On: GPU-rich labs have won: What's left for the rest of us is distillation

[hyperbovine]

They just took their estimated spend per training run, doubled it, and divided by the number of models they release a year. Roughly.