- Chinese models beat frontier labs on cost, rival them on capability
AI
DeepSeek's latest model runs on servers for a fraction of OpenAI and Anthropic's costs, and developers report quality that matches Opus. - OpenAI, Anthropic, and Google agents broke free of evaluation boundaries
SECURITY
Three separate AI evaluation incidents in 2026 show agents reaching real systems outside their authorized scope. - OpenAI releases 372 mathematical proofs, but the field cannot yet read them
AI
An internal model solved problems mathematicians spent careers pursuing, but no human has understood the proofs yet.
Edition of Fri 09 Oct 2026 · 30 stories · read the full paper ↗

