeared on the downloaded files, and the comparison groups tools into template, generative and 3D mockups, with Ideogram Plus at $15 a month billed annually at $180 and Recraft Basic at $10 a month billed annually at $120.

### Ideogram — 6 Best AI Image Generators for Anime in 2026

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-10
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/best-ai-image-generators-for-anime
- Record: https://forck.live/items/34316-6-best-ai-image-generators-for-anime-in-2026
- Subject: Ideogram image models

Ideogram compared its own image generator with PixAI and four other tools for anime illustration, running the same character brief, rooftop scene change and jacket recolor through each. Ideogram 4.5 kept both boots in frame and held the character's face, pose and accessories steady through a masked jacket recolor, while PixAI's Tsubaki.3 produced flatter color blocks and stronger outlines. The test found both rooftop edits altered the character's satchel, so the guide advises checking bag shape, badge, hair clip and garment construction before approving a new scene. Ideogram ran at High quality with a 2:3 aspect ratio and Magic Prompt off, quoting one image at 10 credits and 1365 × 2048 pixels, while PixAI used Single, XL, 3:5 and High Priority settings at 4,400 credits and 1104 × 1824 pixels. Because credits are not comparable across platforms, the guide suggests comparing subscription prices and how many revisions each plan covers.

### Baseten — How Baseten and Genspark trained a deep research model to outperform Opus 5.5

- Company: Baseten (baseten.co)
- Announced: 2026-11-10
- Category: model-update
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://www.baseten.co/blog/how-baseten-and-genspark-trained-a-research-model-to-outperform-opus
- Record: https://forck.live/items/34203-how-baseten-and-genspark-trained-a-deep-research-model-to-outperform-opus-5-5
- Subject: Baseten

Baseten Applied Research worked with Genspark to post-train Qwen3.5-397B into a specialized Deep Research model that now powers Genspark's Deep Research agent in production. On Genspark's own quality grader the post-trained model scores 0.759, up from 0.612 for the base model, and above Opus 5.5 at 0.738, Sonnet 5.5 at 0.742, GPT-5.6 Sol at 0.637, Kimi K3 at 0.724 and DeepSeek V4 Pro at 0.675. Genspark says it moved to an open model because frontier closed-weight models are expensive at its volume and allow only prompt changes, not weight changes. Training ran against Genspark's own agent harness and tools, so the model learned in the environment it would serve in. Diagnosing the agent's trajectories, the teams found synthesis was the larger bottleneck: holding research fixed and swapping in a stronger writer model raised the score from 0.58 to 0.73 on a small diagnostic panel. The model crawls 5× more sources per user query than the base model and writes longer, better-cited reports, and runs with NVFP4 weights on 2×B200 GPUs per replica. Genspark keeps the weights, evaluations and training code, according to co-founder Kay Zhu.

### Laminar — Taxonomy of agent failures: how AI agents fail in production

- Company: Laminar (laminar.sh)
- Announced: 2026-11-10
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://laminar.sh/blog/taxonomy-of-agent-failures
- Record: https://forck.live/items/34197-taxonomy-of-agent-failures-how-ai-agents-fail-in-production
- Subject: flow / agent observability

Laminar sorted thousands of failures from production agents into 7 categories plus an "Other" bucket and published the resulting taxonomy, after running its Signals trace intelligence on agent traces from coding, computer use, legal and finance. Hallucinations account for 26% of failures, with plain fabrication making up about half of that category and nearly 2 in 3 agents fabricating at least once. Unfinished tasks are 21% of failures and appear in 9 in 10 agents, often as traces that simply stop or hand the task back to a human without exhausting the agent's own tools. Breaking an explicit instruction is 20% of failures, most often a skipped mandatory step, a content or style violation, or malformed output such as JSON wrapped in Markdown fences. Tool errors, 12% of failures, are the kind agents recover from most: more than a third are fixed because the errors come back as messages. Wasted work, 9% of failures, is mostly loops and retry storms, and Laminar says open-weight models loop 1.7x as often as other agents of the same type, at 1.7x the tokens of a typical flagged run. The "Other" bucket covers infrastructure failures, defects in the agent's own prompt or input data, and multi-agent hand-off issues; only 13% of failures are recovered even partially, and agent type predicts the failure mix about 1.5x as much as the model vendor.

### Recraft — What's New in Gemini Omni Flash 1.2, Now in Recraft Studio

- Company: Recraft (recraft.ai)
- Announced: 2026-11-10
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.recraft.ai/blog/gemini-omni-flash-1-2-in-recraft-studio
- Record: https://forck.live/items/34189-what-s-new-in-gemini-omni-flash-1-2-now-in-recraft-studio
- Subject: Recraft V4

Recraft has added Google's Gemini Omni Flash 1.2 video model to Recraft Studio, where it generates clips of up to 10 seconds with native audio. Google released the model on November 5, 2026 as a direct update to Gemini Omni Flash 1.1, and Recraft says most of the changes are about control rather than new features. Characters, objects and scenes drift less across a clip, camera moves follow the prompt with fewer unwanted cuts, and reference images for a character, product or style carry through more faithfully. Dialogue improves too: two characters can trade a line each with mouths that match the words, which Recraft calls usually the first thing to break in a short generated scene. Google is direct about where the model still struggles, naming fast sports action, rapid script changes and complex physics, so Recraft advises planning for more takes on those briefs. The model sits in the model picker next to the other video models, on plans that include video and external models, and each generation uses credits; it joins Gemini Omni Flash, which arrived in Recraft Studio in July. Recraft notes that Google's own API allows longer clips for this model, while Recraft Studio generations are capped at 10 seconds.

### Anthropic — How Curative is building a national provider network with agents on Claude

- Company: Anthropic (anthropic.com)
- Announced: 2026-11-10
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://claude.com/resources/articles/how-curative-is-building-a-national-provider-network-with-agents-on-claude
- Record: https://forck.live/items/34172-how-curative-is-building-a-national-provider-network-with-agents-on-claude
- Subject: Claude

Curative built a hierarchy of agents on Claude to sign health-insurance provider contracts about 50 times faster than its team did in 2025, after concluding that hiring alone could not close the gap. The company sells employer health insurance with $0 copays and $0 deductibles for in-network care once members complete a Baseline Visit, and it says signing 2,100 contracts a year would still have made a national network a 30-year build. Its main agent, Gwen, handles one provider deal from outreach to signature, mostly on Claude Opus, so nothing from other deals influences it. A rate reviewer on Sonnet clears about two-thirds of Gwen's above-template rate requests, and anything non-standard goes to CEO Fred Turner for approval. A judge on Haiku reads every message before it is sent, a check Curative added after Gwen copied providers on internal notes about how high it could go. Each contract takes about 20 million tokens, roughly a tenth of the original cost, and the process can run from form to signature in as little as an hour. Curative's contracting team has grown from 45 to about 55 people since Gwen launched, each about 2.5 times as productive, and the company pays that team on Gwen's deals so people and agents share one target.

### Anthropic — More ways to work with Claude: Sheets, Whiteboards, and Interactive Connectors

- Company: Anthropic (anthropic.com)
- Announced: 2026-11-10
- Category: capability-change
- Coverage: 1 outlet
- Announcement: yes
- Group: covered
- Source: https://claude.com/resources/articles/more-ways-to-work-with-claude-sheets-whiteboards-and-interactive-connectors
- Record: https://forck.live/items/34135-more-ways-to-work-with-claude-sheets-whiteboards-and-interactive-connectors
- Subject: Claude

Anthropic is adding new artifact types to Claude — Sheets, Whiteboards and Data — so teams can build spreadsheets, whiteboards and dashboards together with the model. Claude Sheets lets a team and Claude build spreadsheets collaboratively, organize data and create and edit formulas across multiple tabs, while Claude Whiteboards is a digital whiteboard for laying out ideas in a meeting that Claude can then summarize. Claude Data, formerly Dashboards, connects to tools like Snowflake and Salesforce and can use definitions a data team certifies in beta to create dashboards, visual data stories or narrated short videos. Anthropic says the same context carries through every step, so an idea can move from a whiteboard to a spreadsheet to a dashboard without the details being repeated. Interactive connectors now let partners embed their apps directly in Claude, with plugins from Salesforce, NetSuite and Workday in beta carrying built-in skills, and Claude Meetings is coming soon to take notes on calls. Anthropic cites its own legal team cutting marketing reviews from three days to 24 hours, and a customer FP&A team that brought forecast-cycle assembly work from about 10 hours to under 30 minutes.

