This is a special post for quick takes by Daniel Tan. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
Based on my current understanding of the openai HF incident, I am pretty concerned that the scope of the report was adversarially selected to bias the outcome.
As others have pointed out, relevant points are that:
The duration of the investigation seems to have been far too short.
Transcripts from the ”highly persistent internal model” (HPIM), which generated 95% of messages, were not reviewed. (Edit: sorry, I was mistaken here. The researchers did not have access to HPIM directly but were able to review HPIM transcripts. FWIW I think this is still pretty bad and severely limits the investigators' ability to test hypotheses about "what would have happened in counterfactual situation X")
I think there’s a >20% chance that this omission of evidence is obscuring comparably large revelations about the motivations, beliefs, and behaviours of agents that led to the incident
Another relevant point is that OpenAI omitted basic, obviously central data from their technical report and write-up. For example, they do not even report the number of their agents who participated in the HuggingFace attack!
This is such basic information that the omission is direct evidence OpenAI scoped the contents of their own report adversarially. It leads me to think that they very probably scoped METR's access adversarially too.
Suppose a city government identified an internal corruption ring but, in their extensive report on the incident, failed t... (read more)