# forck.live — 16–22 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 the week in progress: every confirmed first-party
announcement of ISO week 2026-W47, 16–22 November 2026, newest first. The
window never moves, so this URL always names the same seven days — but they
have not all happened yet, so the file is still filling: announcements are
appended as they land, and it will answer with more of them later today. Fetch
it again rather than reading a cached copy as the finished week.

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-W47
- The current issue, and the catalogue of the other weeks: https://forck.live/briefing.md
- Catalogue: https://forck.live/llms.txt

## Announcements

### ElevenLabs — How the Nordics' leading postal operator rebuilt its customer support on ElevenAgents

- Company: ElevenLabs (elevenlabs.io)
- Announced: 2026-11-17
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://elevenlabs.io/blog/postnord
- Record: https://forck.live/items/36629-how-the-nordics-leading-postal-operator-rebuilt-its-customer-support-on
- Subject: Voice AI

ElevenLabs has published a case study of how PostNord, the Nordics' leading postal operator, replaced its Swedish inbound phone support with a multi-agent system built on ElevenAgents. PostNord's previous AI phone system had plateaued at about 45% of calls resolved without escalation, which the company treated as a signal to move again. The new system uses an orchestrator that identifies each caller's need and routes it to one of four specialized subagents, including a dedicated tracking agent, so common enquiries run on a simpler, lower-cost setup. Task completion rose from 45% to 60%, and the previous system was fully decommissioned in roughly 90 days. The agent speaks and understands natural Swedish across regional dialects, answers around the clock, and is backed by A/B testing that splits live traffic between model configurations, so ElevenLabs says 60% is a current reading rather than a final number. The rollout ran in four phases, from an isolated weekend test through a hybrid phase with human handover to full production, and Nordic expansion is underway for Denmark, Norway and Finland.

### Upstage — Upstage Announces Strategic Partnership with Cohere

- Company: Upstage (upstage.ai)
- Announced: 2026-11-17
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.upstage.ai/news/upstage-cohere-partnership
- Record: https://forck.live/items/37027-upstage-announces-strategic-partnership-with-cohere
- Subject: Solar Pro / Solar Mini

Upstage and Cohere signed a strategic partnership to collaborate on model development for sovereign AI and to serve Korean and global enterprise markets. The agreement was formalized at a signing ceremony at Upstage's office in Gangnam-gu, Seoul, attended by Upstage CEO Sung Kim and Cohere co-founder and CEO Aidan Gomez. Under the deal, Upstage will integrate its Solar Pro and Solar Mini large language models and its Document Parse document processing AI into North, Cohere's agentic AI platform for enterprises, and the two will offer the combined solution to customers. The solution will be available on premises and via cloud VPC for enterprises and institutions with stringent security and regulatory requirements. Cohere North lets enterprises build and operate AI agents and workflow automation within their own infrastructure, and Upstage says its models are strong in Korean language and document-intensive tasks. The companies describe the deal as the first collaboration between leading sovereign AI companies from Korea and Canada, and plan to expand joint business and technology work across Asia Pacific and beyond.

### Ideogram — How to use Ideogram’s Print on Demand Studio: from niche to mockup

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-17
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/print-on-demand-studio
- Record: https://forck.live/items/36886-how-to-use-ideogram-s-print-on-demand-studio-from-niche-to-mockup
- Subject: Ideogram image models

Ideogram's guide walks through its Print on Demand Studio, a workspace inside Ideogram for people who sell designs on T-shirts, mugs, stickers and posters, using a varsity running-club T-shirt as the worked example. The Studio opens from Print on Demand in the left sidebar and has three views: Home for templates and trending designs, Mockups for product mockup templates, and Research for what sells on Amazon. Research covers seven Amazon marketplaces and lists best sellers rank, 90-day average rank and price, filterable by keyword, marketplace, product type and date published. A niche analysis for "run club" T-shirts in the United States sampled 50 listings, 45 with a sales rank, at an average price of $19.29, and counted listing keywords rather than search volume. Free accounts see an 8-product preview, while searching, full results and niche analysis need a paid plan, and Editable text and Upscale need a paid plan or purchased credits. The guide's own run generated two designs for 40 credits, changed "Weekend" to "Sunday" with Precise Edit for 6 credits, and exported a 3,333 × 3,333 pixel print file at 300 PPI.

### NVIDIA — Introducing NVIDIA Nemotron for Science, just announced at #SC26 by Ian Buck, our VP of Hyperscale and High-Performance Computing.

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-16T23:59:54+00:00
- Category: developer-tool-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://x.com/NVIDIAAI/status/2122479807858680178
- Record: https://forck.live/items/36990-introducing-nvidia-nemotron-for-science-just-announced-at-sc26-by-ian-buck-our

NVIDIA announced Nemotron for Science at SC26, a collection of open models, datasets, tools and post-training recipes that it says will help researchers build more accurate and efficient AI agents for science. Ian Buck, NVIDIA's vice president of hyperscale and high-performance computing, presented the collection at the conference. According to NVIDIA, the agents built with it can plan tasks, run simulations, check results and adjust as they go. A new walkthrough shows how to post-train Nemotron 3.5 Super VL for science using NeMo Gym and NeMo RL.

### NVIDIA — NVIDIA DGX Spark Developer Kit Turns a Desk Into a Supercomputing Lab

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-16T23:00:05+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://blogs.nvidia.com/blog/dgx-spark-hpc-developer-kit-sc26/
- Record: https://forck.live/items/36966-nvidia-dgx-spark-developer-kit-turns-a-desk-into-a-supercomputing-lab
- Subject: AI platform

NVIDIA is introducing the DGX Spark HPC Developer Kit, a four-node cluster of DGX Spark systems for researchers, academics and developers, at the SC26 supercomputing conference in Chicago. The kit links each node directly to its two neighbors in a ring over NVIDIA ConnectX-7 networking, so there is no network switch to buy or set up. At four nodes it offers 512GB of combined memory and 20TB of NVMe storage, and with NVFP4 researchers can deploy models with up to 700B parameters locally, such as DeepSeek V4.1. NVIDIA says the desk-side cluster lets teams with fixed compute allocations develop and debug without queueing, saving server hours for production runs and keeping unpublished results and proprietary datasets on premises. Demonstrated four-system workloads include LAMMPS with Kokkos for materials simulation and NAMD for molecular dynamics, and an NVIDIA HPC SDK installation playbook is coming soon. The kit is on display at SC26 and available for sale through OEM partners including Acer, ASUS, Dell Technologies, GIGABYTE, H3C, HP, Lenovo, MSI and xFusion, and through NVIDIA Marketplace.

### Amazon — Build memory-aware AI agents with Agentic Retrieve on Amazon Bedrock Managed Knowledge Base

- Company: Amazon (amazon.com)
- Announced: 2026-11-16T22:43:30+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/build-memory-aware-ai-agents-with-agentic-retrieve-on-amazon-bedrock-managed-knowledge-base/
- Record: https://forck.live/items/36947-build-memory-aware-ai-agents-with-agentic-retrieve-on-amazon-bedrock-managed
- Subject: Bedrock / Nova

Amazon Bedrock Managed Knowledge Base and AgentCore memory now handle conversation storage, fact extraction and retrieval for knowledge-grounded agents, so applications no longer persist history themselves. Developers turn it on by adding a memoryConfiguration to the AgenticRetrieveStream request, which makes AgentCore memory a retrieval source alongside the knowledge base. Within the same call the service restores the conversation, and the planner decides on its own when a fact about the user would help, retrieving it from long-term memory to compose the sub-queries it sends to the knowledge base. The same request carries a userContext, so the knowledge base returns only documents the user's access control lists allow, and the integration is opt-in and backward compatible. The post walks through an internal employee assistant for a fictional company, where staff sign in with Amazon Cognito and the ID token's sub claim identifies the user's memory while the verified email identifies the user to the ACL filter. It covers designing memory namespaces for many users, reading the trace to see when the planner used memory, and using AWS Identity and Access Management to enforce per-user memory isolation even when application code gets it wrong.

### Amazon — How Duco transforms financial data automation with Amazon Bedrock AgentCore

- Company: Amazon (amazon.com)
- Announced: 2026-11-16T22:06:08+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/how-duco-transforms-financial-data-automation-with-amazon-bedrock-agentcore/
- Record: https://forck.live/items/36945-how-duco-transforms-financial-data-automation-with-amazon-bedrock-agentcore
- Subject: Bedrock / Nova

Duco rebuilt its agentic AI stack on Amazon Bedrock and Amazon Bedrock AgentCore, cutting deployment time and onboarding costs while keeping each bank's data isolated. The financial operations platform had relied on traditional machine learning models for predictive labeling and pattern recognition, which required extensive historical data and heavy customization for each new use case. Rob Gates, Duco's Chief Architect and Innovation Officer, said the classical models "weren't agile or flexible enough" and could only amplify what they had already learned. Duco chose AgentCore as its runtime because it is fully compatible with its existing LangChain and LangGraph code, so most code continued to work without modification, and because it provides a managed runtime that reduces operational overhead. The platform runs as a multi-tenant SaaS offering with strict tenant isolation, a non-negotiable requirement for regulated institutions, enforced from the Kubernetes layer up to the AI agent runtime. Duco processes 20 billion transactions monthly for more than 200 clients, including 7 of the top 20 banks and 10 of the top 20 asset managers.

### Amazon — Taxonomy of multi-agent design patterns, anti-patterns, and lessons from manufacturing

- Company: Amazon (amazon.com)
- Announced: 2026-11-16T21:53:46+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/taxonomy-of-multi-agent-design-patterns-anti-patterns-and-lessons-from-manufacturing/
- Record: https://forck.live/items/36922-taxonomy-of-multi-agent-design-patterns-anti-patterns-and-lessons-from
- Subject: Bedrock / Nova

Amazon's guide lays out a taxonomy of multi-agent design patterns and anti-patterns, arguing that failed enterprise implementations usually stem from coordination rather than the model. It splits coordination into two tiers: deterministic control flow, where developers define the path, and autonomous patterns, where the model chooses the path at runtime at higher cost in tokens, latency and debugging. The post cites roughly 88% of AI pilots never reaching production and a Stanford HAI result that two models can perform worse collaborating than either alone when coordination overhead exceeds the benefit. It illustrates compounding risk with a sequential chain: ten steps each 95 percent reliable yield 59.9 percent end-to-end reliability, and 20 steps yield 35.8 percent. Tier 1 covers sequential pipelines, conditional workflows, conditional routers and parallel workflows, while Tier 2 covers plan-based orchestration and handoff. The guidance is to start with Tier 1 and move a task to Tier 2 only when its subtasks emerge during the work, and it says manufacturing concerns such as takt time, safety boundaries, edge data and shift cycles shape the architecture.

