The current hype phase of artificial intelligence makes it easy to find confirmation for your preconceptions. It’s easy to find examples that make it look extremely capable. It’s just as easy to find examples where it looks embarrassingly clueless. Where it gets really interesting is in between: the techniques humans can use to elicit more accurate answers from the AI when it fails on its own (for example, by telling it to «think step by step» or, funny enough, «don’t make stuff up»).

From Weekly Filet #426, in January 2023.