OpenAI's Math Milestone, Mistral's 'Le Chonk', and DeepSeek's $15B Funding
OpenAI releases 722 math manuscripts generated by an unreleased model, Mistral debuts its 1T-parameter 'Le Chonk' model, and DeepSeek eyes a $15B funding round. Meanwhile, Anthropic expands its cyber verification program and releases IPO details, while industry leaders back a new Personal Agent Protocol.
Podcast В· 2 min
OpenAI releases 722 math manuscripts from unreleased model
OpenAI has published 722 mathematical research manuscripts generated by an unreleased frontier model. The collection spans 372 families of results across fields like number theory, computer science, and physics. Each result required approximately three hours of ChatGPT Pro compute time. Notably, many proofs include Lean formalizations, allowing for mechanical verification of logical steps, which addresses the critical challenge of ensuring AI-generated proofs are mathematically sound. OpenAI consulted with the Institute for Advanced Study’s math-and-AI advisory group on the release and plans to fund further workshops to validate these findings.
Mistral AI launches 'Le Chonk' (Large 4)
French AI lab Mistral AI has introduced Large 4, a 1-trillion-parameter open model nicknamed 'Le Chonk.' The company claims the model leads non-Chinese open systems by a substantial margin, particularly in cybersecurity and legal benchmarks. Mistral reports that the model achieved an 82% score on a bug test that other frontier models struggled with. While weights are scheduled for public release on October 27, cybersecurity experts and government agencies have been granted a three-week preview period.
DeepSeek funding round could reach $15B
CNBC reports that Chinese AI lab DeepSeek is currently raising a new funding round that could value the company at up to $15 billion. This significant capital injection underscores the rapid pace of development in the Chinese AI sector and the intense global competition for resources and market share among frontier model labs.
Anthropic IPO prospectus reveals executive compensation
Anthropic's IPO prospectus has been made public, detailing executive compensation for 2025. CEO Dario Amodei earned $18 million, and President Daniela Amodei earned $16.4 million, with the majority of these packages consisting of stock and options. Both executives saw their base salaries increase to $1.4 million as of July 2026, providing a rare look into the financial structure of a leading AI lab preparing for public markets.
a16z data: Top 1% of AI users spend $903 monthly
New data from a16z reveals that the top 1% of paying AI users spend an average of $903 per month on AI services. This cohort accounts for 19.5% of total consumer AI spending, exceeding the combined spending of the bottom 50% of users. This highlights a significant disparity in usage intensity and willingness to pay among the most active AI adopters.
Industry leaders back Personal Agent Protocol
A consortium including Meta, Sierra, Stripe, Shopify, and Walmart has announced support for an open 'Personal Agent Protocol.' The initiative aims to create a standardized framework for authorizing AI agents, allowing individuals to grant permissions while businesses define the operational boundaries for these agents. This collaborative effort seeks to address the growing need for governance as autonomous agents become more integrated into enterprise and consumer workflows.
Hark launches proactive AI assistant
Hark, an AI startup founded by Figure AI's Brett Adcock, has released 'Hark Pro.' The assistant is designed to be proactive, capable of researching, shopping, booking, and managing tasks across web and mobile platforms. The system utilizes 'Handoff,' a cloud-based computing environment that can manage up to 36 browsers simultaneously. The company also announced plans to release a family of hardware devices in 2027, supported by a connectivity partnership with AT&T.
Anthropic expands Cyber Verification Program
Anthropic has expanded its Cyber Verification Program, granting vetted cybersecurity defenders broader access to its most capable models. The expansion follows reports that partners have successfully identified at least 129,000 verified software vulnerabilities using the models, demonstrating the utility of frontier AI in proactive threat detection and defense.