GPT‑6 is rolling out globally in ChatGPT with Intelligent UI, delivering faster responses with visuals and interactive experiences you can explore and use directly.

OpenAI has released 722 manuscripts from an unreleased frontier model, covering 372 result families and offering solutions to hundreds of open mathematics problems. The release, coordinated with the independent AGMAI advisory group, extends a run of breakthroughs that have both impressed and unsettled mathematicians, while raising questions about research ethics and academic conduct.

Discussion | Link

Simon Willison released llm-openai-decisions 0.1a0, an LLM plugin for OpenAI's new Jev-style Decisions API. Built with GPT-6 Astra reading the official docs and modeled on his earlier llm-typesafe plugin, it supports yes/no, choice, and score questions, and its gpt-6-luna model accepts image input as well as text.

Since 2018, MIT Lincoln Laboratory's Supercomputing Center has run the Lincoln AI Computing Survey, tracking commercial AI accelerators and comparing peak performance and power to help government sponsors and lab researchers understand a fast-moving hardware landscape.

In a comment on the Hacker News thread about EmbeddingGemma 2, Simon Willison welcomes the model's Apache 2.0 license and argues that embedding models in particular should not be closed, proprietary and hosted-only: once a vendor retires a model, users must pay to re-embed the millions of vectors they have already stored. He does not want to self-host, but wants the open weights as a fallback if a provider ever stops serving the model.

Mistral has released a preview of Mistral Large 4, a 1 trillion parameter model with 49 billion active parameters trained on its own cluster of 3,800 NVIDIA Grace Blackwell GPUs. The API preview is live with open weights promised by the end of the month, and its Artificial Analysis score of 38 is a huge jump from Mistral Large 3's 9 — though still roughly six months behind the frontier.

Simon Willison documents a TIL on feeding Datasette's new OpenTelemetry traces into Parseable, a new observability platform, after using Codex to work out the setup and sharing a trace screenshot.

More teams are running their dbt transformations on Databricks Lakehouse: an open...

Google DeepMind's EmbeddingGemma 2 maps text, code, images, video and audio into a single 768-dimensional space. The 740M-parameter model has an 8K token context window and ships under Apache 2.0, targeting on-device search, classification and privacy-first RAG. Weights are live on Hugging Face and Kaggle, with Ollama, llama.cpp GGUF and LiteRT builds available today.

Simon Willison's comment on the Mistral Large 4 discussion at Hacker News, in which he takes a jab about saturated benchmarks and actually runs four frontier models on an absurd SVG prompt.

Richard Tomlinson, who leads product marketing for Databricks' business intelligence...

Discussion | Link

Mistral AI has launched Mistral Large 4 (Le Chonk) in public preview: a granular Mixture of Experts model with 1.05 trillion total parameters, 49B active per token, a 1.6B-parameter vision encoder, and a 1M-token context window, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters. The API is live at $1.36 per 1M input and $4.18 per 1M output tokens, with cached input at $0.14; weights and license are promised for end of October 2026, so self-hosting is not yet possible. Its strongest results are in cybersecurity (93% Cybench, 82% CyberGym-E2E), where Mistral says several closed frontier models score near zero because they refuse the task.

MIT Provost Anantha Chandrakasan announced that Chris Bourg, director of the MIT Libraries since 2015, will become vice provost and Barbara K. Ostrom (1978) Director of the MIT Libraries. The endowed role recognizes her leadership in digital access, open and equitable scholarly publishing, data-intensive research support, and helping MIT navigate the scholarly and ethical implications of generative AI.

Discussion | Link

In a letter, Martyn Thomas asks how any rational person can claim that independent oversight and regulation will be sufficient, after yet another frontier AI model was withdrawn for failing safety tests. He notes that safety-critical software such as aircraft and nuclear power controls has long been governed by international standards requiring rigorous "safety cases".

SAP announced it will acquire Belgian AI company TechWolf, which maps employee work and skills, to enhance HR agents in SAP SuccessFactors. The deal, announced at Connect, is expected to close in Q4 pending approval.

Discussion | Link

With AI hardware, tech companies are pushing the definition of what does and doesn't constitute a recording. Apple's rumored smart home camera saves no video, Apple Watch Audio Intelligence produces transcripts while Apple says it doesn't record audio, and Google is floating cameras that are "just" an image sensor. Meanwhile privacy lights, form factors, and legal norms lag behind, leaving society without an agreed definition of recording — or of who gets to redefine it.

Discussion | Link

Jennifer Neville, partner research manager at Microsoft Research and professor at Purdue, discusses her winding path into AI, the importance of evaluation in pushing AI performance, and the 'surprising failures' that reveal gaps in how models handle complexity.

Discussion | Link

The Guardian is asking readers to share their experiences of using AI agents to handle their finances. The callout follows the release of Meta's Muse and OpenAI's dots, which give such agents the ability to carry out some financial transactions on a user's behalf.

OpenAI and Atlassian are expanding their partnership to bring GPT-6 family frontier models across Atlassian's platform and Rovo, combining them with Atlassian's Teamwork Graph enterprise context layer to power agents that help teams plan, build, and deliver work. The deal expands a collaboration begun in 2023 and includes broader access to frontier models, growing Codex and ChatGPT Enterprise adoption, CLI plugins for ChatGPT and Codex, and planned deeper Jira integrations for assigning, tracking, and reviewing AI agent work.

AbstractIFCO runs one of the world's largest reusable packaging pools with hundreds of millions of crates and pallets...

AWS explains how ISO/IEC 42005:2025 codifies AI system impact assessment best practices, covering what these assessments are, how they fit into enterprise-wide risk management, and how the standard guides their execution and documentation. The post also highlights AWS responsible AI tooling and ISO/IEC 42001 certification support.

A walkthrough of adding a voice travel concierge to an airline app using Amazon Bedrock AgentCore, Amazon Nova 2.5 Sonic for real-time speech, and Amazon Bedrock Knowledge Bases for policy answers. Travelers change seats, check delays, and ask policy questions by voice, while the agent reaches backend services through MCP tools and confirms every change before it writes.

Simon Willison tested whether Claude Opus 5.5 could compose music by asking it to define a simple text-based music format and build an artifact that plays example tracks. The result, Scrimshaw Jukebox, is a browser-based pixel-art player with six surprisingly good adventure-game tracks. Willison wonders whether competent music composition is a newly emergent text-model capability, similar to recent 3D generation, and says cross-model experiments are needed.

MIT launches MIT for America, a nationwide initiative to strengthen STEM education from kindergarten through community college, emphasizing math, hands-on making, and the constructive use of AI through programs such as calculus tutoring, MIT RAISE, PATH, fab labs, and design challenges.