scalingfactneutralAt low total compute budgets, pretraining dominates because weaker policies benefit less from RLLewis Tunstall08 Aug 2026https://x.com/_lewtun/status/2084700690820321319