s1r1us Demands Proof on CoT Monitoring Claims
Founder argues claimants must demonstrate undetectable model exploits from tool calls alone.
TLDR
s1r1us posted that the burden of proof lies with anyone asserting the lack of CoT makes monitoring harder. The statement calls for a concrete case of a model exploiting a vulnerability through sci-fi hacking techniques invisible from its tool calls. Florian Brand, a research engineer at Prime Intellect, retweeted the post. The comment appears amid visible replies on X discussing AI evaluations and monitoring.
Combined views
5.4K
2 Sources, first seen 27d ago