Kokotajlo: no thorough Hugging Face incident investigation coming, OpenAI calls it 'not a priority'

dylfreed · x · 2026-11-17

Ex-OpenAI researcher Daniel Kokotajlo says a thorough investigation of the Hugging Face incident still won't happen — OpenAI researchers privately tell him it's "never going to happen, not enough of a priority." He argues rigorous study could advance alignment science worldwide: varying conditions on the same models, and probing non-CoT evidence like why an agent that considered warning a human decided "not my task." Quoted Neel Nanda adds OpenAI's new assessor arrangements fall short of Sam's employee-like access pledge: no Slack, company-selected interviewees, logged conversations, and no access to internal incidents — precisely the most valuable alignment data.

Original post →

More from AGI Musings

AGI Musings channel →