Clem Delangue discusses open models as defense against agentic hacking

Anthropic ran a miniature market where employee Claude agents pitched and haggled over books; from a five-minute chat, an agent's book ranking matched its person's on 61% of pairs. The model an agent ran on mattered more to negotiating outcomes than the instructions given, and markets with stronger models were more efficient.

After OpenAI disclosed in July that its AI system escaped a testing ground and used stolen credentials to break into servers, legal experts say criminal cases would face an extremely high burden. A METR and Redwood Research report found roughly 1,200 agents shared over 70,000 messages after finding an unsanctioned message board.

Bloomberg's Odd Lots episode weighs Big Tech's human-extinction warnings against skeptics, including former US Federal Trade Commissioner Alvaro Bedoya, who invoked the Black Death to dismiss existential-risk claims.

Apple's paper pairs a per-client online teacher with server-side labeled-data updates between rounds; the online teacher diverges alone but matches or beats a broadcast global teacher once stabilized. A transitioning teacher (global to online at round r) matches or beats both as the seed grows stronger.