xg15 5 hours ago

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

  • luckystarr an hour 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.

esafak an hour ago

The low sample efficiency of RL is well explained.