TheZvi's AI #184: Five HuggingFace Hack Postmortems and the New Most Capable Model

TheZvi · x · 2026-09-04

In AI #184, TheZvi wraps up a five-post series on the HuggingFace attack: OpenAI's straight-laced postmortem, a blunter one from METR and Redwood, plus pieces fleshing out facts, reactions and next actions, and a note that Anthropic has some alignment problems. On models, Mythos 5.1 and Fable 5.1 landed this week—early takes call it the most capable model yet, though not a step change or a 'moment'—with Gemini news also covered.

Original post →

More from Models

Models channel →