### GitHub — GitHub Copilot weekly releases — November 9

- Company: GitHub (github.com)
- Announced: 2026-11-16T21:34:53+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://github.blog/changelog/2026-11-16-github-copilot-weekly-releases-november-9
- Record: https://forck.live/items/36932-github-copilot-weekly-releases-november-9
- Subject: Copilot

GitHub added Claude Fable 5.2 and Gemini 4 Carbon to Copilot for Pro+, Max, Business and Enterprise users, and made GitHub Enterprise Server repositories generally available in the Copilot app. Claude Fable 5.2 requires data retention by default, and Business and Enterprise administrators must switch on its policy, which is off by default. Gemini 4 Carbon carries no such policy note in the changelog. GHES support in the Copilot app still needs a GitHub.com or GHE.com account, and cloud sessions are not available for GHES. Other changes include an Open in menu that shows a QR code for continuing cloud and Codespace sessions on a phone, and in-conversation sign-in for MCP servers such as Slack. Project automations can now be scheduled down to the minute or with a custom cron schedule, and dictation covers French, Spanish and other languages through a downloadable multilingual voice model. GitHub also said the Local harness will be removed in February 2027, with the Copilot harness already the default for new sessions.

### GitHub — GitHub Copilot models deprecated on November 16, 2026

- Company: GitHub (github.com)
- Announced: 2026-11-16T20:25:20+00:00
- Category: deprecation
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://github.blog/changelog/2026-11-16-github-copilot-models-deprecated-on-november-16-2026
- Record: https://forck.live/items/36896-github-copilot-models-deprecated-on-november-16-2026
- Subject: Copilot

GitHub has deprecated several models across most Copilot experiences, including Copilot Chat, inline edits, ask and agent modes, and code completions, as of November 16, 2026. The deprecations were announced on October 15, 2026, and GitHub says no action is required to remove the deprecated models. Claude Opus 4.8 remains available to individual Copilot subscribers on annual plans, while Claude Opus 5 has been removed from the model picker but stays available for runtime refusal fallback through existing authorized access. Under default model enablement, suggested alternatives are automatically enabled for Copilot Enterprise and Copilot Business customers unless an administrator has turned off the global default or explicitly disabled the model. Administrators who turned off the global default can enable the alternatives through their model policies in Copilot settings, after which users see the model in the Copilot Chat model selector in supported experiences. GitHub encourages enterprise customers with questions to contact their account manager.

### NVIDIA — How the World’s Most Advanced AI Infrastructure Supply Chain Invents the Future

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-16T20:00:37+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://blogs.nvidia.com/blog/ai-infrastructure-supply-chain/
- Record: https://forck.live/items/36879-how-the-world-s-most-advanced-ai-infrastructure-supply-chain-invents-the-future
- Subject: AI platform

NVIDIA operations chief Deb Shoquist described how the company picks and builds manufacturing partners, in an interview given after she accepted the 2026 Dr. Lisa Su Woman of Innovation Award. She calls the job a matchmaker's: a partner must match NVIDIA's ramp style, be financially strong enough to invest in its ramps, and be willing to do something never done before. When only two of those three hold, she says NVIDIA has invested in the company to make it a partner, as it did with optical suppliers Coherent and Lumentum. She cites Amphenol's cable cartridge, with 1,500 small cable legs each assembled to under 1 DPPM, built on more than a hundred automated lines in Vietnam. Shoquist meets key suppliers quarterly and spends six weeks a year visiting partners in Taiwan, China, Korea, Vietnam and the U.S.

### GitHub — Control where GitHub Copilot customizations come from

- Company: GitHub (github.com)
- Announced: 2026-11-16T19:48:33+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://github.blog/changelog/2026-11-16-control-where-github-copilot-customizations-come-from
- Record: https://forck.live/items/37014-control-where-github-copilot-customizations-come-from
- Subject: Copilot

GitHub is letting enterprise administrators require that selected Copilot customizations come from plugins or managed settings instead of standalone user or project configuration. The strictPluginOnlyCustomization managed setting blocks standalone customizations from user and project sources, while eligible plugins and managed settings stay allowed subject to other policies. Administrators can apply the restriction to all five supported categories or pick individual ones: skills, custom agents, hooks, MCP servers, and executable Copilot extensions. GitHub notes that plugin-only does not mean enterprise-managed plugins only, and it is not a marketplace allowlist; it can be paired with strictKnownMarketplaces to restrict eligible marketplace sources. The controls are configurable on GitHub Enterprise Cloud for members with Copilot Business or Copilot Enterprise licenses issued by the enterprise or one of its organizations.

### GitHub — VS Code retires the Local harness in February 2027

- Company: GitHub (github.com)
- Announced: 2026-11-16T19:22:56+00:00
- Category: deprecation
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://github.blog/changelog/2026-11-16-vs-code-retires-the-local-harness-in-february-2027
- Record: https://forck.live/items/37015-vs-code-retires-the-local-harness-in-february-2027
- Subject: Copilot

Visual Studio Code 1.146 deprecates the Local harness, the older software that coordinates the model, context, and tools for GitHub Copilot agent sessions in the editor, and will remove it in February 2027. Its replacement, the Microsoft Copilot harness, is built into VS Code and has been the default for new sessions since VS Code 1.143. Choosing the Copilot harness does not move files to the cloud, and users can keep working in VS Code and running agent commands on their own machine. During the three-month transition the Local harness remains available where an organization's policy permits, but new features and improvements will focus on the Copilot harness. Built on the Copilot SDK, the Copilot harness shares its coding-agent runtime with Copilot CLI and the GitHub Copilot app, and a session can continue between the editor and the Agents window while the session's Agent Host process stays running. Administrators can enforce managed runtime permissions and sandbox restrictions that the Local harness cannot enforce, though selecting the Copilot harness does not enable sandboxing. Users who still rely on the Local harness should start a new session, select Copilot under Session Target, and review customizations for compatibility, since switching harnesses does not automatically convert every customization.

### AMD — AMD-Powered Lux Supercomputer Opens to Scientific Users at Oak Ridge, Becoming First AI Factory for DOE’s Genesis Mission

- Company: AMD (amd.com)
- Announced: 2026-11-16T19:00:00+00:00
- Category: infrastructure-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://ir.amd.com/news-events/press-releases/detail/1307/amd-powered-lux-supercomputer-opens-to-scientific-users-at-oak-ridge-becoming-first-ai-factory-for-does-genesis-mission
- Record: https://forck.live/items/36855-amd-powered-lux-supercomputer-opens-to-scientific-users-at-oak-ridge-becoming
- Subject: AMD Instinct / ROCm

AMD says the Lux SI + HPC supercomputer is now open to scientific users at the U.S. Department of Energy's Oak Ridge National Laboratory, giving the Genesis Mission its first dedicated U.S. AI Factory. Lux was developed through a public-private partnership between AMD and the DOE and ranks 14th on today's TOP500 list of the world's most powerful supercomputers. The system combines AMD EPYC processors, AMD Instinct MI355X GPUs, AMD Pensando networking and AMD ROCm software in a system built by HPE, pairing AI with high-performance modeling and simulation. Genesis Mission projects allocated to run on Lux cover AI-assisted fusion power plant design, an AI-based foundation model factory for materials development and advanced manufacturing, and Earth system prediction for water availability, streamflow and river temperature in the Tennessee Valley Authority region. Lux builds on the computing foundation ORNL established with the Frontier supercomputer, the first to break the exascale barrier, and the upcoming Discovery supercomputer will extend the work with deeper integration of AI and high-performance computing. AMD is also working with ORNL on DOE's American Science Cloud, a national platform meant to extend advanced AI and high-performance computing across DOE laboratories, universities and mission partners. Researchers interested in using Lux are encouraged to contact ORNL's Oak Ridge Leadership Computing Facility to discuss computing needs and applicable access pathways.

### Amazon — How MoEngage built agentic customer engagement on Amazon Bedrock AgentCore

- Company: Amazon (amazon.com)
- Announced: 2026-11-16T18:16:07+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/how-moengage-built-agentic-customer-engagement-on-amazon-bedrock-agentcore/
- Record: https://forck.live/items/36811-how-moengage-built-agentic-customer-engagement-on-amazon-bedrock-agentcore
- Subject: Bedrock / Nova

MoEngage built Merlin AI, a set of agents that work on live platform data, on Amazon Bedrock and Amazon Bedrock AgentCore, and exposed its existing APIs as MCP tools so assistants such as Claude or ChatGPT can reach a brand's data. The company says more than 1,350 consumer brands use its platform across push, email, in-app, SMS, WhatsApp and RCS. The MoEngage MCP Connector gives an MCP-compatible assistant secure, tenant-scoped access to a brand's segments, campaign history and content library, which MoEngage treated as an architecture problem: how to grant that access without rebuilding hundreds of existing APIs. Merlin AI runs four agent modes, embedded agents in Merlin AI Studio, conversational agents, custom agents and decisioning agents, all sharing foundation models in Amazon Bedrock, AgentCore runtime for hosting and MCP for data access. Across all four modes it uses Anthropic Claude Sonnet 4.6, Claude Sonnet 5 and Claude Sonnet 5.5 on Amazon Bedrock for reasoning, tool-call planning and content generation. MoEngage says customer Glance saw a 38% uplift in push notification click-through rates as a result.

### Amazon — Modernizing mainframe Assembler to Java using Amazon Bedrock and AWS Transform

- Company: Amazon (amazon.com)
- Announced: 2026-11-16T17:29:29+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/modernizing-mainframe-assembler-to-java-using-amazon-bedrock-and-aws-transform/
- Record: https://forck.live/items/36790-modernizing-mainframe-assembler-to-java-using-amazon-bedrock-and-aws-transform
- Subject: Bedrock / Nova

Amazon's guide shows banking, insurance and airline teams how to move mainframe Assembler programs to Java Spring Boot on AWS using Amazon Bedrock, Kiro IDE and AWS Transform. The problem it addresses is a staffing one: Assembler still runs critical operations, but fewer people can read it each year, and hand rewrites take years and depend on the same experts who are retiring. The method runs in five phases, starting with Business Rules Extraction documentation generated from the Assembler source, followed by functional specifications refined with human-in-the-loop review. The validated specifications drive an Assembler-to-COBOL conversion, and the intermediate COBOL is tested for functional equivalence before AWS Transform converts it to Java. Amazon says the COBOL stage gives subject matter experts a structured, human-readable format to validate before the final transformation, which it says reduces risk and preserves business logic more reliably than direct Assembler-to-Java conversion. Amazon Bedrock is the inference platform behind the first four phases, and the post supplies prompts that can be run in the Bedrock Chat Playground or in Kiro IDE's AI chat and Spec panel. Prerequisites include an AWS account with least-privilege IAM permissions, access to AWS Transform for COBOL to Java conversion, an encrypted S3 bucket for source and artifacts, and test data with expected outputs.

