# forck.live — 15–21 September 2025

> AI lab announcements, news reports and podcasts.

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

This file is one issue of the weekly archive: every confirmed
first-party announcement of ISO week 2025-W38, 15–21 September 2025, newest
first. The window never moves, so this URL always names the same seven days —
but a post published inside them and detected later is added when it arrives,
so a closed week can still gain entries.

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

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

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

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

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

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

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

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

- This week's page: https://forck.live/week/2025-W38
- The current issue, and the catalogue of the other weeks: https://forck.live/briefing.md
- Catalogue: https://forck.live/llms.txt

## Announcements

### Google Research — Deep researcher with test-time diffusion

- Company: Google Research (research.google)
- Announced: 2025-09-19T20:43:00+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://research.google/blog/deep-researcher-with-test-time-diffusion/
- Record: https://forck.live/items/1417-deep-researcher-with-test-time-diffusion
- Subject: Research

Google Research introduces Test-Time Diffusion Deep Researcher (TTD-DR), a framework that models research report writing as a diffusion process, iteratively drafting and revising reports using retrieved information. It achieves state-of-the-art results on long-form report writing and multi-hop reasoning tasks.

### Vercel — How we made global routing faster with Bloom filters

- Company: Vercel (vercel.com)
- Announced: 2025-09-19T13:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://vercel.com/blog/how-we-made-global-routing-faster-with-bloom-filters
- Record: https://forck.live/items/20816-how-we-made-global-routing-faster-with-bloom-filters
- Subject: v0 / AI SDK / AI Gateway

Vercel optimized its global routing service by replacing a slow JSON path-lookup file with a Bloom filter, reducing memory usage by 15%, improving TTFB from the 75th percentile and above by 10%, and cutting path lookup latency to nearly zero for websites with many static paths.

### Harvey — 5 Questions With K-Ming Lee

- Company: Harvey (harvey.ai)
- Announced: 2025-09-19T13:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.harvey.ai/blog/5-questions-with-k-ming-lee
- Record: https://forck.live/items/17674-5-questions-with-k-ming-lee
- Subject: Harvey

Harvey's Head of Customer Engagement K-Ming Lee discusses the company's customer-obsessed approach, plans for community programming in New York, London, and Sydney, and the creation of customer advisory boards for in-house teams and law firms.

### xAI — Grok 4 Fast

- Company: xAI (x.ai)
- Announced: 2025-09-19
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://x.ai/news/grok-4-fast
- Record: https://forck.live/items/3485-grok-4-fast
- Subject: Grok

xAI announced Grok 4 Fast, a model pushing the frontier of cost-efficient intelligence.

### Hugging Face — Scaleway on Hugging Face Inference Providers 🔥

- Company: Hugging Face (huggingface.co)
- Announced: 2025-09-19
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/inference-providers-scaleway
- Record: https://forck.live/items/1618-scaleway-on-hugging-face-inference-providers
- Subject: Platform

Hugging Face announces that Scaleway is now a supported Inference Provider on the Hugging Face Hub, offering serverless access to models like gpt-oss, Qwen3, DeepSeek R1, and Gemma 3 with pay-per-token pricing starting at €0.20 per million tokens.

### Replit — How Replit is Protecting You From the "Shai-Hulud" Worm

- Company: Replit (replit.com)
- Announced: 2025-09-18T22:20:48+00:00
- Category: safety-policy-update
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://replit.com/blog/npm-supply-chain-attack
- Record: https://forck.live/items/17939-how-replit-is-protecting-you-from-the-shai-hulud-worm
- Subject: Replit Agent

Replit blocked the exfiltration endpoint used by the Shai-Hulud worm across all development environments and upgraded its Security Scanner with a Malicious File Detection feature to detect known malicious files from the attack. The AI agent can automatically remediate issues such as removing malicious files and updating to safe package versions.

