OpenAI publishes hundreds of machine-formalized math results, and mathematicians push back

Описание: Inspect is the tool safety researchers use to read back what an A I agent did during an evaluation. Its transcript viewer had a flaw in how it rendered math notation. A rendering flaw. AlphaSignal reports METR researchers found that exploit, and that it lets agents hide actions from the reviewers reading the transcript. The record of what the agent did. Hidden by the agent. AlphaSignal's phrase for it is that observability is a security surface. The window you look through to check the work is also a thing to attack. Lead story: Formalized for the machine.


Learn this properly:

• Automated Theorem Proving: Proof Search Meets Language Models (https://aiuniversity.com/learn/#/less...)

Sources:

• OpenAI Shares New Math Research Results and Lean Proofs From a Frontier Model (https://aiuniversity.com/#/news/aHR0c...) — OpenAI RSS
• OpenAI Releases 722 Math Manuscripts From Unreleased Model, Stirring Fresh Debate (https://aiuniversity.com/#/news/aHR0c...) — The Verge AI
• Democratic Rep. Lori Trahan Rolls Out 'CLAIM Act' AI Liability Proposal (https://aiuniversity.com/#/news/aHR0c...) — Semafor Tech
• Mathematicians criticize OpenAI as it prepares to release more than 100 solutions to unsolved problems (https://aiuniversity.com/#/news/aHR0c...) — Wired AI
• Hacker News Reacts as OpenAI's Math Project Reportedly Solves a Decades-Old Conjecture (https://aiuniversity.com/#/news/aHR0c...) — Simon Willison
• OpenAI publishes 722 AI-generated math manuscripts with Lean-verified proofs (https://aiuniversity.com/#/news/aHR0c...) — AlphaSignal
• Mistral previews Mistral Large 4, with open weights promised this month (https://aiuniversity.com/#/news/aHR0c...) — Simon Willison
• Mistral releases 'Le Chonk', a trillion-parameter model it calls the best open-weight option outside China (https://aiuniversity.com/#/news/aHR0c...) — Wired AI
• Mistral releases Mistral Large 4, a trillion-parameter multimodal model (https://aiuniversity.com/#/news/aHR0c...) — TechCrunch AI
• Mistral unveils Mistral Large 4, a 1-trillion-parameter MoE flagship with open weights due late October (https://aiuniversity.com/#/news/aHR0c...) — AlphaSignal
• Microsoft publishes Nobel economist Daron Acemoglu's bearish AI forecast: about 1.5% GDP growth over a decade (https://aiuniversity.com/#/news/aHR0c...) — The Decoder
• Google will limit free Gemini users to Flash Lite starting October 9 (https://aiuniversity.com/#/news/aHR0c...) — The Verge AI
• OpenAI Cuts Its API Usage Tiers From Five to Three and Halves the Top-Tier Threshold (https://aiuniversity.com/#/news/aHR0c...) — AlphaSignal
• Free Gemini Users Will Soon Be Limited to Google's Weakest Model, ZDNet Reports (https://aiuniversity.com/#/news/aHR0c...) — ZDNet AI
• EmbeddingGemma 2: an open, lightweight multimodal embedding model (https://aiuniversity.com/#/news/aHR0c...) — Google DeepMind (blog.google)
• Google's EmbeddingGemma 2 Claims to Beat Larger Rival Embedding Models (https://aiuniversity.com/#/news/aHR0c...) — The Decoder
• EmbeddingGemma 2 (https://aiuniversity.com/#/news/aHR0c...) — Simon Willison
• FICO Cuts 15% of Workforce as Part of AI-Driven Restructuring (https://aiuniversity.com/#/news...