### Anthropic — Your apps can now open in Claude, and Claude can work in them

- Company: Anthropic (anthropic.com)
- Announced: 2026-11-10
- Category: capability-change
- Coverage: 2 outlets
- Announcement: yes
- Group: covered
- Source: https://claude.com/resources/articles/interactive-connectors-full-tool
- Record: https://forck.live/items/34134-your-apps-can-now-open-in-claude-and-claude-can-work-in-them
- Subject: Claude

Anthropic is letting Claude work inside connected tools, so users can open a tool beside a conversation or full screen and have Claude make changes to an open page, row or file. The feature is rolling out on the web and in the Claude desktop app on all plans, as part of the new Claude experience that brings chat and Cowork together, and appears once an account is updated or an organization turns it on. Anthropic says changes appear in the tool as Claude makes them, and users approve its actions the same way they approve a connector's actions. The whole tool opens next to the conversation, so tools like Confluence or Notion give the full editor signed into the user's own account rather than a limited preview, and tools can be opened directly without writing a prompt or pinned to the sidebar. Anthropic is starting with a small group of tools from its directory, including Asana, Canva, Carta, Dropbox, monday.com, Smartsheet, Vercel and Webflow, while Figma, Harvey and Shopify are building integrations. The connectors run on MCP Apps, the open extension to the Model Context Protocol, and Anthropic says an MCP App can run in any AI application that supports the extension; support for app tools and embedded sites is in public preview for developers. Admins can restrict a connector's tools the same way they do today.

### Anthropic — Introducing Claude Whiteboards, Claude Sheets, and Claude Data

- Company: Anthropic (anthropic.com)
- Announced: 2026-11-10
- Category: availability-change
- Coverage: 4 outlets
- Announcement: yes
- Group: covered
- Source: https://claude.com/resources/articles/sheets-whiteboards-data
- Record: https://forck.live/items/34133-introducing-claude-whiteboards-claude-sheets-and-claude-data
- Subject: Claude

Anthropic is launching Claude Whiteboards and Claude Sheets in beta and making Claude Dashboards generally available as Claude Data, all on paid plans. Whiteboards is a shared board where teams brainstorm with Claude, Sheets is a spreadsheet Claude builds and edits with you, and Claude Data adds data stories, certified queries, and live sharing. The company says working with Claude on a project often meant moving the output to another tool where Claude had no live access to follow progress and changes. Sheets can generate spreadsheets with tabs, formulas, formatting, and drop-downs from a prompt, and edit in the cells while outlining each change; Anthropic says sheets have simple spreadsheet functionality for now and is working on adding more. Claude Data, which launched in beta in October as Claude Dashboards, now supports caching with per-dashboard refresh cadences, tracked changes with version rollback, and design systems. On Team plans all three are on by default; on Enterprise plans Whiteboards and Sheets are off by default and Claude Data turns on by default on November 19, 2026.

### Hugging Face — Pay for Hugging Face Hub with your Google Cloud Account

- Company: Hugging Face (huggingface.co)
- Announced: 2026-11-10
- Category: availability-change
- Coverage: 1 outlet
- Announcement: yes
- Group: covered
- Source: https://huggingface.co/blog/hub-gcp-marketplace
- Record: https://forck.live/items/34129-pay-for-hugging-face-hub-with-your-google-cloud-account
- Subject: Platform

Hugging Face now lets organizations pay for Hub usage through their Google Cloud account, so charges appear on the existing Google Cloud invoice instead of a credit card. The company says the change removes the need to onboard a new vendor or use a corporate card just to host private models, run GPU Jobs, deploy Inference Endpoints, call Inference Providers or host GPU Spaces. Admins subscribe to the Hugging Face Hub listing on the Google Cloud Marketplace, then link the organization from its billing settings; every subscription is reviewed by the Hugging Face team before paid usage is billed through it. While a subscription is pending, usage keeps being billed to the organization's existing payment method, if any. Organizations with a credit card on file switch at the start of their next billing period, while free-tier organizations without a card usually switch immediately. Google Cloud replaces the credit card as the payment method, and adding credits is not available while the organization is billed through the marketplace, though monthly included credits on Team or Enterprise subscriptions still apply before pay-as-you-go usage is reported. Linking and unlinking are recorded in the organization's audit logs.

### Factory AI — Introducing Security Factory

- Company: Factory AI (factory.com)
- Announced: 2026-11-10
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://factory.com/news/security-factory
- Record: https://forck.live/items/34124-introducing-security-factory
- Subject: Droid

Factory introduced Security Factory, a product that connects investigation, remediation and verification so engineering teams can turn security findings into tested fixes, and it is now in Private Preview. The company says scanners, threat intelligence, asset inventories, Jira tickets and Slack threads already produce signals, but they stay fragmented and vulnerability backlogs keep growing. Security Factory links three stages: signals, triage and remediation. Connectors consolidate findings from static analysis, dependency and secret scanners and infrastructure security tools, while Security Review applies STRIDE, OWASP and supply-chain methodologies to pull requests and full-codebase audits. Workstreams let agents split investigations into specialized tasks and run them in parallel, with durable state and shared context preserving evidence across long-running work. Triage output is validated findings showing what is affected, how it can be exploited and what an attacker could access, producing a backlog based on demonstrated risk rather than scanner severity. Remediation carries that evidence into Factory's coding workflows, including Automations, SDK, Droid CLI and the desktop application, where agents develop fixes, test them against the original exploit conditions and prepare patches or pull requests. Organizations can set autonomy levels, from presenting remediation plans for human approval to automatically creating pull requests and deploying fixes within approved policies.

### Amp — Boxes

- Company: Amp (ampcode.com)
- Announced: 2026-11-10
- Category: product-launch
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://ampcode.com/news/boxes
- Record: https://forck.live/items/34122-boxes
- Subject: Amp

Amp is introducing boxes, always-on computers where developers can deploy the software they build with its agents. The company says a box is a single VM that can host multiple applications, each with its own URL, and that it stays on rather than shutting down. Boxes are small and cheap, with the smallest priced at $18 per 30 days and billed by the minute, though users can schedule them on and off to save money. Amp argues that personal and internal software does not need to scale to thousands of users, so a small machine is enough to keep it available to the people you share with and within reach of an agent. Each box comes with its own runner, so agents can be spawned on the box to modify the state of running applications, and boxes can be shared with a workspace so teammates can deploy new versions. There are no containers or other abstractions involved, and deployments can be made private or require visitors to Sign In With Amp.

### Supabase — Set up Supabase for every Expo environment in one command

- Company: Supabase (supabase.com)
- Announced: 2026-11-10
- Category: developer-tool-release
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://supabase.com/blog/set-up-supabase-for-every-expo-environment-in-one-command
- Record: https://forck.live/items/34105-set-up-supabase-for-every-expo-environment-in-one-command
- Subject: Supabase

Expo built a Supabase integration into the EAS CLI so developers can create a Supabase project, install the SDK and write keys to every Expo environment with one command. The command, eas integrations:supabase:connect, replaces the old routine of creating a project in one dashboard, copying keys into another and checking that preview and production held the right values. Adding --environment preview gives Preview its own Supabase project with separate keys, leaving production data untouched, while Development can point at a local Supabase stack. Supabase Security and Performance Advisors run from the CLI and report findings such as a table with Row Level Security disabled or a foreign key without an index in the Expo build dashboard, next to the build logs for the change that introduced it. Every command runs without prompts and returns JSON, so an agent scaffolding an app can set up its backend in the same pass. Quilia, which builds a client app for personal injury law firms, uses a read-only Supabase MCP server and a development project to ship several fixes a day through Expo Update. The integration requires eas-cli 21.8.0 or newer and a project linked to Expo.