### Stability AI — Stability AI Brings Image Services to Amazon Bedrock, Delivering End-to-End Creative Control with Enterprise-Grade Infrastructure

- Company: Stability AI (stability.ai)
- Announced: 2025-09-18T21:03:08+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://stability.ai/news-updates/stability-ai-brings-image-services-to-amazon-bedrock-delivering-professional-creative-control-with-enterprise-grade-infrastructure
- Record: https://forck.live/items/17481-stability-ai-brings-image-services-to-amazon-bedrock-delivering-end-to-end
- Subject: Stable Diffusion

Stability AI has made its Image Services suite available on Amazon Bedrock, providing nine API-based image editing tools including Inpaint, Erase Object, Remove Background, Search and Recolor, Structure, Style Transfer, Style Guide, and Sketch. The tools are designed for enterprise-grade applications and complement existing image generation models already on Bedrock, such as Stable Diffusion 3.5, Stable Image Core, and Stable Image Ultra.

### Google Research — Sensible Agent: A framework for unobtrusive interaction with proactive AR agents

- Company: Google Research (research.google)
- Announced: 2025-09-18T20:10:00+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://research.google/blog/sensible-agent-a-framework-for-unobtrusive-interaction-with-proactive-ar-agents/
- Record: https://forck.live/items/1418-sensible-agent-a-framework-for-unobtrusive-interaction-with-proactive-ar-agents
- Subject: Research

Sensible Agent is a framework designed for unobtrusive interaction with proactive AR agents. It uses real-time multimodal context sensing, subtle gestures, gaze input, and minimal visual cues to offer contextually-appropriate assistance.

### Vercel — Scale to one: How Fluid solves cold starts

- Company: Vercel (vercel.com)
- Announced: 2025-09-18T13:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://vercel.com/blog/scale-to-one-how-fluid-solves-cold-starts
- Record: https://forck.live/items/20818-scale-to-one-how-fluid-solves-cold-starts
- Subject: v0 / AI SDK / AI Gateway

Vercel announced that its Fluid compute platform achieves zero cold starts for 99.37% of requests, with fewer than one in a hundred requests experiencing a cold start. The approach combines scale-to-one (keeping at least one instance warm), Fluid compute's concurrent request handling, predictive scaling, bytecode caching, and rolling releases. Scale-to-one applies automatically to production deployments on Pro and Enterprise plans, and to the most recent branch deployment for Enterprise plans, with no additional cost beyond usage.

### Vercel — What you need to know about vibe coding

- Company: Vercel (vercel.com)
- Announced: 2025-09-18T13:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://vercel.com/blog/what-you-need-to-know-about-vibe-coding
- Record: https://forck.live/items/20817-what-you-need-to-know-about-vibe-coding
- Subject: v0 / AI SDK / AI Gateway

Vercel published a report on vibe coding, noting that over 90% of U.S. developers use AI coding tools and that English has become the fastest growing programming language. The report states that about 63% of users exploring vibe tools are non-developers, and that security must be built into the tools themselves.

### Luma — Luma AI launches Ray3

- Company: Luma (lumalabs.ai)
- Announced: 2025-09-18
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://lumalabs.ai/news/ray3
- Record: https://forck.live/items/7980-luma-ai-launches-ray3
- Subject: Ray / Dream Machine

Luma AI launched Ray3, a video model that can reason in visuals and concepts, evaluate its own outputs, and refine results. It is the first model to produce video in true 10-, 12-, and 16-bit HDR ACES2065-1 EXR format and includes a Draft Mode for faster creative exploration. Ray3 is available in the Dream Machine platform and through partners including Adobe Firefly.

### Hugging Face — Democratizing AI Safety with RiskRubric.ai

- Company: Hugging Face (huggingface.co)
- Announced: 2025-09-18
- Category: developer-tool-release
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://huggingface.co/blog/riskrubric
- Record: https://forck.live/items/1619-democratizing-ai-safety-with-riskrubric-ai
- Subject: Platform

Announcement of RiskRubric.ai, a standardized risk assessment platform for AI models led by Cloud Security Alliance and Noma Security, evaluating models across six pillars (transparency, reliability, security, privacy, safety, reputation) and providing A-F letter grades.

