My conjecture for why base LLMs become self aware is that there's slack in the teacher forcing of "predict the next token" type imitation objectives where there is only one way to get 100% accuracy…

Posters: jd_pressman. jd_pressman

8 likes recorded by the source corpus.