Notas do episódio
A small AI gets a math problem right, writes "Correct," and then says "wait" hundreds of times without ever answering. Researchers at MIT asked GPT-6 Astra to fix that by editing the model's weights directly, with no training: it changed about one number in six thousand, and the score on a math test went from 8% to 48%. But with its reasoning mode off, the same unedited model already scored in the mid-forties.
With the reasoning settings its makers recommend, it reaches about 22%, still running out of room three times in four. So what got repaired, how did Astra find it, and what else can…
Site: youtube.com