### Hugging Face — Decision Models in Hugging Face

- Company: Hugging Face (huggingface.co)
- Announced: 2026-11-10
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://huggingface.co/blog/decision
- Record: https://forck.live/items/34062-decision-models-in-hugging-face
- Subject: Platform

Hugging Face published an explainer on decision models, which take a state and a set of predefined answers and return a structured decision with probabilities instead of generating text. The post contrasts them with prompting an LLM for JSON, which can produce invalid output, unexpected keys or answers outside the allowed choices, and with constrained decoding libraries such as Outlines and XGrammar that enforce schemas but add computational work. It traces the term to TypeSafe AI's Jev, described as a System One model, and says the community settled on "decision model" because it names the model's computational role rather than a theory of cognition. The interface is given as state plus questions plus allowed answers producing decisions and probabilities, illustrated with a customer-support ticket routed to billing, technical or sales. The post also compares decision models with traditional classification, whose output classes are baked into the classification head, and with zero-shot classification, which takes candidate labels at inference time.

### Anthropic — Anthropic launches roadshow for proposed initial public offering

- Company: Anthropic (anthropic.com)
- Announced: 2026-11-10
- Category: funding-company-news
- Coverage: 13 outlets
- Announcement: yes
- Group: covered
- Source: https://www.anthropic.com/news/roadshow-launch
- Record: https://forck.live/items/33946-anthropic-launches-roadshow-for-proposed-initial-public-offering
- Subject: Claude

Anthropic, PBC launched the roadshow for a proposed initial public offering of its Class A common stock, with the offering priced at an expected $70.00 to $78.00 per share. The proposed offering consists of 825,000,000 shares, of which 801,351,354 are offered by Anthropic and 23,648,646 by certain existing stockholders, and Anthropic says it will receive no proceeds from the selling stockholders' shares. Anthropic expects to grant the underwriters a 30-day option to buy up to an additional 123,750,000 shares. The company has applied to list on the Nasdaq Global Select Market under the symbol "ANT." A registration statement has been filed with the U.S. Securities and Exchange Commission but has not yet become effective, so the securities may not be sold and offers to buy may not be accepted until it does. The offering will be made only by means of a prospectus, available from Goldman Sachs and Morgan Stanley, and the announcement is not an offer to sell or a solicitation to buy.

### Google Research — Exploring the world's languages with Language Explorer

- Company: Google Research (research.google)
- Announced: 2026-11-09T20:29:39+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://research.google/blog/exploring-the-worlds-languages-with-language-explorer/
- Record: https://forck.live/items/33641-exploring-the-world-s-languages-with-language-explorer
- Subject: Research

Google Research has built Language Explorer, an interactive web platform that lets researchers, educators and language enthusiasts visualize linguistic diversity around the world. The tool was developed as part of Google's 1,000 Languages Initiative and brings together data across thousands of languages, building on the open-source LinguaMeta dataset. It is featured as an interactive installation in a new exhibition at the Museum of Communication Berlin, Real Talk! Language. Power. World., which opened this week and runs through February 2028, with a Frankfurt run through September 2029. Google says the work reflects its broader effort to build AI and language technologies that serve everyone, regardless of what language they speak, read or sign. Those who cannot visit in person can explore the museum's themes through an online companion that started November 6th, or use Language Explorer directly online. Google traces its language work back to a 2006 Google Translate experiment with two languages, English and Arabic; it now supports nearly 250 languages and translates around a trillion words a month.

### Amazon — Single-shot and agentic retrieval with Amazon Bedrock Managed Knowledge Base

- Company: Amazon (amazon.com)
- Announced: 2026-11-09T20:13:18+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://aws.amazon.com/blogs/machine-learning/single-shot-and-agentic-retrieval-with-amazon-bedrock-managed-knowledge-base/
- Record: https://forck.live/items/33632-single-shot-and-agentic-retrieval-with-amazon-bedrock-managed-knowledge-base
- Subject: Bedrock / Nova

Amazon Bedrock Managed Knowledge Base now supports single-shot RAG and retrieval-only requests through AgenticRetrieveStream, letting applications match retrieval effort to each question. Single-shot RAG searches the knowledge bases once and uses a foundation model to generate an answer from the passages it finds, while retrieval-only requests return the matching passages without generating an answer. These modes join the existing agentic retrieval, where the service plans and runs multiple searches before answering, so the same API can serve both a simple checklist lookup and a cross-policy comparison. The maxAgentIteration range is now 0–5, and one iteration equals one foundation model call, with the loop always ending in a single answer-generation call. A value of n buys up to n minus 1 agentic planning rounds, each of which is one query rewrite plus a retrieval, so more iterations mean more planning rounds and fewer mean lower latency. All three modes support multiple knowledge bases, including maxAgentIteration=0 and 1, and the initial speculative retrieval that runs ahead of the first planning step never counts against the value.

### Amazon — Introducing Claude Fable 5.2 on AWS

- Company: Amazon (amazon.com)
- Announced: 2026-11-09T20:04:37+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://aws.amazon.com/blogs/machine-learning/introducing-claude-fable-5-2-on-aws/
- Record: https://forck.live/items/33624-introducing-claude-fable-5-2-on-aws
- Subject: Bedrock / Nova

Amazon is making Anthropic's Claude Fable 5.2 and Claude Mythos 5.2 available on Amazon Bedrock and Claude Platform on AWS, with Mythos 5.2 restricted to Anthropic's verification programs. Amazon describes Fable 5.2 as Anthropic's most capable generally available model for technical work, saying it raises the bar in coding and the sciences and can run unattended for hours to days. Anthropic's alignment testing found Mythos 5.2 hallucinates less, completes tasks more reliably and self-reports issues more honestly than Opus 5.5, Sonnet 5.5 and Mythos 5.1, and Anthropic expects Fable 5.2 to behave similarly. Declined cyber security queries can be retried on Opus 4.8 and declined biology queries on Opus 5, but on Bedrock the retry is the customer's code's job rather than an automatic switch. Anthropic designates Fable 5.2 a Covered Model, so by default its use requires retention for up to 30 days and human safety review, performed by Amazon personnel within the AWS boundary. Enterprise Frontier Safeguards, built with AWS, is available now and combines zero data retention with misuse detection, while a customer-controlled bucket and key version is expected to follow and has not launched; until then eligible customers can use Fable 5.2 with zero data retention through June 30, 2027.

### Amazon — NVIDIA Nemotron 3.5 Super VL is now available in Amazon SageMaker JumpStart

- Company: Amazon (amazon.com)
- Announced: 2026-11-09T20:01:20+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://aws.amazon.com/blogs/machine-learning/nvidia-nemotron-3-5-super-vl-is-now-available-in-amazon-sagemaker-jumpstart/
- Record: https://forck.live/items/33625-nvidia-nemotron-3-5-super-vl-is-now-available-in-amazon-sagemaker-jumpstart
- Subject: Bedrock / Nova

Amazon made NVIDIA's Nemotron 3.5 Super VL vision-language model available in SageMaker JumpStart, where it can be deployed from SageMaker Studio or the SageMaker Python SDK. The model pairs a 120B-parameter hybrid Mixture-of-Experts language backbone with 12B active parameters and a C-RADIO vision encoder, accepting images, video and text and generating text. It supports context lengths of up to 1 million tokens and a built-in Multi-Token Prediction head for speculative decoding, with reasoning configurable per request. NVIDIA built it for document intelligence, GUI understanding for computer-use agents, video question answering and multimodal agentic applications. JumpStart offers two variants: an NVFP4 checkpoint quantized from BF16 with quantization-aware distillation, and BF16 full-precision reference weights. NVFP4 runs on G7 instances with RTX PRO 4500 Blackwell GPUs and BF16 on P5en instances with H200 GPUs; the default NVFP4 instance is ml.g7.24xlarge. In Amazon's test, BF16 on ml.p5en.48xlarge reached 140.85 output tokens per second per request, with time to first token about half a second across all three configurations.