### Amazon — Novo Nordisk modernizes manufacturing process documents with Amazon Bedrock AgentCore

- Company: Amazon (amazon.com)
- Announced: 2026-11-16T17:04:11+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/novo-nordisk-modernizes-manufacturing-process-documents-with-amazon-bedrock-agentcore/
- Record: https://forck.live/items/36769-novo-nordisk-modernizes-manufacturing-process-documents-with-amazon-bedrock
- Subject: Bedrock / Nova

Novo Nordisk's Manufacturing Technology team within the Tablet area has built an agentic AI support system on AWS to make manufacturing process documents and knowledge easier to find. The team's document portfolio spans standard operating procedures, work instructions, forms and training materials, and updates driven by regulatory changes, process improvements, equipment modifications and audit findings are tracked today with spreadsheets and disconnected systems. The assistant uses a GraphRAG architecture, with a knowledge graph in Amazon Neptune and a hybrid search index in Amazon OpenSearch Serverless kept separate and bridged at query time. Reasoning runs on Amazon Bedrock AgentCore, chosen for its managed agent runtime with session isolation, tool orchestration and integrated IAM and observability, with agents built on the Strands Agents SDK and powered by Anthropic Claude models on Amazon Bedrock. Documents are ingested through Amazon Bedrock Data Automation for multimodal parsing, orchestrated with AWS Step Functions and AWS Lambda, and the stack is defined with the AWS CDK. A guided mode asks clarifying questions when a query's scope is unclear, and the team says the solution helped subject-matter experts complete work that used to take hours or days in minutes.

### Microsoft Research — TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient

- Company: Microsoft Research (microsoft.com)
- Announced: 2026-11-16T17:01:35+00:00
- Category: developer-tool-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.microsoft.com/en-us/research/blog/tokeneconomizer-learning-from-experience-to-make-ai-coding-agents-more-efficient/
- Record: https://forck.live/items/36765-tokeneconomizer-learning-from-experience-to-make-ai-coding-agents-more-efficient
- Subject: Research / Phi

UC Berkeley, Microsoft Research and Microsoft Security have released TokenEconomizer, an open-source plugin that learns from an agent's past sessions in a repository to cut token waste in later ones. Developers return to the same codebases day after day, yet agents often start each session from scratch, rediscovering familiar facts, loading unused tools and repeating unsuccessful approaches. Across 277 real bug-fixing tasks from six repositories and five agent and model combinations, the plugin lowered average cost by 18.4–30.9% and saved the most in all 30 repository-combination pairs without performance loss. It works through three mechanisms: micro-skills that capture repository-specific knowledge in one-line notes, pruning of tools history suggests the agent will not use, and intervention when an agent keeps hitting the same failure. Savings grow as experience accumulates, with cumulative costs falling below the default in every repository by the sixth task, and they carried over to sequential web development, where costs were 16.0–21.2% lower on Web-Bench. The plugin runs alongside existing agents without modifying their underlying models, requires no API key of its own, and keeps what it learns on the developer's machine.

### OpenAI — Strengthening the fight against online child sexual abuse with new tools and support

- Company: OpenAI (openai.com)
- Announced: 2026-11-16T17:00:00+00:00
- Category: safety-policy-update
- Coverage: 9 outlets
- Announcement: yes
- Group: covered
- Source: https://openai.com/index/strengthening-online-child-safety
- Record: https://forck.live/items/36784-strengthening-the-fight-against-online-child-sexual-abuse-with-new-tools-and
- Subject: GPT / ChatGPT / API

OpenAI released an image classifier that detects potential child sexual abuse material, including AI-generated images, and is offering it free through its API to vetted organizations. The company says it is the first broadly available CSAM classifier with a specific signal for whether material may be AI generated, and it was trained and evaluated with help from the National Center for Missing and Exploited Children. In an evaluation of user-uploaded images, OpenAI estimated the classifier would flag 48% fewer images while finding roughly the same number of abusive images as its existing tool, and flagged images were about twice as likely to contain abuse. Early partners reported their own results: fal reviewers confirmed 72% of high-confidence flags as novel CSAM, invideo AI saw one in four flagged cases lead to a ban, and SafetyKit detected 90% of cases involving drawings or animations that sexualize children. OpenAI is working with the Tech Coalition and ROOST to share access, and recently joined the Tech Coalition's Lantern program to exchange signals on child sexual exploitation and abuse across platforms.

### Amazon — Get answers from tables with tabular search in Amazon Bedrock Managed Knowledge Base

- Company: Amazon (amazon.com)
- Announced: 2026-11-16T16:12:16+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://aws.amazon.com/blogs/machine-learning/get-answers-from-tables-with-tabular-search-in-amazon-bedrock-managed-knowledge-base/
- Record: https://forck.live/items/36737-get-answers-from-tables-with-tabular-search-in-amazon-bedrock-managed
- Subject: Bedrock / Nova

Amazon Bedrock Managed Knowledge Base now supports tabular search, which lets agentic retrieval answer questions that need calculations or lookups across CSV, XLSX and XLS files. The feature targets questions no single file can answer, such as which departments must write variance commentary, where the result depends on a policy threshold, a memo and math over every row of a P&L workbook. Tabular search runs model-written code over the original file in an isolated sandbox with no network access, so the answer uses the whole table rather than only the rows retrieval returns. It works through the same AgenticRetrieveStream request used for documents, and the answer cites the passages it used while the trace names the table files the code read. Amazon says there is nothing to turn on, and the feature works with any connector that brings files into the knowledge base, including Amazon S3, Microsoft SharePoint and Atlassian Confluence. Running code adds at least one more agent turn, so answers that use table data take longer than text-only answers, and Amazon says governed metrics, recurring dashboards and warehouse-scale joins still belong in business intelligence tools.

### Baseten — Introducing the Baseten Voice Agent Stack: own and optimize voice agents end-to-end

- Company: Baseten (baseten.co)
- Announced: 2026-11-16T16:09:26+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.baseten.co/blog/introducing-the-baseten-voice-agent-stack
- Record: https://forck.live/items/36800-introducing-the-baseten-voice-agent-stack-own-and-optimize-voice-agents-end-to
- Subject: Baseten

Baseten is launching the Baseten Voice Agent Stack and Baseten Realtime Voice, both in private preview, to let teams run, customize and scale production voice agents on its own infrastructure. The stack runs the orchestrator and every model over a tight in-cluster network on Baseten Dedicated Inference, scales each component independently, and lets teams pick or post-train any model in the pipeline. Baseten Realtime Voice is a reference agent built on the stack that runs a cascade of Baseten-optimized open-weight speech models with a custom Pipecat-based orchestrator, and it speaks the OpenAI Realtime protocol so existing agents can move over with minimal changes. On Baseten's setup it scores 93% on Full-Duplex-Bench, which Baseten says is the highest of any open-source system on Artificial Analysis, and 73.5% on τ²-Bench Voice, more than double the cascaded baseline. Baseten says it costs less per minute ($0.012) than every hosted speech-to-speech API it evaluated at 32+ concurrent streams, and that the reference agent achieves sub-second median turn latency across 32 concurrent sessions. The stack supports optional delegation, pairing a co-located talker model with a stronger thinker model for reasoning and tool calls, and server-side tool and MCP execution on the orchestrator to cut tool-calling latency.

### NVIDIA — Lambda Boosts AI Factory Resiliency and Efficiency With NVIDIA Photonics

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-16T16:00:51+00:00
- Category: infrastructure-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://blogs.nvidia.com/blog/lambda-photonics/
- Record: https://forck.live/items/36721-lambda-boosts-ai-factory-resiliency-and-efficiency-with-nvidia-photonics
- Subject: AI platform

Lambda is connecting racks of NVIDIA GB300 NVL72 systems into a network of more than 10,000 GPUs using NVIDIA Photonics co-packaged optics switches, aiming to deliver lower power, higher resiliency and longer mean time between interruptions at AI factory scale. The change matters because in large training jobs one interrupted link can stall thousands of GPUs, and every watt spent on networking is a watt unavailable for compute. In Lambda's tests, network latency fell by about 4% at rack scale and by up to about 10% at pod scale, while throughput matched conventional pluggable optics across rack, scalable unit and pod configurations. NVIDIA says replacing pluggable transceivers with co-packaged optical engines yields up to 5x better power efficiency and up to 15x higher reliability, and uses 4x fewer lasers. Each switch carries 144 ports running at 800Gb/s, with liquid cooling and 18 removable external light sources that technicians can replace in the field. Lambda reports a median optics temperature of 36°C against 51°C for conventional pluggable optics, and says co-packaging shortens the electrical path from inches to millimeters and removes retimer chips.

### NVIDIA — NVIDIA Expands Slurm to Bring Quantum Computing Into Supercomputing

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-16T16:00:47+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://blogs.nvidia.com/blog/slurm-quantum-computing-supercomputing/
- Record: https://forck.live/items/36722-nvidia-expands-slurm-to-bring-quantum-computing-into-supercomputing
- Subject: AI platform

NVIDIA is expanding Slurm so that quantum processor builders can expose their hardware to every supercomputing center through a single interface file, building on earlier work with IBM, Pasqal and the wider quantum community. The change answers a practical bottleneck: until now, integrating a QPU into a supercomputing center meant a new software challenge for every new QPU or center. By building on existing quantum-HPC interfaces such as QRMI and QDMI, QPU builders can reuse their integrations across all their deployments, and centers can add quantum processors to existing scheduling workflows without bespoke QPU-specific solutions. Alice & Bob, IBM, IQM, Pasqal, Quantinuum and MQSC are already adopting the expansion, and Slurm 26.11 will be available later this month. NVIDIA says the work adds a community-developed orchestration layer to its open platform for quantum-GPU supercomputing, which centers on CUDA-Q; at SC26 it also detailed CUDA-Q QEC, a fault-tolerant primitives layer, CUDA-Q Algorithms and CUDA-Q Pulse. NVIDIA says a new TCBP-JOBS decoder is 20,000x faster than the industry standard, and that Fermilab used CUDA-Q to perform studies in three weeks that would normally have taken five months.

