“‘we have input and output classifiers now… moving more towards having these sorts of contextual classifiers… gives the classifier a much better sense’”
Alan Chan, GovAI · "The Cognitive Revolution" · Governing AI Agents: Alan Chan of GovAI on Monitoring, Agent IDs, Rogue Agents & Preparing for RSI
Play from 7:00
7:00
Audio streams from the publisher’s own feed. Verbatim does not host or copy it.
What was described
recommended using contextual classifiers that evaluate the entire conversation history rather than single input/output pairs
Said nearby on the same subject
“I think we've seen a pattern of employees being evaluated on AI usage… tracking usage of tools is the absolute wrong way of going about evaluating output.”
“I’d go through every fix a person made to an agent’s work… propose a rule… add each approved rule as a test using the original input and the accepted output.”
“We rather think it's useful to instead use the direct carbon prices that sectors are subject to rather than just using a blanket shallow carbon price.”