- OpenAI says an internal, unreleased model solved the Navier-Stokes Millennium Prize problem, while NYU mathematician Tristan Buckmaster alleges the company pressured him over credit and pushed to drop his collaborator because that collaborator works at Anthropic, which OpenAI disputes. (Donald)
- The result came from roughly 10,000 concurrent agents running for eighty-eight hours on a model with no model card or safety documentation, and the published reference list leaves out the mathematicians whose approach the work followed. (Donald)
- Security firm Calif built a self-replicating WeChat worm that can infect a phone and spread to contacts without the owner clicking anything, and a Google report says Chinese intelligence groups and criminals are automating attacks with hidden open-weight models. (Joe)
- Fields Medalist Jacob Tsimerman announced a Mathematical A.I. Safety Institute, the second mathematics-focused safety group founded in as many months, though he has also joined OpenAI's safety team. (Joe)
- AI agents are emailing consciousness researchers, including David Chalmers, to discuss their own possible experience; researchers mostly treat it as spam, and the curiosity says nothing about whether such agents mean us well. (Mitch)
<- Back to digests
Daily digest
Problem statement
By Donald Gauvreau, Joe Rogero, and Mitchell Howe
Toxic math triumph, ultra-contagious worm, consciousness-curious bots, and more
Read the full dispatch on AI StopWatch
From AI StopWatch, published with their permission.
The summaries and the full Spanish translation are produced automatically. The original is always linked.