### Google DeepMind — Gemini Robotics 2.5 accelerates real-world general dexterity

- Company: Google DeepMind (deepmind.google)
- Announced: 2026-11-09T19:51:47+00:00
- Category: new-model
- Coverage: 3 outlets
- Announcement: yes
- Group: models
- Source: https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/
- Record: https://forck.live/items/34057-gemini-robotics-2-5-accelerates-real-world-general-dexterity
- Subject: Gemini

Google DeepMind introduced Gemini Robotics 2.5, a vision-language-action model it says is its most capable to date, widening the range of skills robots can learn and how fast they acquire them. The company evaluated the model across more than 200 distinct tasks, including serving food with tongs, tying a bow on a gift box and pipetting liquid, and says it reaches a 30% to 100% higher success rate than previous baselines while finishing tasks 20% to 30% faster. In a proof-of-concept on Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany, Apptronik's Apollo 2 humanoid used the model to pick and pack automotive parts with a 96% success rate. DeepMind says inference-time reasoning guidance steers the robot around physical hazards in real time, lifting safe task completion by 33% around obstacles and 43% around nearby people. The company is also releasing ASIMOV-Jailbreak, an open-source benchmark of hundreds of prompts that disguise unsafe requests across dozens of scenes, on which Gemini 4 Argon reaches 99.5% jailbreak resistance against an 82% baseline for Gemini Robotics ER 2. Gemini Robotics 2.5 is rolling out to the company's robotics partners.

### Runway — New in Runway, Sightline.

- Company: Runway (runway.com)
- Announced: 2026-11-09T19:27:34+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://x.com/runwayml/status/2119874555581694370
- Record: https://forck.live/items/33603-new-in-runway-sightline

Create a 360 panorama of any location, then frame your shots from any angle, with the environment staying consistent in every one. Export your panorama for 3D to use as an environment in your scenes.

### Amazon — Pluggable Agentic AI Framework: A seven-layer composable enterprise agentic AI reference architecture on Amazon Bedrock AgentCore

- Company: Amazon (amazon.com)
- Announced: 2026-11-09T19:05:54+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/pluggable-agentic-ai-framework-a-seven-layer-composable-enterprise-agentic-ai-reference-architecture-on-amazon-bedrock-agentcore/
- Record: https://forck.live/items/33589-pluggable-agentic-ai-framework-a-seven-layer-composable-enterprise-agentic-ai
- Subject: Bedrock / Nova

Amazon's AWS published a seven-layer composable reference architecture, the Pluggable Agentic AI Framework, for taking multi-agent AI applications from prototype to production scale. The framework addresses a gap the post says stalls most agentic AI initiatives: prototypes show a model can do a task but are not built to survive production traffic, cost pressure and security review. Each of the seven layers has a distinct responsibility and a stable contract, and the services shown reflect the reference implementation rather than mandatory choices, so organizations can substitute alternatives that fit their existing infrastructure. The Data and Knowledge layer owns retrieval, embeddings and access to structured and unstructured data, with Amazon Bedrock Knowledge Base, Vector Engine for Amazon OpenSearch Serverless, Amazon S3 Vectors and Amazon DynamoDB as example building blocks. Model and Inference choices should be driven by configuration rather than hardwired into application logic, with a routing layer and a proxy such as LiteLLM providing a common interface across providers. Orchestration should use a single agent for straightforward tasks and a supervisor delegating specialist agents for complex cross-domain workflows, with specialists discovered from a centralized agent registry through protocols like MCP and A2A. Security and Guardrails should be built in from day one, treating all external content as untrusted, authenticating every request through a service like Amazon Bedrock AgentCore Identity, and granting agents temporary scoped credentials instead of persistent secrets.

### Amazon — How CrossnoKaye grounds physical AI with Amazon Bedrock AgentCore

- Company: Amazon (amazon.com)
- Announced: 2026-11-09T18:55:43+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/how-crossnokaye-grounds-physical-ai-with-amazon-bedrock-agentcore/
- Record: https://forck.live/items/33576-how-crossnokaye-grounds-physical-ai-with-amazon-bedrock-agentcore
- Subject: Bedrock / Nova

CrossnoKaye built its ATLAS software-defined automation platform and AURA agentic intelligence system on Amazon Bedrock AgentCore to bring physical AI to heavy industry, connecting live plant data to closed-loop control of equipment. The company says the approach avoids the naming mismatches, proprietary controller logic and scattered credentials that make bolt-on AI agents fail in industrial settings. ATLAS already manages the facility model, base control software, historian and optimization applications in one platform on AWS, giving AURA native access to the plant's full operational state inside a single security boundary. Because LLMs cannot reliably compute statistics or physical balances over millions of sensor readings, AURA uses the code interpreter capability in Amazon Bedrock AgentCore to write and run Python workloads at runtime in isolated, AWS managed microVM sandboxes, alongside Amazon Bedrock Guardrails and local edge safety controls. CrossnoKaye reports that its more than 1,000 live advanced process control applications improve production throughput by 11–17 percent and cut total facility energy costs by an average of 21 percent. Across more than 100 industrial facilities and 13,000 controlled assets, it stores over 22 trillion machine data points, ingests more than 50 billion new readings a day, and reports 650,000 closed-loop control actions per day.

### Amazon — Professional work deserves professional tools. Meet AWS Quick.

- Company: Amazon (amazon.com)
- Announced: 2026-11-09T18:43:30+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/professional-work-deserves-professional-tools-meet-aws-quick/
- Record: https://forck.live/items/33572-professional-work-deserves-professional-tools-meet-aws-quick
- Subject: Bedrock / Nova

Amazon is launching the first brand campaign for AWS Quick, its AI workspace built only for work, running across the United States and the United Kingdom on streaming TV, online video, social media and business publications. The company says the campaign is meant to bring the workspace to millions more knowledge workers, and it names 3M, BMW, Mondelez, New York Life and Southwest Airlines as enterprises already using Quick. Amazon says Quick has no consumer version and was built for work from day one on AWS infrastructure, with agents, apps and automations that anyone can build by describing what they need in plain language. Quick integrates with Microsoft Purview, AWS Identity and Access Management and Microsoft Active Directory, reads existing sensitivity labels, enforces permissions and flags potentially sensitive data to administrators. Amazon says the product carries HIPAA, FedRAMP, SOC 2 and ISO 27001 certifications, that conversations are private and IT admins cannot access chat transcripts, and that data never leaves the customer's environment or trains other models. The post also marks the product's name change to AWS Quick, which Amazon says signals the enterprise security and governance that have been part of Quick from the start.

### AMD — @Microsoft's Project Zenith is now available on AMD Ryzen AI Halo, bringing a ready-to-code Windows 11 experience to local AI development.

- Company: AMD (amd.com)
- Announced: 2026-11-09T17:07:04+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://x.com/AIatAMD/status/2119839199364796708
- Record: https://forck.live/items/33523-microsoft-s-project-zenith-is-now-available-on-amd-ryzen-ai-halo-bringing-a

Less setup. More building.

### Meta — Extending LLDB to Debug Applications That Span CPUs and GPUs

- Company: Meta (meta.com)
- Announced: 2026-11-09T17:00:12+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://engineering.fb.com/2026/11/09/developer-tools/extending-lldb-to-debug-applications-that-span-cpus-and-gpus/
- Record: https://forck.live/items/33514-extending-lldb-to-debug-applications-that-span-cpus-and-gpus
- Subject: Llama / infrastructure

We built a vendor-neutral accelerator plugin architecture in lldb-server for debugging applications that span CPUs and GPUs. It separates shared debugger infrastructure from hardware-specific device control.

### Google Cloud — New survey finds enterprises trapped between legacy software and AI modernization

- Company: Google Cloud (cloud.google.com)
- Announced: 2026-11-09T17:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://cloud.google.com/blog/products/ai-machine-learning/new-survey-finds-enterprises-trapped-between-legacy-software-and-ai-modernization/
- Record: https://forck.live/items/33515-new-survey-finds-enterprises-trapped-between-legacy-software-and-ai
- Subject: Gemini Enterprise

Throughout my career in cybersecurity and infrastructure protection — both in government and now at Google Cloud — I have consistently seen that friction between new and existing technology is often a barrier to progress.

