They simply do not believe that LLaMa 2 70B outperforms FLAC if you tokenize audio and stick it in there, implying the model learns the causal trace of every modality implied by text.…

Posters: jd_pressman. jd_pressman

5 likes recorded by the source corpus.