AI Daily · 2026-09-02

Today’s main thread was simultaneous model updates from Google and Anthropic: Google launched Gemini 3.7 Flash, a low-cost model aimed at coding and agentic workloads, alongside agentic video understanding, the Google Pics image tool and the Pixel 11 lineup, with Gemini app monthly active users passing one billion. Anthropic released Claude Fable 5.1, showing large gains on Terminal-Bench-Science and other long-horizon reasoning benchmarks, though Simon Willison’s tests found that output quality depends heavily on the reasoning tier and still trails Gemini 3.7 Flash on style. In applied AI, OpenAI opened ChatGPT to EHR and healthcare data connections, while NVIDIA and CrowdStrike introduced the SafeMind agentic cybersecurity system. On the open-source and research side, Python 3.15 hit RC2, datasette-mcp 0.2 shipped, major AI projects are using in-house agents to triage and merge instead of external PRs, and Google published MAPL-EMIT for detecting methane plumes from satellite data.

North America · First-hand

OpenAI

⭐⭐ [Product Update] Healthcare organizations can now connect EHR and additional industry data to ChatGPT

OpenAI announced that ChatGPT can now connect to trusted healthcare data, helping clinicians securely access patient context, medical research, and more. The feature is aimed at healthcare organizations and supports connecting electronic health records (EHR) and additional industry data to ChatGPT.

Google

⭐⭐⭐ [Model Release] The latest AI news we announced in August 2026

Google recapped its August 2026 AI announcements, highlighting the release of Gemini 3.7 Flash, a cost-efficient model for coding and agents launched just three weeks after 3.6 Flash with an introductory price half that of its predecessor per million tokens. The company also unveiled the Pixel 11 series, powered by the Tensor G6 chip running the latest Gemini Nano. The Gemini app surpassed 1 billion monthly active users. Additional updates include a free one-year Google AI plan for eligible students and new AI-powered learning features in Search.

⭐⭐ [Product Update] Introducing agentic video understanding with Gemini

Google DeepMind has launched agentic video understanding across Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite. Instead of static fixed-frame-rate processing, the feature lets Gemini dynamically decide which video segments to watch, at what speed, and through which modality using native video tools. On standard benchmarks, it cuts token consumption by up to 88% and costs by up to 66% while improving accuracy by up to 7%, with Gemini 3.7 Flash landing at the accuracy-to-cost Pareto frontier. The feature is available today for video uploads and YouTube videos via the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform by setting the API configuration to "agentic". Use cases include sub-second moment retrieval, anomaly detection, and precise counting.

⭐⭐ [Product Update] Try Google Pics: Easy image creation and editing in Google Workspace

Google announced Google Pics, an AI image creation and editing tool built on the Nano Banana model, rolling out in the coming weeks to Google AI Pro and Ultra subscribers and most Workspace business customers. It will be both a standalone product and integrated into Workspace apps such as Slides, Docs, and Drive, allowing users to edit images without switching tabs. Key features include object segmentation and targeted edits, in-image text editing and translation, real-time collaboration, and multiple image generations from a single prompt. Users can start using it today in Docs and Slides via pics.new.

⭐⭐ [Research] Mapping global methane emissions from space with deep learning

Google Research scientists published a paper in PNAS introducing MAPL-EMIT, a deep-learning framework that automates detection, enhancement quantification, and source estimation of methane plumes from NASA EMIT hyperspectral satellite data. It achieves 84% recall on expert-annotated plumes and higher signal-to-noise than existing matched-filter-based enhancement methods. Google is releasing a global plume database on Earth Engine, the trained model and synthetic plumes on Kaggle, and an inference library on GitHub. The post notes that methane has about 30 times the 100-year warming potential of carbon dioxide and accounts for roughly 25% of industrial-era human-induced warming, framing point-source detection as support for the Global Methane Pledge.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐⭐ [Model Release] Claude Fable 5.1 made me a really nice animated pelican

Anthropic released Claude Fable (and Mythos) 5.1 today, claiming it sets a new standard for coding, knowledge work, and long-running problem-solving. On the Terminal-Bench-Science 0.1 benchmark, Fable 5.1 scored 52.6%, up from 24.7% for Fable 5, 29.0% for Opus 5 and 22.4% for GPT-5.6 Sol, while other benchmarks improved only modestly. Simon Willison tested the model by asking it to generate an SVG of a pelican riding a bicycle across five reasoning levels: low and medium produced no visible reasoning, high only a little, and xhigh and max scaled up to about 36.8k and 65.9k output tokens, taking 7m51s and 13m54s and costing $1.83 and $3.30 respectively. He says max yielded the best pelican he has seen from any Anthropic model, though the result still lacked the flair of Gemini 3.7 Flash.

