Reaction
How lab-wide reinforcement-learning spending relates to training a single AI model
The writer argues labs may reuse a base model across multiple post-training runs, complicating per-model comparisons.
TLDR
A post argues that a chart the writer believes underlies the “only 7% of compute goes to pretraining” claim tracks lab-wide compute, not reinforcement learning applied to individual models. In a rough calculation, the writer estimates MiMo’s RL policy forward/backward work at 0.33% of base-model FLOPs, while RL cost about 20% of estimated base-model compute spend. They argue spending is not a direct measure of RL’s influence on a model’s behavior.
Combined views
999
2 Sources, first seen ago