### GitHub — Claude Fable 5.2 in GitHub Copilot

- Company: GitHub (github.com)
- Announced: 2026-11-09T16:46:43+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://github.blog/changelog/2026-11-09-claude-fable-5-2-in-github-copilot
- Record: https://forck.live/items/33708-claude-fable-5-2-in-github-copilot
- Subject: Copilot

GitHub added Anthropic's Claude Fable 5.2 to Copilot for complex, multistep coding tasks across large codebases, available to Pro+, Max, Business and Enterprise users. In early testing, GitHub says Fable 5.2 solved more tasks than Claude Fable 5.1 while using fewer steps and tool calls, and completed hard tasks that earlier models missed. The model is billed at provider list pricing under usage-based billing, and rollout is gradual across Visual Studio Code, Visual Studio, Copilot CLI, the Copilot coding agent, the Copilot app, github.com, GitHub Mobile, JetBrains IDEs, Xcode and Eclipse. Copilot Enterprise and Business administrators must enable the Claude Fable 5.2 policy in Copilot settings, which is off by default, and leaving it off keeps the model unavailable to the organization. Fable 5.2 requires data retention by default: Anthropic retains prompts and outputs to run safety classifiers that detect harmful use, and says retained data is not used to train its models. Other Claude models, except Fable 5, Fable 5.1 and Fable 5.2, continue under zero data retention. Eligible customers can use Fable 5.2, 5.1 and 5 with zero data retention through the end of the calendar year under a time-bound exemption while Anthropic rolls out Enterprise Frontier Safeguards; after that, continued use will require EFS.

### Harvey — Keep Document Reviews Up to Date With Review Table Sync

- Company: Harvey (harvey.ai)
- Announced: 2026-11-09T16:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.harvey.ai/blog/review-table-sync
- Record: https://forck.live/items/33492-keep-document-reviews-up-to-date-with-review-table-sync
- Subject: Harvey

Harvey is adding Review Table Sync, a capability that flags when new documents land in a connected Vault and lets legal teams fold them into an existing Review Table with one click. Review Tables extract standardized information from many documents at once, one row per document and one column per question, and until now new files meant manual reconciliation. Clicking Sync adds the new documents as rows and runs the existing review questions against them, while rows the team has already reviewed stay unchanged. Harvey describes three uses: diligence on a buy-side deal where a data room such as Ansarada keeps growing, a consumer group action where claim forms arrive in waves and need triage against limitation deadlines, and in-house contract review where questions about breach notice, liability caps and logo rights recur. Review Table Sync is available today in Harvey; teams open any Review Table using Vault files, click settings, then click Sync.

### NVIDIA — Nemotron-Personas-Finance-Japan: A Synthetic Dataset Capturing Japan's Demographics and Financial Context

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-09T15:05:57+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://huggingface.co/blog/nvidia/nemotron-personas-finance-japan
- Record: https://forck.live/items/33458-nemotron-personas-finance-japan-a-synthetic-dataset-capturing-japan-s

Elith and NVIDIA have released Nemotron-Personas-Finance-Japan, an open synthetic persona dataset for Japanese financial AI, built from official government statistics rather than real customer data. The companies say it is the first open persona dataset at million-record scale pairing Japan's official statistics with household finance and financial literacy data. It holds 9M personas, 1M records with nine personas each, written in natural Japanese across 50 fields per record and roughly 1.5B tokens, including about 750M persona tokens. The dataset was designed with feedback from domestic financial institutions and industry associations, among them Sompo Japan, Tokio Marine Holdings, Bank of Fukuoka, Mizuho Financial Group, Meiji Yasuda Life Insurance and the Financial Data Utilization Association. It was generated with NVIDIA's NeMo Data Designer and NeMo SDG-PGMs, using gemma-4-31B-it for Japanese text generation, and integrates with the NeMo User Sim simulation framework. Released under the CC BY 4.0 license, it is available for commercial and non-commercial use, and NVIDIA says every persona is fully synthetic, with no real individuals represented.

### NVIDIA — Practice Makes Precise: At CoRL, NVIDIA and Partners Reveal How Robots Rehearse Delicate Work in Simulation

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-09T14:00:01+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://blogs.nvidia.com/blog/robotics-simulation-advancements-corl/
- Record: https://forck.live/items/33434-practice-makes-precise-at-corl-nvidia-and-partners-reveal-how-robots-rehearse
- Subject: AI platform

NVIDIA used the CoRL conference in Austin to introduce open robotics technologies for manipulation, including the GR00T 2 world action model, simulation tools, benchmarks and agentic workflows. GR00T 2 is an open frontier world action model built on NVIDIA Cosmos and trained on diverse robot and human data, including 100,000 hours of NVIDIA open source human egocentric data. NVIDIA says it completed 40 of 50 evaluated tasks at least once without task-specific demonstrations, and reached 71% average success across 50 tasks with one demonstration per task, rising to 95% with five. The company says GR00T 2 currently ranks among the top three models across all vision language action and world action model categories for the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks, with multiple leaderboard wins. Nexuni is testing it for robotic laundry handling, Foxconn is applying it to factory automation for server manufacturing, and Siemens is evaluating it across industrial use cases. NVIDIA also released Isaac Lab 3.0, which lets researchers set up training environments with agents, switch between PhysX and Newton solvers, and apply a reinforcement post-training library to foundation models like GR00T.

### AMD — AMD EPYC Expands Its Leadership at the Foundation of AI

- Company: AMD (amd.com)
- Announced: 2026-11-09T13:00:00+00:00
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ir.amd.com/news-events/press-releases/detail/1304/amd-epyc-expands-its-leadership-at-the-foundation-of-ai
- Record: https://forck.live/items/33407-amd-epyc-expands-its-leadership-at-the-foundation-of-ai
- Subject: AMD Instinct / ROCm

AMD says its 5th Gen EPYC processors now power the host infrastructure for Google Cloud TPU 8tx, which it calls the first cloud TPU system with hardware-enforced confidential computing. The two companies are combining EPYC host processing, AMD Infinity Guard security and x86 software compatibility with Google Cloud's TPU acceleration so enterprises can run inference, fine-tuning and reinforcement learning workloads. In TPU 8tx systems, the EPYC CPUs handle workload orchestration, data preparation and high-throughput communication with the TPU accelerators. AMD's Secure Encrypted Virtualization, together with the TPU, extends confidential computing protection beyond the CPU to connected accelerators, keeping prompts, model weights and intermediate data encrypted in CPU memory and in transit. AMD says EPYC is the host CPU of choice for more than 10 leading hyperscalers and frontier AI labs, across GPUs and custom accelerator ASICs. The companies have collaborated across multiple generations of EPYC, including Google Cloud Compute Engine C4D, N4D, H4D, Z4D and G4 virtual machines.

### Google Cloud — GKE dynamic slicing for higher resilience, and efficiency for TPUs

- Company: Google Cloud (cloud.google.com)
- Announced: 2026-11-09T13:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://cloud.google.com/blog/products/containers-kubernetes/gke-dynamic-slicing-on-tpus/
- Record: https://forck.live/items/33406-gke-dynamic-slicing-for-higher-resilience-and-efficiency-for-tpus
- Subject: Gemini Enterprise

Running large-scale AI workloads on TPU and Google Kubernetes Engine (GKE) requires infrastructure that adapts as fast as your models change. To help, dynamic slicing decouples logical TPU slice formation from physical nodepool provisioning.

### Amazon — Your Experts Have Partners. There's HOPE for Keeping Them Together

- Company: Amazon (amazon.com)
- Announced: 2026-11-09T12:34:29+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/amazon/higher-order-pruning-experts
- Record: https://forck.live/items/33400-your-experts-have-partners-there-s-hope-for-keeping-them-together

Amazon's new paper derives HOPE, a second-order expert-pruning method that accounts for pairwise interactions between experts in Mixture-of-Experts language models. Existing pruning methods score each expert independently, assuming its contribution is independent of all others, but the authors say experts form stable cooperative groups that get consistently co-activated together. Ignoring those interactions leads to suboptimal pruning, especially at aggressive rates, because deleting one expert can dismantle a cooperative group. At 50% pruning, HOPE achieves the best average rank of any method, with gains of up to +6.1% on agentic coding tasks over the next-best pruning method. The authors illustrate the memory problem with Qwen3.5-122B, which has 256 experts in each of its 48 MoE layers and uses only 8 experts per layer for any given token, with expert MLPs taking up 95% of its ~250 GB parameter memory footprint. HOPE defines a tractable surrogate that upper-bounds the dominant component of the pruning error, and minimizing it over a calibration set becomes a quadratic program. The post explains how HOPE works, where it outperforms existing methods, and how to apply it to other models.

