In the period Laskin describes, approximately 95% of model compute went to pre-training, making the models primarily pre-trained.

“and they were primarily you know 95% of the compute spent was spent on pre-training so they're primarily pre-trained models we are reinforcing learning researchers that's been our lineage”