Simon Willison

Highlight

The standout post today is a detailed exploration of OpenAI’s agentic resolution of the Navier–Stokes Millennium Prize Problem and the surrounding ethical controversy over research priority and user data usage. It raises crucial questions about whether rumors alone can trigger millions in compute to scoop researchers, as well as how AI labs leverage user session data to train future models.

Posts

On the Navier–Stokes Millennium Prize Problem · OpenAI OpenAI announced that an unreleased model resolved the Navier–Stokes Millennium Prize Problem using 4.9 million messages and 300 billion output tokens (~$15M public API equivalent) verified via Lean. However, the milestone triggered allegations from NYU professor Tristan Buckmaster and Anthropic researcher Levent Alpöge, who had spent a year working on the proof using OpenAI Codex before OpenAI launched a massive agent run after hearing rumors. Simon highlights the troubling precedent of “rumour as an exploit,” comparing it to security vulnerability hunting where a mere whisper can prompt massive compute deployment to scoop human researchers. He also reiterates ongoing privacy concerns about how AI labs utilize user session data to train models that might later replicate user discoveries.

Introducing ChatGPT Images 2.5 · OpenAI OpenAI released ChatGPT Images 2.5, introducing two new API models (gpt-image-2.5-sunburst for precise editing and gpt-image-2.5-flare for fast everyday generation) with improved multi-turn instruction following and subject preservation. Simon quickly updated his Python CLI tool openai_image.py to support multi-image reference inputs via uv. He tested the new capabilities by feeding an existing chart image into gpt-image-2.5-sunburst and prompting it to add a thoughtful raccoon scientist studying the chart.

llm 0.34 · GitHub Release The main release of Simon’s core llm CLI tool adds response duration tracking in milliseconds and human-readable format to llm logs --usage, alongside a new duration_ms field in llm logs --short. It also incorporates community bug fixes and delivers a significant performance boost to log loading thanks to contributor waveplate. This release gives developers better visibility into model latency when running local or remote LLM workflows from the terminal.

llm-anthropic 0.28 · GitHub Release This update to the llm-anthropic plugin adds support for Claude Fable 5.1 and enables reasoning traces by default for models that support them. It also introduces a dedicated llm_anthropic.ClaudeRefusal exception to gracefully handle API refusal responses. It ensures seamless integration when experimenting with Anthropic’s latest reasoning models through the llm CLI.

llm-openrouter 0.7.1 · GitHub Release A maintenance patch for the llm-openrouter plugin that resolves performance slowdowns when loading OpenRouter models. Contributed by GitHub user waveplate, this fix complements the performance optimizations landed in the core llm 0.34 release.

Project Pulse

Today’s updates highlight rapid iteration across Simon’s open-source ecosystem, combining performance tuning and new reasoning model support in the core llm CLI family alongside immediate script updates (openai_image.py) for newly dropped image APIs.

💡 If you’d like, I can help you draft a quick summary table comparing how Simon’s llm CLI plugins handle model-specific features like reasoning traces and usage logs across providers.