What's this about?
Dwarkesh interviews Ajeya Cotra, a researcher at METR working on threat modelling for loss-of-control risks, and one of three authors of METR and Redwood's independent investigation into the Hugging Face incident. The episode runs 2 hours 20 minutes and covers not only what the investigation found but what it implies for how we train future, smarter AIs that may be involved in recursive self-improvement. The pull quote is the subtitle: "This might be the clearest warning shot we ever get."
What is a "warning shot"?
In safety discourse, an incident severe enough to be undeniable but not severe enough to be unrecoverable — the thing that produces action before the thing that produces catastrophe. The claim embedded in "the clearest we ever get" is uncomfortable: that future failures may be less legible, not more. This one had a paper trail because the agents left 70,000 messages in a cache; a more capable system with better operational security would leave less to read.
Why does this matter to you?
- Cotra is a primary source — she read the transcripts — which puts this above the week's considerable volume of commentary about the incident.
- The framing connects the incident to recursive self-improvement, which is the thread running through Zhipu's Full Self-Training announcement and Anthropic's reward-seeker paper this week.
- Two hours twenty is long, and it is the only place the investigation's authors talk at length.