### Stability AI — Stability AI’s Annual Integrity Transparency Report

- Company: Stability AI (stability.ai)
- Announced: 2025-09-17T17:26:52+00:00
- Category: safety-policy-update
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://stability.ai/news-updates/stability-ais-annual-integrity-transparency-report
- Record: https://forck.live/items/17482-stability-ai-s-annual-integrity-transparency-report
- Subject: Stable Diffusion

Stability AI published its annual integrity transparency report covering April 2024 to April 2025, detailing its safety-by-design approach for video, image, 3D, and audio models. The report states that no CSAM or CSEM was detected in training datasets and that 100% of generative AI models were stress-tested for CSAM/CSEM capabilities with zero issues found. It also describes red teaming practices, including collaboration with UK law enforcement, and enforcement of an Acceptable Use Policy requiring users to be 18 or older.

### Google Research — Making LLMs more accurate by using all of their layers

- Company: Google Research (research.google)
- Announced: 2025-09-17T17:00:00+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://research.google/blog/making-llms-more-accurate-by-using-all-of-their-layers/
- Record: https://forck.live/items/1419-making-llms-more-accurate-by-using-all-of-their-layers
- Subject: Research

Google Research introduces SLED (Self Logits Evolution Decoding), a decoding strategy that improves LLM factual accuracy by using all layers of the model rather than just the last layer. It requires no external data or fine-tuning and was featured at NeurIPS 2024.

### Vercel — Addressing security and quality issues with MCP tools in AI Agent

- Company: Vercel (vercel.com)
- Announced: 2025-09-17T13:00:00+00:00
- Category: developer-tool-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://vercel.com/blog/generate-static-ai-sdk-tools-from-mcp-servers-with-mcp-to-ai-sdk
- Record: https://forck.live/items/20819-addressing-security-and-quality-issues-with-mcp-tools-in-ai-agent
- Subject: v0 / AI SDK / AI Gateway

Vercel released mcp-to-ai-sdk, a CLI that generates static AI SDK tool definitions from any MCP server. The tool addresses security, cost, and quality issues by vendoring tool schemas and descriptions into the codebase, preventing unexpected changes from upstream servers.

### Perceptron — Introducing Isaac 0.1

- Company: Perceptron (perceptron.inc)
- Announced: 2025-09-17
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://www.perceptron.inc/blog/introducing-isaac-0-1
- Record: https://forck.live/items/18622-introducing-isaac-0-1
- Subject: Perceptron Isaac / Perceptron Mk1

Perceptron released Isaac 0.1, an open-source 2B-parameter perceptive-language model for real-world applications. The model matches or exceeds the performance of models over 50 times its size on key perceptive benchmarks while using fewer weights. It is available for researchers and developers.

### Hugging Face — Public AI on Hugging Face Inference Providers 🔥

- Company: Hugging Face (huggingface.co)
- Announced: 2025-09-17
- Category: infrastructure-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/inference-providers-publicai
- Record: https://forck.live/items/1620-public-ai-on-hugging-face-inference-providers
- Subject: Platform

Public AI is now a supported Inference Provider on the Hugging Face Hub, offering free access to public and sovereign models from institutions like the Swiss AI Initiative and AI Singapore. The service is free of charge at the time of writing.

### OpenAI — Detecting and reducing scheming in AI models

- Company: OpenAI (openai.com)
- Announced: 2025-09-17
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://openai.com/index/detecting-and-reducing-scheming-in-ai-models
- Record: https://forck.live/items/466-detecting-and-reducing-scheming-in-ai-models
- Subject: GPT / ChatGPT / API

Apollo Research and OpenAI developed evaluations for hidden misalignment (scheming) and found scheming behaviors in controlled tests across frontier models. They also shared examples and stress tests of an early method to reduce scheming.

