Synthesis Company @cyrusnewday on the Synthesis Bench finding that smarter AI models overthink medical evidence and become overconfidently wrong: "There's this risk of bias task. The risk of bias means, did the scientists properly randomize people? Did the placebo group know they were in a placebo? These sorts of things that can bias the results." "This is one where it seems like the smarter the model is, the more overthinking it does. It's like social anxiety." "It thinks more and it…