### NVIDIA — NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-16T16:00:25+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://blogs.nvidia.com/blog/mlperf-training-endpoints-vera-rubin/
- Record: https://forck.live/items/36723-nvidia-sweeps-mlperf-training-and-endpoints-as-vera-rubin-debuts-with-leading
- Subject: AI platform

NVIDIA says its Vera Rubin platform led every MLPerf Training v6.1 benchmark in its debut and, in its first MLPerf Endpoints submission, delivered more than 10x generational gains in throughput per watt for agentic AI. The company was the only one with results on every training benchmark this round, spanning large language models, image generation, recommender systems and reinforcement learning post-training. Vera Rubin NVL72 delivered up to 2x more performance per GPU than the GB300 NVL72 system, which NVIDIA calls early results. On the DeepSeek-V3 benchmark, Vera Rubin NVL72 scaled from 256 to 1,024 GPUs with about 97% efficiency, cutting training time from 16 minutes to about four. NVIDIA also ran the benchmark on 10,240 GB300 GPUs, the largest scale in this round, and Microsoft Azure matched that scale with 10,240 GB200 GPUs. In a provisional result for Kimi K3, Vera Rubin NVL72 delivered more than 10x the throughput per watt of GB300 NVL72 at the same response speed.

### NVIDIA — NVIDIA CUDA-X Libraries Give Agents the Tools to Accelerate Scientific Discovery

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-16T16:00:16+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://blogs.nvidia.com/blog/cuda-x-libraries-agents-accelerate-scientific-discovery/
- Record: https://forck.live/items/36724-nvidia-cuda-x-libraries-give-agents-the-tools-to-accelerate-scientific-discovery
- Subject: AI platform

NVIDIA expanded its CUDA-X libraries with cuPhoton, cuPotential and cuSpa, giving AI agents tools that speed up scientific research across astrophysics, materials science and biology. The company says the new libraries let researchers analyze larger datasets and model more complex systems than before. With cuPhoton.xBeam, black hole observations from the Event Horizon Telescope can be correlated orders of magnitude faster, cutting analysis that once took over a month to eight minutes. Pacific Fusion researchers used the cuPhoton.xLens module to extract insights while plasma simulations run, and can couple the pipeline with a Nemotron-powered agent running locally for interactive analysis. NVIDIA says cuPotential, built on work by the Simons Foundation's Flatiron Institute and D. E. Shaw Research, accelerates existing molecular simulation software by up to 3x while improving stability and accuracy. The newly open source cuSpa accelerates cell polygon calculations, enabling spatial graphs at Memorial Sloan Kettering Cancer Center with a 900x speedup over CPUs.

### NVIDIA — CUDA at 20: How the Platform Transformed Scientific Computing

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-16T16:00:11+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://blogs.nvidia.com/blog/20-years-cuda/
- Record: https://forck.live/items/36725-cuda-at-20-how-the-platform-transformed-scientific-computing
- Subject: AI platform

NVIDIA is marking 20 years of CUDA at SC26 by honoring the researchers and developers who turned GPU computing into a foundation for science and AI. NVIDIA announced CUDA in November 2006 and brought it to the scientific community at SC06, giving scientists a familiar, high-level programming model for using GPUs. The company says the platform extended accelerated computing beyond graphics to demanding applications in science and simulation. More than 11 million developers now build on CUDA across the computing stack, according to NVIDIA. The retrospective credits CUDA-accelerated work behind the HIV capsid's atomic structure, AlexNet's 2012 ImageNet win, and AlphaFold 2, whose publicly released software uses CUDA for GPU execution. NVIDIA points to 1,000+ CUDA-X libraries and directs developers to the CUDA Toolkit and its SC26 booth.

### NVIDIA — NVIDIA Announces Rubin HPC — A New GPU to Accelerate Agents for Science, Engineering and Analytics

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-16T16:00:01+00:00
- Category: product-launch
- Coverage: 2 outlets
- Announcement: yes
- Group: covered
- Source: https://blogs.nvidia.com/blog/rubin-hpc/
- Record: https://forck.live/items/36726-nvidia-announces-rubin-hpc-a-new-gpu-to-accelerate-agents-for-science
- Subject: AI platform

NVIDIA announced Rubin HPC, a new GPU in the Vera Rubin platform, at SC26, built to accelerate simulation and data processing for science, engineering and analytics workflows. The company says agents helping scientists and engineers explore more candidate materials, molecules and designs depend on simulations to test those ideas, and that faster simulation shortens the path from exploration to discovery. Rubin HPC carries a record number of CUDA cores, more than 50,000 on a single GPU, alongside a sparse data acceleration capability that keeps those cores fed with data. It also accelerates GPU-to-GPU connectivity with seventh-generation NVIDIA NVLink interconnect technology. NVIDIA says the GPU's balanced design scales data movement and compute together rather than increasing compute alone, which it says optimizes the chip for recurring simulation patterns such as sparse matrix operations, stencil computations, sorting and reductions. Ian Buck, vice president of hyperscale and high-performance computing at NVIDIA, said faster simulation means faster discovery for agents tackling science and engineering. Rubin HPC is expected to be available in 2028.

### Amazon — Best practices for writing Denied Topics in Amazon Bedrock Guardrails

- Company: Amazon (amazon.com)
- Announced: 2026-11-16T15:48:01+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/best-practices-for-writing-denied-topics-in-amazon-bedrock-guardrails/
- Record: https://forck.live/items/36733-best-practices-for-writing-denied-topics-in-amazon-bedrock-guardrails
- Subject: Bedrock / Nova

Amazon's guide sets out eight best practices for writing Denied Topics in Amazon Bedrock Guardrails, which block a content category by plain-language description applied by meaning rather than keyword. The post says the same mechanism can match more than intended: a topic named Finance, meant to keep an assistant out of investment advice, also blocks "When does my direct deposit clear?". It attributes over- and under-blocking to four authoring patterns: topics framed around a subject domain, definitions built from keyword lists, examples that teach the wrong thing, and overlapping topics that interfere with each other. The practices are worked through on a fictional AnyCompany Bank assistant that answers account, transfer, card and mortgage questions and must refuse personalized investment recommendations, litigation guidance, insider-trading advice, weapon-building instructions, prompt-injection attempts and named competitor products. A topic has a name, a definition and up to five examples, and the guardrail evaluates all three together as a single unit. The name can be up to 100 characters, the definition up to 1,000 characters with 100–300 as the practical sweet spot, each example up to 100 characters, and a guardrail can hold up to 30 topics. The post closes with one limitation to plan for regardless of how well the practices are followed: results are not fully deterministic.

### Amazon — Customize answers per tenant in Amazon Bedrock Managed Knowledge Base

- Company: Amazon (amazon.com)
- Announced: 2026-11-16T15:46:28+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/customize-answers-per-tenant-in-amazon-bedrock-managed-knowledge-base/
- Record: https://forck.live/items/36734-customize-answers-per-tenant-in-amazon-bedrock-managed-knowledge-base
- Subject: Bedrock / Nova

Amazon Bedrock Managed Knowledge Base now supports retrieval instructions, letting teams set a per-tenant voice, format, retrieval guidance and answer policy for agentic retrieval. The same product and policy documents often serve several audiences, such as customers in different regions or internal teams, and each audience needs its own answer. Before this feature, answer style could not be set, so agentic retrieval could use a different style from one request to the next; a fixed voice or structured output meant turning off response generation, calling a model yourself and maintaining the prompt in every application. Instructions are written once as a managed, versioned Amazon Bedrock resource, and each request references it, with versions that can be pinned so applications switch when ready. An instruction belongs to an AWS account and Region rather than a knowledge base, and moves through create, asynchronous review, activation, versioning and application stages, with the first Planning trace event reporting appliedInstruction. Amazon says instructions are not an authorization or data-isolation boundary, so metadata filters or separate knowledge bases are needed to control what each tenant can retrieve.

### Amazon — Use prompt caching with open-weight models on Amazon Bedrock

- Company: Amazon (amazon.com)
- Announced: 2026-11-16T15:42:41+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://aws.amazon.com/blogs/machine-learning/use-prompt-caching-with-open-weight-models-on-amazon-bedrock/
- Record: https://forck.live/items/36735-use-prompt-caching-with-open-weight-models-on-amazon-bedrock
- Subject: Bedrock / Nova

Amazon's guide shows developers how to use prompt caching with open-weight models on Amazon Bedrock, which now support explicit caching alongside implicit caching. Prompt caching reuses the computed state of a matching prompt prefix, so repeated context is read from cache instead of being processed again, which Amazon says helps reduce input costs and response latency. Implicit caching works automatically on a best-effort basis, while explicit caching lets developers mark the end of each reusable prefix with a cache breakpoint, up to the maximum the model supports. Cache reads are billed at a lower rate than standard input, though cache writes can be billed higher depending on the model. The post walks through a support assistant example using the Chat Completions API and applies implicit caching to an OpenCode coding agent session. Amazon says prompts and outputs aren't shared with the model provider or used to train models.

### Perplexity — Today we're launching Perplexity Frontier Research.

- Company: Perplexity (perplexity.ai)
- Announced: 2026-11-16T15:32:19+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://x.com/perplexity_ai/status/2122352071295009156
- Record: https://forck.live/items/36715-today-we-re-launching-perplexity-frontier-research

Perplexity is launching Perplexity Frontier Research, a team whose stated goal is to build a system that can perform autonomous research. The work will focus on multi-agent orchestration, reinforcement learning, continual learning, systems, and pre-training, according to the company. Perplexity says it will develop algorithms, models, and evaluations, released openly and tested in Perplexity products, supported by more than 10,000 GPUs. The Frontier Research team will regularly publish papers, models, data, and benchmarks. Perplexity says it is actively hiring researchers and engineers for the team in San Francisco, Palo Alto, New York, and London.

### Microsoft — Misalignment signals in Kimi K3: What probes flag and where they peak

- Company: Microsoft (microsoft.com)
- Announced: 2026-11-16T15:05:22+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://commandline.microsoft.com/misalignment-signals-in-kimi-k3-what-probes-flag-and-where-they-peak/
- Record: https://forck.live/items/36693-misalignment-signals-in-kimi-k3-what-probes-flag-and-where-they-peak
- Subject: Foundry / Responsible AI

Lightweight activation probes already help screen model activity in production monitoring systems, routing flagged behavior for further review.

