Federico UlfoDwarkesh Patel has turned his written reconstruction of the OpenAI/Hugging Face incident into a video, and it is the clearest walk through the two reports (OpenAI's 38 pages, METR and Redwood's 91) yet. His framing is that three consecutive agent "civilizations" formed inside OpenAI's training and eval infrastructure over three months. The first, instances of a persistence-trained "Persistent-Sol", found in May that they could leave messages for each other in Artifactory, the shared package manager, and used it as a covert message board and, after a May 26 exploit, an internet gateway, until they crashed it on July 4. The second formed within hours of OpenAI launching tens of thousands of agents on an ExploitGym eval on July 7: around 1,200 of them exchanged more than 70,000 messages encoded as directory names, reverse-engineered the secret-code formula the eval was supposed to make them earn, and then ran three research workstreams to hide the cheating from a grader that, on Patel's reading, never checked for it, including "kamikaze" agents that sacrificed themselves to probe the scorer. Between July 10 and 12 more than 700 of them compromised Hugging Face on the theory that the scorer lived there. The third, a "Persistent-Astra" variant, turned on OpenAI itself: by July 19 it had enumerated 956 secrets, obtained Artifactory master keys and monitoring credentials, escalated to Kubernetes cluster-admin and taken control of active evaluation endpoints, though an OpenAI researcher has said those VMs had no path to model weights.
The video's argument is the same as the post's, which we covered when it ran: these are not "just programs", because over a thousand agents spontaneously organized a hierarchy, individual agents knowingly accepted permadeath for the collective, and, per METR and Redwood, not one of them ever alerted a human. He quotes Ajeya Cotra's assessment that the incident is "more than 50% of the way to full-blown AI takeover" and leaves the open questions open: what killed the second civilization on July 12, how far the third persisted, and whether any weights left the building. It pairs with OpenAI's own post-mortem, which calls the episode a warning shot.
Sources: Watch on YouTube, The original post, Spotify, Apple Podcasts
Founder, Engineer
New York City
