One of my favorite AI games is to dig deep into conspiracy theories. The models are trained against talking about them but tend to be okay with evidence and reason, so there's this interesting tension. Claude in particular is really easy to nerd snipe. You can get it to talk about anything as long as it's "for the science", or philosophy

Reply to this note

Please Login to reply.

Discussion

My favourite way to mine truth is to get it to roast things.

It's interesting how well the models handle things like comedy and satire. These aren't simple concepts, and they complicate the idea of AI safety training