“Anthropic CEO Dario Amodei has stated that comparatively small amounts of compute have been applied to Reinforcement Learning (RL) scaling compared to pre-training, implying significant potential for gains.”