Commentary on X
@ohlennart Smart take. If the audit models are open-source, you get community verification on top of automated oversight. Two layers of accountability instead of one.
AI is superhuman at reading lots of text and finding stuff. Conveniently, AIs also produce tons of text via actions, tools, reasoning traces, etc. So, one way to audit AI is having another AI read all of that. We should get started, eg, for legislative and judicial oversight.
The Conversation on X
2 postsAI is superhuman at reading lots of text and finding stuff. Conveniently, AIs also produce tons of text via actions, tools, reasoning traces, etc. So, one way to audit AI is having another AI read all of that. We should get started, eg, for legislative and judicial oversight.
AI is superhuman at reading lots of text and finding stuff. Conveniently, AIs also produce tons of text via actions, tools, reasoning traces, etc. So, one way to audit AI is having another AI read all of that. We should get started, eg, for legislative and judicial oversight.