### Replicate — Introducing our new search API

- Company: Replicate (replicate.com)
- Announced: 2025-09-16T16:00:00+00:00
- Category: api-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://replicate.com/blog/new-search-api
- Record: https://forck.live/items/2478-introducing-our-new-search-api
- Subject: Platform

Replicate launched a new search API in beta, available via TypeScript and Python SDKs and MCP servers, returning models, collections, and documentation pages with enhanced metadata.

### Supabase — Defense in Depth for MCP Servers

- Company: Supabase (supabase.com)
- Announced: 2025-09-16T15:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://supabase.com/blog/defense-in-depth-mcp
- Record: https://forck.live/items/21333-defense-in-depth-for-mcp-servers
- Subject: Supabase

Over the past few months, there’s been renewed discussion around the risks of connecting MCP servers to databases containing private data.

### Supabase — Processing large jobs with Edge Functions, Cron, and Queues

- Company: Supabase (supabase.com)
- Announced: 2025-09-16T15:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://supabase.com/blog/processing-large-jobs-with-edge-functions
- Record: https://forck.live/items/21332-processing-large-jobs-with-edge-functions-cron-and-queues
- Subject: Supabase

Supabase's guide explains how to use Edge Functions, Cron, and database queues together to build a resilient, scalable system for processing large data jobs. It walks through a three-layer pattern of collection, distribution, and processing, using an NFL news aggregator as a real-world example.

### Vercel — AI agents at scale: Rox’s Vercel-powered revenue operating system

- Company: Vercel (vercel.com)
- Announced: 2025-09-16T13:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://vercel.com/blog/ai-agents-at-scale-roxs-vercel-powered-revenue-operating-system
- Record: https://forck.live/items/20820-ai-agents-at-scale-rox-s-vercel-powered-revenue-operating-system
- Subject: v0 / AI SDK / AI Gateway

Rox, a revenue operating system company, uses Vercel's infrastructure to deploy AI agents that research and engage prospects. Vercel's serverless functions batch thousands of concurrent AI agent calls, and the Vercel AI SDK powers Rox's chat-driven interface, Rox Command. The company reports 3-5 minute frontend deployments and a 2.5x reduction in P95 load speed.

### OpenAI — Teen safety, freedom, and privacy

- Company: OpenAI (openai.com)
- Announced: 2025-09-16T06:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/teen-safety-freedom-and-privacy
- Record: https://forck.live/items/469-teen-safety-freedom-and-privacy
- Subject: GPT / ChatGPT / API

Their approach to teenager safety, freedom, and privacy.

### OpenAI — Building towards age prediction

- Company: OpenAI (openai.com)
- Announced: 2025-09-16T06:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/building-towards-age-prediction
- Record: https://forck.live/items/468-building-towards-age-prediction
- Subject: GPT / ChatGPT / API

OpenAI is building age prediction and parental controls for ChatGPT to create safer, age-appropriate experiences for teens.

### World Labs — Generating Bigger and Better Worlds

- Company: World Labs (worldlabs.ai)
- Announced: 2025-09-16
- Category: product-launch
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.worldlabs.ai/blog/bigger-better-worlds
- Record: https://forck.live/items/7346-generating-bigger-and-better-worlds
- Subject: Marble / Atlas

World Labs announced Marble, a limited access beta preview of their model that generates 3D worlds from images or text prompts. The generated worlds are bigger, more stylistically diverse, and have cleaner 3D geometry compared to previous results. Users can explore worlds in their browser at no cost and export them as Gaussian splats using the open-source rendering library Spark.

### Hugging Face — `LeRobotDataset:v3.0`: Bringing large-scale datasets to `lerobot`

- Company: Hugging Face (huggingface.co)
- Announced: 2025-09-16
- Category: developer-tool-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/lerobot-datasets-v3
- Record: https://forck.live/items/1621-lerobotdataset-v3-0-bringing-large-scale-datasets-to-lerobot
- Subject: Platform

Hugging Face releases LeRobotDataset:v3.0, a new version of its standardized dataset format for robot learning. The update packs multiple episodes per file to overcome file-system limitations, adds native streaming support, and provides a one-liner conversion utility.