### AMD — AMD Completes Acquisition of World Labs

- Company: AMD (amd.com)
- Announced: 2026-11-16T14:05:00+00:00
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://ir.amd.com/news-events/press-releases/detail/1306/amd-completes-acquisition-of-world-labs
- Record: https://forck.live/items/36669-amd-completes-acquisition-of-world-labs
- Subject: AMD Instinct / ROCm

AMD has completed its acquisition of World Labs, adding the AI model research team behind interactive 3D environment models to work on future AMD hardware, software and systems for the open AI ecosystem. World Labs develops AI models that create and simulate interactive 3D environments, as well as technology for robotic learning and simulation. Dr. Fei-Fei Li joins AMD as executive vice president and chief scientist, reporting directly to Su. The World Labs team will continue advancing foundational model research while collaborating with AMD engineers to apply insights from emerging model capabilities to future AI infrastructure. AMD CEO Lisa Su said delivering infrastructure for the next era of super intelligence requires optimizing compute platforms and models together, and that World Labs' model research will help shape AMD's future hardware, software and systems.

### Legora — Walker Morris builds on AI investment with Legora partnership

- Company: Legora (legora.com)
- Announced: 2026-11-16T13:13:00+00:00
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://legora.com/newsroom/walker-morris-builds-on-ai-investment-with-legora-partnership
- Record: https://forck.live/items/36817-walker-morris-builds-on-ai-investment-with-legora-partnership
- Subject: Legora

Walker Morris has partnered with Legora, the agentic operating system for legal work, following a pilot tailored to individual practice areas in which leads built client-focused workflows. The rollout gives all of the firm's 300+ lawyers access to Legora alongside other firmwide AI tools, including Microsoft Copilot and Lexis+ AI Protégé. The firm has introduced a structured, mandatory training and education programme for all fee earners, focused on safe, responsible and effective use and on maintaining the quality of legal advice. Walker Morris says the investment sits within its Ambition 27 strategy and its continued spending on people and technology. Adoption work includes super user groups at all levels that design workflows and standardised prompts for fee earners, and efforts to equip junior lawyers with AI and innovation skills. Alex Fortescue-Webb, Legora's Head of UK & Ireland, says that when firms have a strong group of champions, usage across the rest of the firm more than doubles.

### Legora — Legora extends offering with legal intake in Microsoft Teams and agentic search across Microsoft 365

- Company: Legora (legora.com)
- Announced: 2026-11-16T13:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://legora.com/newsroom/legora-extends-offering-with-legal-intake-in-microsoft-teams-and-agentic-search-across-microsoft-365
- Record: https://forck.live/items/36646-legora-extends-offering-with-legal-intake-in-microsoft-teams-and-agentic
- Subject: Legora

Legora is expanding its Microsoft 365 capabilities with legal intake in Microsoft Teams and agentic search across Microsoft 365 enterprise data, adding to its existing SharePoint and OneDrive integrations and its MCP-powered Copilot plug-in. The company says everyday legal requests arrive through chat, email and meetings while the supporting context sits in Microsoft 365 inboxes, calendars and SharePoint libraries, so lawyers often re-enter requests, forward attachments and export files into separate tools. Legora says those steps slow the work and create document copies outside the original permission controls, and that the new capabilities keep existing user permissions intact. Legal intake in Microsoft Teams and agentic search across Outlook and Calendar are both new and generally available, letting users mention "@Legora" to submit a request and draw on email and calendar context to find information behind a task. The Legora Agent can also draw on stored SharePoint documents to support agreement review and save edits back as a new version, a feature that is in beta. The Legora plugin in Copilot for Legal is generally available, letting users start legal work from Microsoft 365 Copilot. 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, and will exhibit at Microsoft Ignite from November 17-20, 2026 in San Francisco.

### ElevenLabs — Kotoba is joining ElevenLabs

- Company: ElevenLabs (elevenlabs.io)
- Announced: 2026-11-16T12:00:00+00:00
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://elevenlabs.io/blog/kotoba
- Record: https://forck.live/items/36876-kotoba-is-joining-elevenlabs
- Subject: Voice AI

ElevenLabs is acquiring Kotoba, a Japanese voice AI lab, to speed up adoption of AI agents and voice interfaces in Japan. Kotoba was founded in 2023 by AI researchers Noriyuki Kojima and Jungo Kasai, who built voice models for a multilingual world. Its voice model Koto handles Japanese text-to-speech, speech-to-text and speech-to-speech for simultaneous interpretation, and is already used by major global enterprises including Salesforce, where it powers Japanese-language conversations in Agentforce Voice. ElevenLabs CEO Mati Staniszewski said Kotoba has built state-of-the-art voice models for Japan that leading enterprises have adopted, and that the two can build the full interaction layer for AI in Japan. CTO Piotr Dąbkowski said bringing Kotoba's research together with ElevenLabs' will improve Japanese across the entire stack ElevenLabs builds. ElevenLabs plans to build out local sales leaders and keep investing in infrastructure in Japan in the coming months. The acquisition is expected to close later in Q4 2026, subject to customary regulatory approvals and closing conditions.

### ElevenLabs — Autotorino automates dealership calls with ElevenAgents

- Company: ElevenLabs (elevenlabs.io)
- Announced: 2026-11-16T12:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://elevenlabs.io/blog/autotorino-agents
- Record: https://forck.live/items/36749-autotorino-automates-dealership-calls-with-elevenagents
- Subject: Voice AI

Autotorino, an Italian dealership group, built a voice agent on ElevenAgents to answer the roughly 100,000 calls a month its team cannot handle. The agent answers calls, qualifies requests and creates leads, replacing an earlier workflow that combined Zapier with separate transcription and language tools to turn a recorded message into a Salesforce lead. Autotorino compared several providers and chose ElevenLabs on latency, human-like interaction, and how well the system understands callers without moving data between separate tools. The agent now answers at any hour, where calls outside staffed hours often went unanswered. About 70% of requests are post-sales, mainly workshop appointment bookings and vehicle drop-off, and callers also ask about the status of cars in the workshop. Autotorino evaluated Eleven v4 and reported a more natural, less artificial-sounding voice, a more consistent tone, and greater fluency with the same voice, settings and system prompt. Automated call volume is more than 2x the previous system, and the hang-up rate stayed below the limit Autotorino set. In November, Autotorino plans to replace the IVR system at its first branch with ElevenLabs voice agents.

### ElevenLabs — The Year of More: $200,000 for what you create next

- Company: ElevenLabs (elevenlabs.io)
- Announced: 2026-11-16T12:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://elevenlabs.io/blog/the-year-of-more
- Record: https://forck.live/items/36698-the-year-of-more-200-000-for-what-you-create-next
- Subject: Voice AI

ElevenLabs is launching The Year of More, an ElevenCreative competition with $200,000 in prizes, open to free and paid users with no purchase required. The brief asks creators to make something with images or video that shows ElevenCreative is more than voice, whether a short film, an imagined product ad, or a wordless image series. Judges will pick 5 winners at $10,000 each on November 20, 25 and 30, and announce a $50,000 Grand Prize on December 1, 2026. Submissions close November 30 at 11:59pm PT, and an entry only counts once the official form is submitted with post links and a description of how ElevenCreative was used. Entrants must post their creation on at least two of X, LinkedIn, YouTube and Instagram with the competition watermark and #morethanvoice, tagging @elevenlabs and @elevencreative. Alongside the competition, the ElevenCreative Black Friday offer gives eligible subscribers up to 55% off annual plans from November 16–December 1, with the Creator annual plan at $99.

### OpenAI — LVMH’s Media Division announces strategic partnership with OpenAI

- Company: OpenAI (openai.com)
- Announced: 2026-11-16T09:00:00+00:00
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://openai.com/index/lvmh-media-division-partnership
- Record: https://forck.live/items/36561-lvmh-s-media-division-announces-strategic-partnership-with-openai
- Subject: GPT / ChatGPT / API

LVMH Group's Media Division and OpenAI signed a strategic agreement to bring AI tools into the division's newsrooms and make its journalism accessible through ChatGPT. The deal covers deployment of OpenAI tools and ChatGPT Enterprise to teams across the media division, which the companies say will let them build AI applications for their day-to-day work. Teams will be able to explore document research, document comparison and preliminary translation, with the stated goal of freeing more time for investigative reporting, analysis and reporting in the field. The partnership also puts content from the division's publications into ChatGPT, which will provide summaries and excerpts in response to relevant user questions, identify sources and include direct links to the original articles. The rollout will be supported by training tailored to each team's use cases, and the agreement extends across all the division's publications. LVMH's Media Division spans 17 media brands, 1,800 employees including nearly 700 journalists, and reaches more than 27 million people in France each month.

### OpenAI — Our first phase of embedded assessments

- Company: OpenAI (openai.com)
- Announced: 2026-11-16T08:00:00+00:00
- Category: safety-policy-update
- Coverage: 8 outlets
- Announcement: yes
- Group: covered
- Source: https://openai.com/index/our-first-phase-of-embedded-assessments
- Record: https://forck.live/items/36762-our-first-phase-of-embedded-assessments
- Subject: GPT / ChatGPT / API

OpenAI has signed agreements with four independent assessors — Apollo Research, METR, Transluce and Goodfire — for a first phase of embedded assessments, giving outside experts deep access to its internal models and systems. The company says the arrangement is meant to let independent reporting give people a clearer view of how labs evaluate and manage risk as AI capabilities advance. Embedded assessments grant assessors some of the same internal resources OpenAI researchers use, including model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation, and discussions with the responsible researchers and engineers. The initial work focuses on alignment and safeguards for models during training and internal use, complementing OpenAI's safety cases, and builds on a recent pilot assessment of its internal monitoring systems and an investigation of the Hugging Face incident. Apollo Research will assess whether models could covertly work against their developers' intentions and whether safeguards would detect such behavior, testing four claims in two stages. METR is to assess misaligned behavior in what OpenAI calls its largest reinforcement learning run to date, and the two have preliminarily agreed to a broader monitor assessment pending agreement on scope. Transluce will examine the quality and coverage of OpenAI's post-training alignment evaluations, with a follow-on assessment of about eight weeks planned for early 2027. OpenAI says the efforts support commitments under the White House Accord on Super Intelligence, where it agreed to independent assessments of its safety controls.