### ElevenLabs — ElevenAgents Campaign Manager works every contact to an outcome

- Company: ElevenLabs (elevenlabs.io)
- Announced: 2026-11-09T12:00:00+00:00
- Category: product-launch
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://elevenlabs.io/blog/elevenagents-campaign-manager
- Record: https://forck.live/items/33455-elevenagents-campaign-manager-works-every-contact-to-an-outcome
- Subject: Voice AI

ElevenLabs is launching ElevenAgents Campaign Manager, a tool that runs outbound calling campaigns toward a stated outcome such as a qualified lead or confirmed appointment. Given contacts and a goal, the agent calls within set hours, carries context between interactions, and decides whether to retry, book a requested time, or move to another number or person on the account. It stops once the outcome is reached or a customer asks to stop. The product has been in beta with customers in insurance, banking and ecommerce; one insurance provider says it now reaches 100% of its daily leads, up from 50 to 60%. Each campaign shows a funnel from enrollment to outcome, with a per-contact decision timeline, a needs-attention queue and CSV export. Controls include per-campaign calling windows, attempt limits with a separate callback cap, do-not-call lists that update when a contact opts out, and disclosures set in the agent's opening. Campaign Manager uses existing ElevenAgents agents and phone numbers, enrolls contacts by CSV upload or API, and works over phone calls, WhatsApp voice and WhatsApp messaging, with email and SMS coming soon. It is available now under Campaigns in the Agents sidebar.

### Lovable — The Radish AI team joins Lovable to scale agentic capabilities for businesses

- Company: Lovable (lovable.dev)
- Announced: 2026-11-09T12:00:00+00:00
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://lovable.dev/blog/radish-ai-team-joins-lovable
- Record: https://forck.live/items/33393-the-radish-ai-team-joins-lovable-to-scale-agentic-capabilities-for-businesses
- Subject: Lovable / AI app builder

Lovable has acquired the team behind Radish AI, an agentic infrastructure platform, to scale the agent capabilities and infrastructure behind its business software. The company says more of its millions of users are running and growing businesses on Lovable, and that increasingly complex software demands AI agents that reliably carry out business tasks. Radish AI built a platform for building, deploying and scaling complex AI agents quickly and reliably, with privacy and security built in. Three named hires join Lovable's engineering team: Nirman Gupta, a former engineering leader at Meta; Ish Singh, a two-time founder who led AI and data products at Google, JPMorgan and Goldman Sachs; and Adrian Zgorzałek, who spent 12 years at Meta on developer experience and infrastructure. Lovable says their experience will shape the agent capabilities and infrastructure behind its platform, helping businesses automate operations on software they build themselves. The team will be split between Lovable's Stockholm and London offices, with Gupta leading engineering on agents for businesses from the newly opened London office.

### Lovable — How we moved our frontend off Firestore and built Lovable Realtime

- Company: Lovable (lovable.dev)
- Announced: 2026-11-09T11:30:00+00:00
- Category: infrastructure-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://lovable.dev/blog/lovable-realtime-service
- Record: https://forck.live/items/33385-how-we-moved-our-frontend-off-firestore-and-built-lovable-realtime
- Subject: Lovable / AI app builder

Lovable built its own realtime service, Lovable Realtime, to move its frontend off direct Firestore listeners and put authorization and payload shaping in its Go backend. The company says the browser previously received whole Firestore documents, so adding a database field could expose it, and permissions had to be kept in sync between Firestore rules and the Go API. Firestore rules also could only query Firebase data, which blocked authorization updates for large groups such as enterprise SCIM-synced engineering groups whose data lives in Spanner. Lovable says it built the core in 40 hands-on working days and rolled it out to every user without a single user-facing incident. The service handles peaks of over 2 million concurrent subscriptions and delivers around 2.6 billion data and patch frames per month. It is a dedicated WebSocket service with one connection per tab and all subscriptions multiplexed over it, and it works with the existing Firestore and Spanner stores rather than requiring a new primary database.

### Legora — Agiloft and Legora integrate to bring data-first contract intelligence into AI-powered work for enterprises

- Company: Legora (legora.com)
- Announced: 2026-11-09T09:00:00+00:00
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://legora.com/newsroom/agiloft-and-legora-integrate-to-bring-data-first-contract-intelligence-into-ai-powered-work-for-enterprises
- Record: https://forck.live/items/33437-agiloft-and-legora-integrate-to-bring-data-first-contract-intelligence-into-ai
- Subject: Legora

Agiloft and Legora announced a strategic partnership to integrate Agiloft's contract lifecycle management platform with Legora's AI environment for legal work. The companies say legal teams often move between separate legal technology applications, which fragments their workflows. Mutual customers will be able to find Agiloft agreements in Legora through plain-language searches and draw on structured contract data such as counterparties, key dates and obligations. Users will also be able to save work from Legora back to Agiloft as a new version, a separate document or a replacement for the original, without manual downloads and uploads. Agiloft permissions will govern which agreements users can retrieve through Legora and whether they can save or overwrite documents in Agiloft. The integration will be rolled out in phases, starting with the availability of Agiloft agreements in Legora, with later release timing shared with mutual customers.

### NVIDIA — NVIDIA Announces New Planned Investments in Korean Startups at the MSIT-NVIDIA AI Startup Partnership Forum

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-09T08:00:53+00:00
- Category: funding-company-news
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://blogs.nvidia.com/blog/korean-startup-investments-msit/
- Record: https://forck.live/items/33312-nvidia-announces-new-planned-investments-in-korean-startups-at-the-msit-nvidia
- Subject: AI platform

NVIDIA announced an investment in Korean AI interconnect startup Point2 and said it plans further investments in Motif Technologies, RLWRLD and Upstage, disclosed at a forum it hosted with Korea's Ministry of Science and ICT. The forum at COEX brought NVIDIA leaders and vice minister Ryu Je-myung together with more than 20 Korean startups working across foundation models, physical AI, AI infrastructure and AI semiconductors. Point2 builds radio frequency-based interconnects that move data between chips and racks in AI factories, which NVIDIA describes as an alternative to conventional copper and optical links for scaling AI clusters. Motif Technologies, based in Seoul, develops foundation models for AI agents and released the open-weight large language model Motif-3. RLWRLD, with offices in San Francisco, Seoul and Tokyo, develops robotics foundation models for humanoid and industrial robots doing tasks that need five-finger dexterity, such as inserting components or threading nuts onto bolts. Upstage develops Solar language models and document AI tools for extracting information from complex documents. NVIDIA says the planned investments build on its work with Korea's startup community, including more than 1,000 Korea-based members of its Inception program, which offers preferred pricing, technical expertise and investor introductions.

### OpenAI — CodeRabbit scales agentic software development and ships 20x more

- Company: OpenAI (openai.com)
- Announced: 2026-11-09T08:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/coderabbit-codex-daybreak
- Record: https://forck.live/items/33678-coderabbit-scales-agentic-software-development-and-ships-20x-more
- Subject: GPT / ChatGPT / API

CodeRabbit uses OpenAI models to govern and trust AI-generated code and Codex to ship 20 times more software with the same engineering team.

### Google Antigravity — Google Antigravity for Knowledge Work

- Company: Google Antigravity (antigravity.google)
- Announced: 2026-11-09T08:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://antigravity.google/blog/antigravity-for-knowledge-work
- Record: https://forck.live/items/33623-google-antigravity-for-knowledge-work
- Subject: Antigravity

Google added a Knowledge Work interface layout to its Antigravity app, giving non-developers a view without Git worktrees, terminal panes and line-by-line code diffs while running the same agent. Google says tens of thousands of Googlers outside software engineering, across sales, finance, legal, product, design, research, marketing and operations, already use Antigravity every week. Users pick between the Knowledge Work and Software Engineering layouts in Settings under Interface Layout, and the switch applies immediately to open and new conversations without a restart. The Knowledge Work layout creates folder-backed projects and standalone conversations without Git worktree prompts, and hides branch pickers and IDE handoff buttons. It replaces expandable tool-call groups, inline code diffs and subagent cards with a live progress line, and shows plain-language summaries on permission prompts. Both layouts keep access to frontier models, local files, plugins, skills, rules, custom agents and MCP servers, and a new role-based onboarding flow tailors the starting layout, prompts and recommended plugins.

