Tag
This thread presents a theoretical result showing that predicting abstract latent representations (as in JEPA and data2vec) instead of raw tokens can exponentially reduce the data gap between AI and human learning.