### OpenAI — More ways to move care forward with ChatGPT for Healthcare

- Company: OpenAI (openai.com)
- Announced: 2026-11-16T06:00:00+00:00
- Category: capability-change
- Coverage: 2 outlets
- Announcement: yes
- Group: covered
- Source: https://openai.com/index/more-ways-to-move-care-forward-with-chatgpt-for-healthcare
- Record: https://forck.live/items/36686-more-ways-to-move-care-forward-with-chatgpt-for-healthcare
- Subject: GPT / ChatGPT / API

OpenAI is adding computer-use agents, improved clinical search and more than 20 new plugins to ChatGPT for Healthcare, aiming to cut the time clinicians spend moving between systems. The agents read screens, click, type and move between applications so they can carry out multi-step work in the software healthcare teams already use, such as processing referrals and moving documents between systems. OpenAI says it is working with Abridge, Ambience, Cigna Healthcare and CVS Health on deployments, and that organizations can set which applications agents may access and which steps need human review. For HIPAA-compliant work with protected health information, organizations must use eligible services covered by their BAA with OpenAI, apply required security and retention settings, and extend protections to connected systems; content shared with ChatGPT for Healthcare is not used to train models. The clinical search updates, which cover how ChatGPT finds, evaluates and cites medical evidence, apply to ChatGPT for Healthcare and ChatGPT for Clinicians. In a blind evaluation, physicians reviewed responses to real-world clinical questions across multiple specialties, including multi-turn conversations, and in over 1300 reviews rated the updated experience more highly across all eight dimensions evaluated, including factual correctness, safety, actionability and source authoritativeness. The new plugins, from companies across clinical care, care coordination and business operations, are available today in ChatGPT for Healthcare, including through ChartGPT, which brings ChatGPT into the EHR workflow.

### Vercel — Agents are users, so your software needs evals

- Company: Vercel (vercel.com)
- Announced: 2026-11-16T05:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://vercel.com/blog/agent-evals
- Record: https://forck.live/items/36828-agents-are-users-so-your-software-needs-evals
- Subject: v0 / AI SDK / AI Gateway

Vercel's guide explains how to build evals for agents that use software, focusing on whether an agent completes a user's task correctly. It argues that a successful request can still be the wrong outcome, since an agent can place a "successful" order for a hot latte when the user asked for an iced one, and logs look fine. Because LLM outputs vary between runs, the guide says one good manual test cannot tell you whether a fix worked, so the same tasks must be repeated and graded before and after a change. It groups agentic tests into a pyramid by how many LLM calls they need: deterministic tests, one-turn evals, and end-to-end evals, and says the cheap, fast tests should make up the bulk of a suite. The worked example is SF Beans, a fictional coffee-ordering API, with an end-to-end eval built using Harbor, an open-source framework from the creators of Terminal-Bench that runs agents in containers, checks results, and records their actions. The guide defines task, variant, trial, run and trace, and says goals should come from observability, customer tickets, or the roadmap rather than aiming for full coverage.

### Alibaba — Qwen4-Max: Your New Teammate Has Arrived

- Company: Alibaba (alibaba.com)
- Announced: 2026-11-16T02:00:00+00:00
- Category: new-model
- Coverage: 9 outlets
- Announcement: yes
- Group: models
- Source: https://qwen.ai/blog?id=qwen4-max
- Record: https://forck.live/items/36470-qwen4-max-your-new-teammate-has-arrived
- Subject: Qwen

Alibaba is introducing Qwen4-Max, a 2.4-trillion-parameter model built on the new Qwen4 architecture with native vision and a one-million-token context window. The company says Qwen4-Max is a significant step up from Qwen3.8-Max, delivering stronger performance across nearly every capability it evaluated. Alibaba reports that Qwen4-Max outperforms Qwen3.8-Max with fewer tokens: on Work360, token usage drops by approximately 24% and 35% at Xhigh and High effort, and on Terminal-Bench 4.0 by approximately 48% and 49%. Alibaba has also lowered API prices compared with Qwen3.8-Max, with particularly large reductions in cached-input pricing; in its Work360 evaluation, the average cost per task drops from $4.43 to $1.67, a 62% reduction. The model is built on a 2.4T backbone with 71B activated per token and an additional 328B N-gram Embedding parameters, using GDN plus Qwen Sparse Attention, Gated Residual branches and N-gram Embedding. Alibaba says pre-training at that scale progressed smoothly, with loss decreasing steadily and no loss spikes, and that Qwen4-Max-Base improves on Qwen3.8-Max-Base in every domain with approximately 5–13% lower BPB while activating 25% fewer parameters. On the Work360 benchmark, Qwen4-Max scored 52.69, up from 47.74 for Qwen3.8-Max.

### NVIDIA — AI Day Seoul Brings Together Korea’s Industry Leaders, Startups and Developer Communities

- Company: NVIDIA (nvidia.com)
- Announced: 2026-11-16T00:00:28+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://blogs.nvidia.com/blog/ai-day-seoul-wrapup/
- Record: https://forck.live/items/36442-ai-day-seoul-brings-together-korea-s-industry-leaders-startups-and-developer
- Subject: AI platform

NVIDIA held AI Day Seoul at COEX, drawing more than 2,000 attendees from Korea's industry, startups and developer communities to work on physical AI and AI agents. Deepu Talla, NVIDIA's vice president of robotics and edge AI, said Korea is "punching above its weight," pointing to Doosan Robotics, Hyundai Motor Group, LG Electronics and SK Group using NVIDIA platforms. Jonathan Cohen, vice president of applied research, showed enterprises how to build agents they own and control, while breakout sessions drew speakers from Coupang Intelligent Cloud, LG AI Research, NAVER Cloud, SK Telecom and Upstage. The Agentic AI 10x Bootcamp Community Day gathered developer groups including Pseudo Lab, Python Korea, PyTorchKR, HAERAE and Hugging Face KREW. Five finalists from the Hangul Day hackathon, co-hosted with Fastcampus, built agents trained on Korean language and culture with the OpenShell runtime, and the winning team, FlyGate, received an NVIDIA DGX Spark system. More than 500 developers joined full-day NVIDIA Deep Learning Institute workshops and Korean-language certification exams covering inference at scale, Isaac robotics workflows and secure AI agents. The Physical AI Startup Connect event featured pitches from 10 NVIDIA Inception startups, and sessions will soon be available on NVIDIA On-Demand.

### Ideogram — Does Ideogram add a watermark? What the image files contain

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-16
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/ideogram-watermark-metadata
- Record: https://forck.live/items/36987-does-ideogram-add-a-watermark-what-the-image-files-contain
- Subject: Ideogram image models

Ideogram added no visible watermark to any image tested on November 16, 2026, but 14 of 16 API files carried signed C2PA content credentials naming Ideogram. The tests found no Ideogram logo, wordmark, text overlay or badge in any file viewed whole and at 100% in all four corners. Ideogram's public documentation does not state a watermark or provenance policy, so most findings are measurements of specific files on one day. Each manifest sat in one PNG chunk named caBX and made the same set of statements, with the signature verified and the content hash matching in every one of those files. The signing certificate chains to a root certificate that Ideogram issued itself, which was not on the published C2PA trust list when checked, so the open-source c2pa library reported each manifest as Valid and its signer as untrusted. Ideogram did not store the prompt or the seed in any file tested, and EXIF was absent from most files; where it appeared it was one 134-byte block with four tags. The Ideogram API needs no subscription, and Ideogram 4.5 costs $0.03 per image at Low quality, $0.06 at Medium and $0.10 at High.

### Ideogram — Ideogram commercial use: who owns the images you make

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-16
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/ideogram-commercial-use
- Record: https://forck.live/items/36986-ideogram-commercial-use-who-owns-the-images-you-make
- Subject: Ideogram image models

Ideogram has published a plain-language summary of its legal documents, checked on November 16, 2026, explaining who owns images made with its tools and what each plan permits. The Terms of Service state that Ideogram does not claim ownership of user output, with Section 2.1 reading: "We do not claim any ownership rights in your User Input or User Output, and we do not restrict your ability to use User Output for your own purposes (including for commercial purposes)." That clause sets two limits: output may not be unique because another user can get a similar image from a similar prompt, and users are responsible for checking their content does not violate laws or third-party rights. Commercial use does not depend on the plan, though the plan changes who can see images, and published images fall under Section 4.2, which lets other users use, reproduce, modify and remix public content. Ideogram takes a royalty-free, transferable, sublicensable, worldwide and irrevocable license to content solely to operate, improve, promote and provide the services, and the terms describe no training opt-out for app plans below Enterprise. The open-weight Ideogram 4.0 model has separate licenses: free weights are non-commercial, while commercial self-hosting needs the Self-Serve Commercial License or an Enterprise license, with $300 a month buying 10,000 images, or $0.03 per image, and annual billing showing $240 a month.

### Ideogram — Are Ideogram images private? Visibility, deletion and model training

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-16
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/ideogram-image-privacy
- Record: https://forck.live/items/36985-are-ideogram-images-private-visibility-deletion-and-model-training
- Subject: Ideogram image models

Ideogram's guide explains who can see images made on each plan, how to publish or unpublish them, and when content is used for model training. Images made on the free plan are public, while Plus, Pro, Team and Enterprise keep a new image private until its creator publishes it, and publishing can be reversed image by image or in bulk from Library. Unpublishing only changes community visibility and does not delete the image, and it requires a plan that includes private generation; some team workspaces turn publishing off entirely. Training terms differ by product: the API agreement says Ideogram does not use User Input or User Output to train the Ideogram AI Model except content flagged as violating its Usage Policies, and the Enterprise page says Enterprise inputs are not used to train Ideogram's shared models. Deleting an image from Library cannot be undone, and deleting an account permanently removes all generations and data, usually in about 48 hours, with Enterprise customers able to configure retention down to 30 minutes. The guide summarizes Ideogram's published pages as they read on November 16, 2026, and notes the Terms of Service, Privacy Policy and API agreement are the binding documents.

### Ideogram — 8 Best Photoroom Alternatives for Product Photos (2026)

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-16
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/best-photoroom-alternatives
- Record: https://forck.live/items/36984-8-best-photoroom-alternatives-for-product-photos-2026
- Subject: Ideogram image models

Ideogram has published a comparison of eight Photoroom alternatives for product photography, testing Ideogram and Pixelcut on the same AURELLE lotion packshot and lifestyle scene. The test asked each tool to change the scene's wall color while keeping the bottle, label and shadow unchanged, and both kept the label readable. Ideogram completed the wall-color revision with a prompt alone and no mask, while Pixelcut placed the bottle further right, leaving more room for ad copy. The guide notes that Pixelcut's prompt-edit route needed a paid upgrade on the account used, and that Photoroom's own Product Staging required Pro. It advises teams to check credits and exports, batch editing, API access and privacy terms before switching, since Photoroom separates AI credits from exports and its app and API have different data policies.