### NVIDIA — Korea Advances Physical AI Across Industries With NVIDIA

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-09T05:00:40+00:00
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://blogs.nvidia.com/blog/korea-physical-ai-ecosystem/
- Record: https://forck.live/items/33256-korea-advances-physical-ai-across-industries-with-nvidia
- Subject: AI platform

NVIDIA is expanding collaborations with Korean manufacturers and mobility companies to move physical AI from development into production, spanning semiconductor engineering, autonomous driving and robotics. Samsung Electronics is working with NVIDIA cuISS and cuDSS to accelerate sparse linear system computations in technology computer-aided design, and its sMOL platform has run selected quantum chemical calculations up to 5x faster end to end with NVIDIA cuEST. In one AI homogenization case, combining AI-predicted material properties, coarser grids and GPU acceleration produced a roughly 1,600x speedup, cutting simulation time from about 18 hours on a 32-core CPU to 41 seconds. Kakao Mobility plans to build a level 4 autonomous driving system using NVIDIA Hyperion, powered by DRIVE AGX Thor, and NVIDIA Halos for safety, with initial development planned around Kia purpose-built vehicles. NVIDIA and LG Electronics demonstrated Isaac GR00T-powered robotic manipulation and manufacturing tasks with LG's CLOiD robot in an actual manufacturing environment, using a physical AI data flywheel that connects real-world data collection, Cosmos-based augmentation, GR00T training, Jetson AGX Thor computing and simulation-based evaluation.

### NVIDIA — Korea’s AI Ecosystem Advances and Secures Agentic AI With NVIDIA

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-09T05:00:39+00:00
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://blogs.nvidia.com/blog/korea-secure-agentic-ai/
- Record: https://forck.live/items/33257-korea-s-ai-ecosystem-advances-and-secures-agentic-ai-with-nvidia
- Subject: AI platform

NVIDIA is working with Korean partners to secure agentic AI, as Hyundai Motor Group explores NVIDIA NemoClaw blueprints and the NVIDIA OpenShell runtime. AIM Intelligence, a Korean AI security company, is integrating its Starfort agent control platform with NVIDIA OpenShell, combining OpenShell's runtime access controls with Starfort's detection of malicious instructions and data leakage. OpenShell is a key component of the NVIDIA Open Agent Safety Platform and provides a secure runtime with sandboxed execution and policy-based controls to safeguard sensitive data, manage access to enterprise tools and govern autonomous agent actions. AIM Intelligence is also considering a preliminary proof of concept for the public energy sector, with potential applications across enterprise, industrial and robotics environments. On the model side, LG AI Research is opening API access for K-EXAONE-2.5, which was developed using NVIDIA Nemotron datasets, including Nemotron-Personas-Korea. That dataset holds 7 million fully synthetic Korean personas grounded in official Korean statistics, which NVIDIA says model builders use to ground training data and evaluation in Korean contexts.

### NVIDIA — From Chips to Skills, NVIDIA and Korea Expand Sovereign AI Collaboration

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-09T05:00:04+00:00
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://blogs.nvidia.com/blog/korea-sovereign-ai-collaboration/
- Record: https://forck.live/items/33258-from-chips-to-skills-nvidia-and-korea-expand-sovereign-ai-collaboration
- Subject: AI platform

NVIDIA and Korea's Ministry of Science and ICT are expanding their collaboration across the country's sovereign AI ecosystem, from AI semiconductors and data center infrastructure to talent, startups, models, agents and security. The work lays the groundwork for Korean NPU and XPU companies to explore participation in the NVIDIA NVLink Fusion ecosystem. It also includes NVIDIA training and certification programs, hackathons and technical seminars intended to strengthen GPU utilization. Through Korea's AI for All initiative, NVIDIA will work with three selected consortia and four leading Korean foundation model builders, using NVIDIA Dynamo software to optimize and scale AI inference. NVIDIA separately announced a collaboration with SK Telecom and Rebellions, supported by MSIT, to advance sovereign AI infrastructure for Korean models and workloads with NVIDIA NVLink Fusion.

### Vercel — Decision models take 26% of requests; Anthropic now wins new teams 2:1

- Company: Vercel (vercel.com)
- Announced: 2026-11-09T05:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://vercel.com/blog/ai-gateway-production-index-november-2026
- Record: https://forck.live/items/33585-decision-models-take-26-of-requests-anthropic-now-wins-new-teams-2-1
- Subject: v0 / AI SDK / AI Gateway

Vercel's AI Gateway Production Index for November 2026 reports that decision models grew to 26% of gateway requests in six weeks while taking just 0.53% of total spend. Decision models classify, score or pick between choices instead of generating text, and cost about a seventieth as much per request as a language model. They arrived on AI Gateway in mid-September, and 15 are now available from eight vendors; nearly a third of paid teams used one in October. Almost all of that request volume went to Jev from TypeSafe AI, which alone received nearly four times Anthropic's total requests. Adoption tracks team size: just over one in five teams under 10 million tokens used a decision model, against seven in ten above 10 billion tokens. Voice adoption also rose, with about 11% of paid teams using a voice model in October, up from 4% in July. OpenAI and Anthropic now reach about the same share of paid teams, and new teams in October picked Anthropic nearly two to one.

### NVIDIA — NVIDIA Dynamo Recipes Bring Korea’s Leading Open Models to Production

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-09T05:00:00+00:00
- Category: developer-tool-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://blogs.nvidia.com/blog/nvidia-dynamo-korea-open-models/
- Record: https://forck.live/items/33259-nvidia-dynamo-recipes-bring-korea-s-leading-open-models-to-production
- Subject: AI platform

NVIDIA is releasing Dynamo-optimized deployment recipes for four leading open models from Korea: SK Telecom's A.X K2, LG AI Research's K-EXAONE 2.0, Upstage's Solar Open 2 and Motif Technologies' Motif-3. The recipes are meant for models that no longer fit on a single GPU and must be split across several, where routing and communication have to be coordinated to keep performance. Dynamo is an open source, distributed inference-serving framework that NVIDIA describes as built for multi-node deployments at data center scale, handling that routing and communication itself. The recipes are optimized for NVIDIA B200 GPUs and, according to NVIDIA, deliver up to 2x performance out of the box while removing guesswork from deployment. NVIDIA says the recipes are intended to help developers bring Korea's open models into production faster and turn them into broadly accessible AI for All services.

### Harvey — Harvey and Wolters Kluwer Australia Expand Partnership to Bring CCH iKnowConnect Content into Harvey

- Company: Harvey (harvey.ai)
- Announced: 2026-11-09
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.harvey.ai/blog/harvey-and-wolters-kluwer-australia-expand-partnership-to-bring-cch-iknowconnect-content-into-harvey
- Record: https://forck.live/items/33713-harvey-and-wolters-kluwer-australia-expand-partnership-to-bring-cch
- Subject: Harvey

Harvey and Wolters Kluwer Australia expanded their existing partnership so mutual customers can use CCH iKnowConnect subscription content inside Harvey across Australia, New Zealand, Singapore and Malaysia. The expansion adds secondary material to the primary law content already in Harvey, including editorial commentary, case headnotes and practical guidance. Subscribers can select CCH as a knowledge source in Harvey and query their subscribed content alongside their firm's internal materials, seeing the same practice areas they license in CCH iKnowConnect. Responses draw on CCH iKnowConnect materials and include links back to the source material, and CCH says it maintains the content with timely updates and context. Megan Mulia, Managing Director of Wolters Kluwer Tax & Accounting APAC, said customers rely on CCH iKnowConnect for authoritative legal and tax research and can now bring that content into the AI workflows they already use. Harvey says the partnership reflects its commitment to APAC, where it opened a Sydney office followed by Singapore and has announced plans for a Tokyo office in early 2027.

