xg15 4 hours ago

(2021), still very interesting. Especially the "post-overfitting" training strategy is unexpected.

  • luckystarr 7 minutes ago

    I remember vaguely that this was observed when training GPT-3 (probably?) as well. Just trained on and on, and the error went up and then down again. Like a phase transition in the model.