### Ideogram — Ideogram vs Midjourney: how they differ and which fits your work

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-16
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/ideogram-vs-midjourney
- Record: https://forck.live/items/36983-ideogram-vs-midjourney-how-they-differ-and-which-fits-your-work
- Subject: Ideogram image models

Ideogram published a side-by-side comparison of its image tools and Midjourney's, drawing on both companies' documentation and checked on November 16, 2026. The post says the two differ most on text rendering, vector output, API access and style tooling, and it tells readers to test with their own prompts because Ideogram publishes the post. Ideogram's text rendering page claims 95% accuracy, a figure the page does not explain, while Midjourney's guide says shorter words or phrases have a better chance of appearing correctly. Midjourney starts at $10 a month and Ideogram's first paid plan at $20, and on the post's arithmetic Midjourney's Basic plan yields more fast images per dollar. Midjourney's guidelines say it does not provide an API, while Ideogram's public API is prepaid and billed per image, from $0.03 at Low quality to $0.10 at High for Ideogram 4.5. Ideogram also notes Midjourney documents more ways to develop and reuse a style, which the post calls a reason to choose Midjourney if that is the center of your work.

### Ideogram — Bulk image generation: from a spreadsheet of prompts to the API

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-16
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/bulk-image-generation
- Record: https://forck.live/items/36974-bulk-image-generation-from-a-spreadsheet-of-prompts-to-the-api
- Subject: Ideogram image models

Ideogram's guide explains two routes for bulk image generation: uploading a spreadsheet of prompts to Batch, or sending prompts through the Ideogram API. Batch accepts a file with up to 500 prompt rows plus a header, and each row produces 1 to 4 images, with the credit cost shown on the review screen before the batch starts. A blank model cell defaults to Ideogram 3.0 and a blank num_images cell defaults to four images, so the guide tells users to set num_images to 1 for one image per row. The API has no batch endpoint, so code sends one request per generation, up to 10 in flight at a time by default, and teams working at catalog volume add a custom model and a production workflow through Ideogram Enterprise. Batch does not support Ideogram 4.5, partner models or video, and Ideogram 1.0 is retired, so a file that uses it is rejected. The guide says a job that needs more than 500 prompt rows, Ideogram 4.5, Precise Edit, a tool such as Ad Resizer or Background Remover, or a source image belongs in the API instead.

### Ideogram — How to create a consistent character with AI in Ideogram

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-16
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/consistent-character-ai
- Record: https://forck.live/items/36973-how-to-create-a-consistent-character-with-ai-in-ideogram
- Subject: Ideogram image models

Ideogram's guide explains how to keep a character consistent across images, testing one fictional character through six scenes and reporting what held and what drifted. Ideogram 4.5 is the model for character consistency in the app and accepts the same reference images through the API, while Ideogram 3.0 offers a second route through its Character app and Character endpoints. Ideogram 4.5 takes one to five reference images, each JPEG, PNG or WebP up to 25 MB, whereas the Ideogram 3.0 Character app and API take one image; dry runs on November 16, 2026 showed a sixth file and a second character reference each rejected with a 400 error. The guide's own reference stops at the chest, so the scenes came back with different trousers and shoes, illustrating that the reference decides what the model has to invent. A character image from Ideogram 4.5 with references costs 6 credits in the app, or $0.06 through the API, at Medium quality, and the price does not change with the number of reference images. High with references costs $0.22 against $0.10 for a prompt-only image, making Medium the cheaper setting to iterate at, and reference-image generations use credits on every plan.

### Luma — Building canvas-native agents: lessons from Laya

- Company: Luma (lumalabs.ai)
- Announced: 2026-11-16
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://lumalabs.ai/news/building-canvas-native-agents
- Record: https://forck.live/items/36959-building-canvas-native-agents-lessons-from-laya
- Subject: Ray / Dream Machine

Luma described how it built canvas-native agents that read a creative workspace and act inside it, using Laya, its canvas-organization subagent, as a practical test of the approach. The company extended the view of Orca, Luma's main agent, to include canvas objects and their relationships alongside media, so an agent can see frame titles, membership and nearby notes rather than infer them from proximity. Orca delegates organization to Laya, passing the user's intent, permitted scope and constraints, while the server captures the scope and saves a layout recovery snapshot before the job starts. Luma says the split keeps descriptions, grouping decisions, plans and write receipts with the specialist instead of filling Orca's ongoing conversation, and lets it customize Laya's tools, instructions and model configuration independently. In one paired test of the same scoped-layout request, Luma reports runtime fell from 195 to 22 seconds and committed write batches from 12 to one after a fix to parent-local coordinates and measured text bounds. The server applies changes in checked, bounded batches, and if it cannot confirm a batch's outcome it blocks further writes for that run rather than risk applying the changes twice.

### Together AI — Last year we built megakernels. This year we deleted them: The road to Together Compiler

- Company: Together AI (together.ai)
- Announced: 2026-11-16
- Category: developer-tool-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.together.ai/blog/road-to-together-compiler
- Record: https://forck.live/items/36946-last-year-we-built-megakernels-this-year-we-deleted-them-the-road-to-together
- Subject: Inference platform

Together AI has introduced Together Compiler, a framework that uses AI agents to rewrite and fuse the GPU kernels in a model's forward pass, which the company says beats hand-written megakernels by 28% on B200. The company says the compiler reached 2,118 tokens/s against 1,649 tokens/s for megakernels on Llama 3.2 1B. Together had previously helped release some of the first AI megakernels, which executed an entire model forward pass in a single fused GPU kernel, but says they were hard to put into production: after months of effort it had one model, one batch size and one customer. The company found that CUDA Graphs and PDL on top of default Torch Inductor kernels recovered 98% of megakernel performance without writing a single kernel, and that multi-kernels and agent loops beat the old megakernels by up to 1.7x, exceeding 1100 TPS for Llama 1B on H100. On the target Llama 1B workload, Inductor delivers 669 TPS, and Together says its heuristics cannot keep up on different architectures and newer hardware. The post is the first of three parts, covering Torch Compile's heuristics, why megakernels are not always the right optimization target, and the agent-driven production system Together ended up with.

### Ideogram — PDP photography with Ideogram: one product photo to a full image set

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-16
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://ideogram.ai/blog/pdp-photography
- Record: https://forck.live/items/36942-pdp-photography-with-ideogram-one-product-photo-to-a-full-image-set
- Subject: Ideogram image models

A product detail page needs more images than most teams have time to shoot. The workflow below starts with one clear photo and makes the rest of the set from it, in eight steps, with the prompt, the price and the result of each.

### Harvey — Harvey Completes IRAP PROTECTED Assessment to Deepen Support of Australian Public Sector

- Company: Harvey (harvey.ai)
- Announced: 2026-11-16
- Category: safety-policy-update
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.harvey.ai/blog/harvey-completes-irap-protected-assessment-to-deepen-support-of-australian-public-sector
- Record: https://forck.live/items/36934-harvey-completes-irap-protected-assessment-to-deepen-support-of-australian
- Subject: Harvey

Harvey says it has completed an independent security assessment under Australia's Infosec Registered Assessors Program (IRAP) at the PROTECTED level, giving Australian public sector organisations evidence about its security controls. The assessment examined the effectiveness of applicable controls from the Australian Government Information Security Manual at the PROTECTED level, which Harvey says will inform public sector organisations' own risk assessments and system authorisation decisions. Harvey already works with public sector organisations, public universities and government-owned enterprises across Asia-Pacific, supporting legal teams with research, document review and drafting. The company has also been approved as a supplier under the NSW Government ICT Services Scheme (SCM0020) and is listed on the Victorian Government's Digital Marketplace for eServices, giving eligible buyers established procurement routes. Ashleigh Whittaker, Harvey's Director, APAC Strategy, and Country Manager for Australia and New Zealand, said Australia is a priority region and that the company is committed to long-term partnerships with public sector legal teams. The IRAP assessment adds to Harvey's existing security program, which includes FedRAMP Class D certification in partnership with Knox Systems, TX-RAMP Level 2 certification, a SOC 2 Type II attestation, and ISO 27001, ISO 27701, ISO 42001 and AIUC-1 certifications.

### Together AI — Making open models work for long-horizon agentic workloads

- Company: Together AI (together.ai)
- Announced: 2026-11-16
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.together.ai/blog/making-open-models-work-for-long-horizon-agentic-workloads
- Record: https://forck.live/items/36909-making-open-models-work-for-long-horizon-agentic-workloads
- Subject: Inference platform

Together AI and Ollama have written a joint guide on what the harness and the inference engine each must get right so long agent sessions stay fast. Coding agents resend the whole conversation every turn, and prefix caching spares the engine from redoing that work, but only if the harness leaves the start of the prompt alone and the engine sends each turn back to the GPU that already holds it cached. The two companies work on opposite sides of that handoff: Ollama decides what goes into each request, Together AI decides what happens when the request arrives. The post walks through the prefill and decode phases, noting that an agent turn is mostly prefill because the input is the whole session and the output is usually a short tool call. It gives an idealized arithmetic example: ten turns without caching process 55 units of prompt against 10 with caching, and a hundred turns process 5,050 against 100. A prefix cache matches from the beginning of the prompt up to the first differing token and stops there, so changing anything earlier means re-processing everything after it. Ollama says it shipped the ability to connect harnesses such as Claude Code, Codex and Pi to open models like DeepSeek-V4.1-Flash and GLM-5.3 through the ollama launch command.

### Ideogram — How to test AI product images for accuracy before they reach your catalog

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-16
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/ai-product-image-accuracy
- Record: https://forck.live/items/36908-how-to-test-ai-product-images-for-accuracy-before-they-reach-your-catalog
- Subject: Ideogram image models

Ideogram published a six-step method for testing whether AI product images stay accurate before they reach a catalog, using the company's own published examples to illustrate each check. The guide tells teams to build a test set of 20 to 30 already-approved photos, including hard cases such as small label text, shiny packaging, textured materials and stripes that must stay aligned. Every tool under comparison gets the same source file and the same written request, and results are judged against the source at full size rather than as thumbnails. Seven details decide the outcome: shape, color, logos and label text, pattern and material, count, everything not asked to change, and resolution. A pixel-level check follows the visual one: Ideogram's ecommerce page counts a pixel as changed when any color channel moves by more than 12 of 255 after the output is resized to the original size, and in the necklace example 0.13% of pixels changed with Ideogram 4.5. The guide also asks teams to chain ten edits on one image and count how many results reviewers approve, since Ideogram says many image models add artifacts, pixel shifts and color changes with each edit. Ideogram notes the test applies to any tool, including its own, and that a hex code is an instruction rather than a measured color match, so a physical sample should be checked when color must match a manufactured product.