Product Update

⭐⭐ [Product Update] NVIDIA and CrowdStrike Strengthen Agentic Cybersecurity Frontier

NVIDIA and CrowdStrike announced a partnership at Fal.Con 2026 to strengthen the agentic cybersecurity frontier. NVIDIA CEO Jensen Huang joined CrowdStrike CEO George Kurtz to unveil CrowdStrike SafeMind, an agentic cybersecurity system developed by CrowdStrike. Huang noted that cybersecurity is at an inflection point where attacks are now automated, and so defense must be as well.

⭐⭐ [Product Update] Codex bundles LibreOffice

Simon Willison noticed while browsing his ~/.cache/ directory that the OpenAI Codex desktop app (now rebranded as ChatGPT) has about 1.7GB of runtime data in a folder called codex-primary-runtime, including full Python and Node.js installations plus native binaries for Poppler, git, and LibreOffice, the open-source office suite forked from OpenOffice.org in 2010. A documents plugin folder under that cache contains skills that tell Codex how to locate and use those binaries.

⭐⭐ [Product Update] datasette-mcp 0.2

Simon Willison released datasette-mcp 0.2, the first non-alpha release of the Datasette plugin. It adds a /-/mcp MCP server endpoint to any Datasette instance. The rows returned by execute_sql are now an array of objects instead of an array of arrays, which helps weaker models keep track of column mappings. The plugin now depends on mcp>=2.1.1.

⭐⭐ [Product Update] Python 3.15.0 candidate 2 is here!

Python 3.15.0 release candidate 2 (RC2) has been published, with the final release scheduled for October. Release manager Hugo van Kemenade says only clear bug fixes are allowed after this RC and encourages third-party maintainers to prepare projects and publish Python 3.15 wheels on PyPI. Simon Willison recommends testing during the RC period and shares a GitHub Actions matrix configuration using allow-prereleases and check-latest so the latest RC is picked up automatically. He reports that Datasette and sqlite-utils pass, while LLM is currently blocked by a missing scikit-learn 3.15 wheel.

Research

⭐⭐ [Research] BenchMIRT: What are LLM benchmarks actually measuring?

Ai2 introduces BenchMIRT, a method to audit LLM benchmarks at the level of individual prompts. Built on multidimensional Item Response Theory (MIRT) from psychometrics, it estimates model capabilities as well as item difficulty and discrimination at both the model and question level. BenchMIRT was trained on evaluation results from 100 LLMs across 16 benchmarks and over 34K questions, without being told which benchmarks target which capabilities, and it independently recovered safety and general reasoning as the two dominant dimensions. The analysis shows that BBQ, a social-bias benchmark often grouped with safety, aligns much more strongly with general reasoning, suggesting that averaging question scores into a single benchmark number can obscure such differences.

Opinion

⭐⭐ [Opinion] PRs NOT Welcome: How Top AI Open Source Projects Are Managing Thousands of Contributors

The article reports that top AI-native open source projects such as Vercel's AI SDK, Astro, Flue, and tldraw are moving away from open community pull requests, adopting a software factory model in which maintainer-run agent teams triage issues, reproduce bugs, implement and review fixes, and hand them to humans for merging. Vercel says that four weeks after deploying its factory, agents authored 25-35% of merged PRs and closed 70-80% of issues, after the AI SDK backlog had grown to over 1,000 open issues and nearly 800 PRs. Astro adopted auto-triage, and its founder says the situation reversed within six months as issues became a prioritized weekly task list rather than an ever-trimmed backlog. Vercel engineer Lars Grammel argues that maintainers can trust their own optimized and historically validated agent configurations more than community-run agents, which reduces review time.

Other

⭐⭐ [Other] Quoting Rick Brewster

Rick Brewster, author of Paint.NET, says Direct2D has always been the biggest obstacle to running Paint.NET on WINE and cannot simply be disabled. He explains that Paint.NET now ships an internal, from-scratch, clean-room reverse-engineered rewrite of Direct2D, used on WINE via the /wine flag and living in PaintDotNet.Windows.Direct2D1.Managed.dll. The roughly 180,000 lines of code were mostly written by Claude and are largely described as 'vibe coded' and not thoroughly reviewed, compared with the rest of Paint.NET's ~700,000 lines maintained over 20 years. Rick notes he had to supervise Claude closely, including fixing COM reference-counting issues and correcting bad design or architecture decisions, while praising Claude's reverse-engineering work on the formulas behind Direct2D's built-in effects library.