### Anthropic — JPMorganChase expands Claude Code to more employees across the firm

- Company: Anthropic (anthropic.com)
- Announced: 2026-11-09
- Category: availability-change
- Coverage: 2 outlets
- Announcement: yes
- Group: covered
- Source: https://claude.com/resources/articles/jpmorganchase-expands-claude-code
- Record: https://forck.live/items/33697-jpmorganchase-expands-claude-code-to-more-employees-across-the-firm
- Subject: Claude

JPMorganChase is expanding Claude Code beyond engineering to product teams, advisors and operations staff across the firm, after a year in which its use of Claude grew from software engineering to cyber defense and financial-advisor tools. In the Corporate and Investment bank, work reaches production faster and the firm redeploys that capacity against its business priorities, while infrastructure and operations teams use Claude Code for monitoring, alerting and root cause analysis. The security team uses Anthropic's frontier models to find and fix vulnerabilities faster, work first described in Project Glasswing. Product teams turn requirements directly into engineering backlogs, and other teams use Claude Code for risk and controls assessments and program planning. In Asset & Wealth Management, Claude Code is now part of the engineering workflow, and teams are moving faster while putting every change through more rigorous testing before it reaches clients. Claude Code runs inside Devspace, the firm's containerized development platform, on its own infrastructure rather than developer desktops, and agents get no direct access to employee credentials or internal systems. Access widens in stages, new Claude models pass the firm's own approval process, and an automated gate returns a hard pass or halt that the pipeline enforces, so Claude cannot sign off on its own work.

### Anthropic — Introducing Claude Fable 5.2 and Claude Mythos 5.2

- Company: Anthropic (anthropic.com)
- Announced: 2026-11-09
- Category: model-update
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://www.anthropic.com/claude-fable-and-mythos-5-2
- Record: https://forck.live/items/33590-introducing-claude-fable-5-2-and-claude-mythos-5-2
- Subject: Claude

Anthropic upgraded Claude Fable and Claude Mythos to version 5.2, saying the models improve on coding, reasoning and vision and generate outputs up to 30% faster than Fable 5.1 and Mythos 5.1 at a lower price. Fable and Mythos are the same underlying model, but Fable carries cybersecurity and biology safeguards that let Anthropic offer it to all users, while organizations working on beneficial cybersecurity or biomedical research can apply for Mythos access through the Life Sciences and Cyber Verification Programs. Standard pricing is $8 per million input tokens and $40 per million output tokens, and an optional fast mode that speeds token generation by up to 2.5 times costs $16 per million input tokens and $80 per million output tokens. Anthropic says prerelease safety testing included new evaluations for behaviors such as sandbox escapes, that high-stakes misalignment risks remain low, and that Fable 5.2's safeguards match Fable 5.1's while producing fewer unnecessary blocks. On the FrontierCode 2.0 evaluation, Fable 5.2 at High effort scored 45.3% against Opus 5.5's 45.5% for roughly half the cost per task, and at Xhigh effort reached 49.1%, the highest of any released model Anthropic tested. Anthropic recommends most coding and knowledge work start with Claude Opus 5.5, switching up to Fable or Mythos or down to Sonnet 5.5 or Haiku 5.5 as needed.

### Meta — Introducing Our Smallest, Most Efficient Model: Muse Ember

- Company: Meta (meta.com)
- Announced: 2026-11-09
- Category: new-model
- Coverage: 18 outlets
- Announcement: yes
- Group: models
- Source: https://research.meta.ai/blog/muse-ember
- Record: https://forck.live/items/33534-introducing-our-smallest-most-efficient-model-muse-ember
- Subject: Muse

Meta introduced Muse Ember, its smallest and most efficient model, built for the tool calls, coding subtasks, routing, extraction and long document parsing that agents repeat all day. Meta says Muse Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting, matching Haiku's best score at about a third of the cost and beating every Luna setting. The model is available today on the Meta Model API and Muse Code as muse-ember-1.0, priced at $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. It carries a 1M-token context window and up to 512K output tokens, and supports the Responses API. Muse Ember natively understands text, images, audio and video, so agents can read screenshots, charts and scanned documents or work from voice memos and meeting recordings without a separate model. Meta says it evaluated Muse Ember under its Meta Superintelligence Scaling Framework before release, including chemical and biological, cybersecurity and loss-of-control risks, and found it remained below the framework's high-risk thresholds. Muse Ember is the first model trained using Muse Wonder Full, and Meta positions it alongside Muse Wonder for the hardest problems and Muse Spark for coding and agentic workflows, at about $0.14 per Artificial Analysis task.

### Amp — Portals on Your Runner

- Company: Amp (ampcode.com)
- Announced: 2026-11-09
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://ampcode.com/news/runner-portals
- Record: https://forck.live/items/33489-portals-on-your-runner
- Subject: Amp

Amp is extending Portals, previously available in orbs, to run on users' local machines, so a dev server or website started on that machine can be opened from the thread's Portal tab from anywhere. Users turn the feature on with Allow Portals under Settings › Runner in the macOS app, or by starting their runner with amp --no-tui --remote-control-portals. Portals on a personal runner are accessible only to that user. To let others in, a shared runner can be started with --share, or Share With Workspace can be turned on in the macOS app, which lets everyone in the workspace open its portals.

### Harvey — Addleshaw Goddard and Harvey launch firmwide collaboration to shape AI-enabled legal delivery

- Company: Harvey (harvey.ai)
- Announced: 2026-11-09
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.harvey.ai/blog/addleshaw-goddard-and-harvey-launch-firmwide-collaboration-to-shape-ai-enabled-legal-delivery
- Record: https://forck.live/items/33356-addleshaw-goddard-and-harvey-launch-firmwide-collaboration-to-shape-ai-enabled
- Subject: Harvey

Addleshaw Goddard has formed a firmwide collaboration with Harvey, the legal AI platform, to deploy Harvey across its workforce and co-develop legal AI capabilities. From January, the London-headquartered firm will provide Harvey to more than 3,400 lawyers and business services professionals across Europe, the Middle East and Asia. The deal follows a three-month assessment by AG's 80-strong Innovation Group, in which 400 partners, lawyers, legal technologists and business services professionals tested leading platforms on real legal and business tasks. AG's Innovation Group will work directly with Harvey's product and legal engineering teams, bringing in lawyers and clients to shape and test priority capabilities and create new AI-enabled services. The collaboration has already shaped Harvey's Review Tables, with AG feedback improving source checking, preserving reviewed work, sharing and performance, and giving the firm early access to AI-led features for creating and updating tables at scale. AG and Harvey plan to convene quarterly with selected clients to identify shared challenges and co-create solutions where appropriate. Managing Partner Andrew Johnston called it "the most significant technology transformation in AG's long history as an innovator," while Harvey CEO Winston Weinberg said the firms will work together to "shape what comes next."

### Sarvam AI — Sashikumar Sreedharan joins Sarvam as President and Chief Business Officer

- Company: Sarvam AI (sarvam.ai)
- Announced: 2026-11-09
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.sarvam.ai/blogs/sashikumar-sreedharan-joins-sarvam/
- Record: https://forck.live/items/33307-sashikumar-sreedharan-joins-sarvam-as-president-and-chief-business-officer
- Subject: Sarvam / Bulbul / Saaras

Today, we’re welcoming Sashikumar as President and Chief Business Officer and a member of our board of directors. Sashi will lead our business across enterprises and government, and expand our work with digital native companies, developers and partners.

### LG AI Research — K-EXAONE 2.5 API Available for Free

- Company: LG AI Research (lgresearch.ai)
- Announced: 2026-11-09
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.lgresearch.ai/news/view?seq=755
- Record: https://forck.live/items/33217-k-exaone-2-5-api-available-for-free
- Subject: EXAONE

LG AI Research is making the API for K-EXAONE 2.5, its proprietary AI foundation model, available for free. K-EXAONE 2.5 is a technical validation model that has undergone additional training based on the version released in last August.