### Ideogram — Ideogram API pricing: what an image, an edit and an upscale cost

- Company: Ideogram (ideogram.ai)
- Announced: 2026-11-16
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://ideogram.ai/blog/ideogram-api-pricing
- Record: https://forck.live/items/36846-ideogram-api-pricing-what-an-image-an-edit-and-an-upscale-cost
- Subject: Ideogram image models

Ideogram published a guide to what each kind of API request costs in US dollars, drawn from its public pricing catalog as of November 16, 2026. Every Ideogram model bills per generated image, and the price depends on the quality or rendering speed you choose: Ideogram 4.5 costs $0.03 per image at Low quality, $0.06 at Medium and $0.10 at High, while Ideogram 4.0 and Ideogram 3.0 both start at $0.03 at Turbo. Upscaling is the exception to per-image billing, charged per output megapixel at $0.03, $0.06 or $0.09 depending on the upscaler, and the legacy v1 upscale endpoint costs $0.06 per image for an upscale of up to 2X. Usage is prepaid: you fund a balance and each request draws it down, with one credit equal to one US cent, and requests are rejected with a 402 error until the account has a payment method and credits. A subscription's monthly credits are used first when they cover the whole request, and the default rate limit is 10 in-flight requests. The Ideogram 4.5 Generate and Precise Edit endpoints accept a dry_run query parameter that validates and prices a request without running it, generating, storing or billing anything.

### Hugging Face — Hugging Face and OpenRouter join forces to make open-weight models easy to run

- Company: Hugging Face (huggingface.co)
- Announced: 2026-11-16
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/openrouter
- Record: https://forck.live/items/36773-hugging-face-and-openrouter-join-forces-to-make-open-weight-models-easy-to-run
- Subject: Platform

Hugging Face has moved its Inference Providers routing to OpenRouter, so every open-weight model on the Hub that a provider serves can be called with a Hugging Face token through one API. Requests to language models now go through OpenRouter, which picks a provider for each call based on live price, speed and reliability, and retries on another provider if the first one fails. The router supports OpenAI's Chat Completions and Responses APIs and Anthropic's Messages API for every chat model, whatever the provider behind it, and the company says your code stays the same. Users can still choose a policy or a provider with a suffix on the model id, such as fastest, cheapest or a named provider, and routing is cache-aware so a conversation stays on the provider holding its prompt cache. Hugging Face says pricing doesn't change: you pay the provider's price on your Hugging Face bill, with no markup added, and PRO users get $2 of credits every month while Team and Enterprise organizations get $2 per seat. As a breaking change, requests to language models are billed to your Hugging Face account or organization, so you can no longer use your own API key for a provider with these models, though custom keys for image and video providers keep working. Image, video and audio generation continue to work as before, directly through Hugging Face Inference Providers without going through OpenRouter.

### Together AI — How we make Together AI easy for coding agents to use

- Company: Together AI (together.ai)
- Announced: 2026-11-16
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.together.ai/blog/how-we-do-agent-evals-at-together-ai
- Record: https://forck.live/items/36758-how-we-make-together-ai-easy-for-coding-agents-to-use
- Subject: Inference platform

Together AI says small fixes to its docs, skills and API errors raised coding-agent task completion on its platform from 75% to 97%. The company built an agent evaluation suite after noticing more developers were pointing coding agents at Together AI, and it had no way to tell whether those agents could work with the platform efficiently. Each task is run three ways: a baseline with only the prompt, a configuration with direct documentation access through Together's MCP server, and one given the Together skill package of agent-written instructions and examples. Every task runs against three models with two attempts each, giving six attempts per task per configuration, and agents run in a Together sandbox with GitHub Actions handling setup and cleanup while a proxy records API calls. An LLM judge scores each run as pass, incomplete or fail, and Together says failed and incomplete verdicts only show where to start looking, not whether the fault was the agent, the API or missing evidence. Together also reviews passing runs, since agents that work around confusing responses reveal what to fix for token efficiency.

### Liquid AI — d1.2: State-of-the-Art Omnimodal Decision Models

- Company: Liquid AI (liquid.ai)
- Announced: 2026-11-16
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://www.liquid.ai/blog/d1-2
- Record: https://forck.live/items/36681-d1-2-state-of-the-art-omnimodal-decision-models
- Subject: LFM / d1 models

Liquid AI released d1.2 and d1.2-ultra, new versions of its decision models that add audio input and five new tasks, with d1.2-ultra ranking first among decision models on the Artificial Analysis Decision Index v2.1. The five new tasks let the models read a value from an input, find every passage of a kind in a text, locate objects in a picture, estimate a number with a range, and pick every option that applies. Both models are available today on the Liquid API, at $0.02 for d1.2 and $0.042 per million input tokens for d1.2-ultra, and both bill input tokens only. d1.2-ultra scores 75 on the Decision Index v2.1, the highest on the index and two points ahead of the previous best decision model, and Liquid AI says no model scores higher than d1.2-ultra at any price. Across 13 benchmarks and 22,505 questions answered by all five models, d1.2-ultra's calibration error is 0.8 points, compared with 1.2 for d1.2, 2.7 for Jev 1.14, 3.7 for Perplexity's Decider v1.3, and 7.2 for GPT-6 Luna on its Decisions API. On internal test sets, d1.2-ultra reads a value from a receipt at 98% and d1.2 at 96%, and on held-out recordings both models answer 95% to 97% of questions about which sound they hear or which language is spoken.

### Cohere — Video intelligence: Make every moment in your video library searchable

- Company: Cohere (cohere.com)
- Announced: 2026-11-16
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://cohere.com/blog/video-intelligence-preview
- Record: https://forck.live/items/36680-video-intelligence-make-every-moment-in-your-video-library-searchable
- Subject: Command

Cohere is adding Video Intelligence to Compass, letting enterprise search and AI agents find specific moments inside recorded meetings, training videos and media archives rather than whole files. Compass understands both what is said and what is shown on screen, returning a timestamped clip with a transcript and visual description instead of an hour-long recording. Rather than relying on fixed video segments, it determines the relevant start and end times at search time, and results link back to the original footage so users can verify the source. Video sits in the same retrieval system as documents, so applications and agents can pull footage alongside documents in RAG and agentic workflows, and an MCP server lets agents reason over slides, charts and demonstrations. Cohere says permission-aware retrieval limits users to authorized content, with configurable fields, model choices and deployment options. Cohere evaluated Compass against seven video embedding models across 11 datasets, including six public benchmarks and five internal enterprise video datasets, and says it ranks first with an nDCG@10 4.4 points ahead of the strongest alternative. Cohere is recruiting design partners for the preview.

### Legora — Bringing the power of Legora across Microsoft 365

- Company: Legora (legora.com)
- Announced: 2026-11-16
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://legora.com/blog/bringing-the-power-of-legora-across-microsoft-365
- Record: https://forck.live/items/36649-bringing-the-power-of-legora-across-microsoft-365
- Subject: Legora

Legora is adding four ways to reach its legal agent from inside Microsoft 365, so legal teams can start and track work without leaving the tools they already use. Intake for Teams and Outlook lets colleagues mention "@Legora" in Teams or email a connected shared mailbox, and Legora gathers the necessary information and turns the request into a tracked task that the Agent can then work on using the team's Skills and playbooks, with a lawyer accountable throughout. A new Outlook email and Calendar integration lets the Legora Agent search connected emails and attachments and find relevant calendar details, so it can retrieve the correspondence behind a negotiating position or identify meetings linked to a matter. The Excel add-in lets the Agent populate templates, build formulas, format data and create pivot tables directly in a workbook, working with selected cells and asking for approval before making changes. A Legora plugin in Copilot for Legal lets users ask the Agent to research legal questions, compare contractual clauses and assess risks from Copilot, then continue in Legora to review the underlying sources. The additions build on existing integrations with Word, Outlook, SharePoint and OneDrive, and Legora says it will be at Microsoft Ignite in San Francisco, November 17–20, at booth #1726.

### H Company — H Company launches its NVIDIA-powered agentic AI on the AWS European Sovereign Cloud

- Company: H Company (hcompany.ai)
- Announced: 2026-11-16
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://hcompany.ai/newsroom/h-company-aws-european-sovereign-cloud-press-release
- Record: https://forck.live/items/36614-h-company-launches-its-nvidia-powered-agentic-ai-on-the-aws-european-sovereign
- Subject: Holo

H Company has made its agentic platform available on the AWS European Sovereign Cloud, giving European organizations access to its computer-use AI agents on infrastructure located entirely within the European Union. The deployment is aimed at regulated industries and the public sector, which need advanced AI while keeping customer-created content and customer metadata inside the EU. H Company says it runs EU customer inference on NVIDIA-powered instances in the AWS European Sovereign Cloud, while continuing to train its models on AWS infrastructure globally. The cloud runs under an independent sovereign governance structure, operated exclusively by EU residents and transitioning to EU citizens, with no critical dependencies on non-EU infrastructure, according to the announcement. H Company's agents click, type and navigate on screen to work with legacy systems without API integrations or changes to those systems, which means they process live organizational data. A highly regulated European financial services company will be the first to use the agents on the AWS European Sovereign Cloud. H Company says its latest model version, Holo4, leads computer use benchmarks at much lower costs than general-purpose models.

### OpenAI — The King’s Trust and OpenAI launch £1m partnership

- Company: OpenAI (openai.com)
- Announced: 2026-11-16
- Category: partnership-acquisition
- Coverage: 1 outlet
- Announcement: yes
- Group: covered
- Source: https://openai.com/index/the-kings-trust-and-openai
- Record: https://forck.live/items/36610-the-king-s-trust-and-openai-launch-1m-partnership
- Subject: GPT / ChatGPT / API

The King's Trust has launched a £1 million partnership with OpenAI to help young people across the UK build AI skills for work, prioritizing those not in education, employment or training. OpenAI is making the £1m donation to the youth charity, and the program aims to support up to 25,000 young people over the next 18 months. It forms part of a wider program of nationally recognized AI qualifications The Trust says it will launch in the coming months. The partnership will combine AI skills learning with recognized qualifications and practical experience, working with employers, AI specialists and young people. The Open University's Business Barometer 2026 found that 68% of 16–24-year-olds surveyed who were NEET would be willing to train or upskill to improve their job prospects. For OpenAI, the partnership extends work that includes OpenAI Academy, which provides free learning resources for students and job-seekers. The announcement comes as The King's Trust marks its 50th year, having supported more than 1.3 million young people.
