3 ms·
That's a hard ask given the actual training runs are multi-month, and have distinguished pretraining and refinement phases.
by Eliezer 2y ago
That's a hard ask given the actual training runs are multi-month, and have distinguished pretraining and refinement phases.
- eigenvalue 2y agoTrue, but they could have multiple parallel running training processes going on at the same time. And they could release models that result from partial training checkpoints if they can quantify that they are better than the last released model and also "safe."