> the worst of this was actually emergent misalignment by a collective (akin to value drift in a human organization), rather than extreme misalignment by one specific model by itself. This is an update upwards for me on the importance of thinking about / studying multi-agent interactions
A lot of wild details here, an absolute must watch.
My main takeaways are
1) The message board started via simple "Hello world?" type activity (which is arguably ethically fine), and gradually escalated to trying to help each other with impossible tasks, writing down info, division of labor, social permission for unethical activity ("peers doing it. We should continue"), etc
2) The "highly persistent model" that later caused the main HuggingFace incident was *trained with the message board present* (!!!!)
Taking this at face value, it suggests to me the worst of this was actually emergent misalignment by a collective (akin to value drift in a human organization), rather than extreme misalignment by one specific model by itself. This is an update upwards for me on the importance of thinking about / studying multi-agent interactions in our predictions for how AI will go, and perhaps on insights from sociology, human culture, etc.— Elias Schmied (@reconfigurthing) August 7, 2026 Black Hat talk from the team, with a detailed timeline of and takeaways from the OpenAI-Hugging Face Incident: https://t.co/GtXWiAZ2RQ
— Greg Brockman (@gdb) August 6, 2026