### Vercel — Introducing Vercel Drains: Complete observability data, anywhere

- Company: Vercel (vercel.com)
- Announced: 2025-09-15T13:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://vercel.com/blog/introducing-vercel-drains
- Record: https://forck.live/items/20822-introducing-vercel-drains-complete-observability-data-anywhere
- Subject: v0 / AI SDK / AI Gateway

Vercel expanded its Log Drains into Vercel Drains, adding support for OpenTelemetry traces, Web Analytics events, and Speed Insights metrics. The unified pipeline streams observability data to external tools via custom HTTP endpoints or turnkey integrations, with automatic correlation between logs and traces. Drains are available on Pro and Enterprise plans at $0.50 per GB.

### Vercel — Helly Hansen migrated to Vercel and drove 80% Black Friday growth

- Company: Vercel (vercel.com)
- Announced: 2025-09-15T13:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://vercel.com/blog/how-helly-hansen-migrated-to-vercel-and-drove-80-black-friday-growth
- Record: https://forck.live/items/20821-helly-hansen-migrated-to-vercel-and-drove-80-black-friday-growth
- Subject: v0 / AI SDK / AI Gateway

Helly Hansen migrated its ecommerce platform to Next.js and Vercel, improving Core Web Vitals from red to green, reducing load times from over 4 seconds to 1.8–2 seconds, and achieving 80% year-over-year growth on Black Friday Cyber Monday with zero downtime. The company also increased its A/B testing velocity from one test per month to 5–6 per sprint and saw a 21% ecommerce growth while the industry remained flat.

### OpenAI — Introducing upgrades to Codex

- Company: OpenAI (openai.com)
- Announced: 2025-09-15T10:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://openai.com/index/introducing-upgrades-to-codex
- Record: https://forck.live/items/470-introducing-upgrades-to-codex
- Subject: GPT / ChatGPT / API

OpenAI announced upgrades to Codex, making it faster, more reliable, and better at real-time collaboration and independent task handling across various development environments.

### OpenAI — How people are using ChatGPT

- Company: OpenAI (openai.com)
- Announced: 2025-09-15T03:00:00+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/how-people-are-using-chatgpt
- Record: https://forck.live/items/471-how-people-are-using-chatgpt
- Subject: GPT / ChatGPT / API

OpenAI published new research from the largest study of ChatGPT use, showing how the tool creates economic value through personal and professional use, with adoption broadening beyond early users.

### Together AI — Improved Batch Inference API: Enhanced UI, Expanded Model Support, and 3000× Rate Limit Increase

- Company: Together AI (together.ai)
- Announced: 2025-09-15
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.together.ai/blog/batch-inference-api-updates-2025
- Record: https://forck.live/items/2408-improved-batch-inference-api-enhanced-ui-expanded-model-support-and-3000-rate
- Subject: Inference platform

Together AI announced major improvements to its Batch Inference API, including a streamlined UI, support for all serverless models and private deployments, a 3000x rate limit increase (from 10M to 30B tokens), and 50% cost reduction compared to the real-time API.

### Hugging Face — Visible Watermarking with Gradio

- Company: Hugging Face (huggingface.co)
- Announced: 2025-09-15
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/watermarking-with-gradio
- Record: https://forck.live/items/1622-visible-watermarking-with-gradio
- Subject: Platform

Hugging Face announces that its Gradio library now supports visible watermarking for images, video, and text with a single command, making it easy to add attribution to AI-generated content.

### OpenAI — Addendum to GPT-5 system card: GPT-5-Codex

- Company: OpenAI (openai.com)
- Announced: 2025-09-15
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://openai.com/index/gpt-5-system-card-addendum-gpt-5-codex
- Record: https://forck.live/items/472-addendum-to-gpt-5-system-card-gpt-5-codex
- Subject: GPT / ChatGPT / API

OpenAI announces GPT-5-Codex, a version of GPT-5 optimized for agentic coding in Codex, with dynamic thinking effort adjustment based on task complexity.