Yesterday's email.

1. 01
   OpenAI let METR and Redwood Research embed for three weeks to test its internal monitoring of misaligned agents, and they found several new vulnerabilities

   It is one of the first times outside evaluators got deep inside access to a frontier lab's safety systems, and they found real gaps.

   The post by @OpenAI on X
2. 02
   Mustafa Suleyman says Microsoft's MAI Diagnostic system is 1.7x more accurate than emergency department physicians, with the results published in Science

   It is a peer reviewed medical AI paper in Science, but the accuracy figures are Microsoft's own, come from a text only benchmark and not a clinical trial, and the paper itself says they do not establish superiority in real world practice.

   The post by @mustafasuleyman on X
3. 03
   Anthropic's Frontier Red Team says Claude formally verified security properties in over 100 open source libraries, with Theorem

   If AI can write machine checked proofs about real code at this scale, proven security guarantees could move from rare, expert only work toward a routine part of software development.

   The post by @AnthropicAI on X
4. 04
   MiniMax publishes open weights for M3.1 Flash, a 428B parameter model with 23B active per token, citing benchmark scores that are its own figures

   Downloadable weights for a large multimodal coding and agent model let outside teams run and tune it themselves, but every performance number so far comes from MiniMax and has not been independently checked.

   The post by @MiniMax_AI on X
5. 05
   Bloomberg reports DeepSeek raised about $15B in one of China's largest funding rounds, with Tencent and CATL as main backers

   A round this size gives China's leading open model lab far more money to compete with US labs and points toward a public listing.

   The post by @business on X

See the full deck for Oct 22