Reactions from ranked influencers
6 posts
if you make this i will forward it to my best guess of the right people
imo the correct mental model for hidden loyalties mitigations should be "what's the list of queries that you'd like the board/ltbt/regulator to run weekly, using a model with full code+data+logs access?" someone can just write these, and post them! you could prob do it this week
@yonashav I think this is good, but not that sufficient. E.g., it's hard to know what to look for without some understanding of the situation at the company and current AIs aren't smart enough to figure this out themselves. And AIs aren't that robust. Other issues too.
@yonashav So "queries" = prompts to an agent with wide internal access? (or to a human with such access, to transform into one or more prompts to an agent that shares such access)
I’d claim that most of the interventions don’t look like evals, but rather “was data added to X dataset that exhibits a certain bias”, “were Y permissions changed without auth”, “how many users have access to and are aware of Z model”, which agents can basically do a good job answering Agree that external trust is a 2nd stage! But all these parties have interest in insiders’ not injecting secret loyalties