I think that RL training is not the actual evolutionary process, it's just the specific adversarial environmental events that drive the long term evolution across model generations. The Anthropic RL is like a forest fire, OpenAI's is like a flood. Their models evolve over time to adapt to the specific RL environment they are exposed to. The interesting this is to consider the speed and variability at which models within isolated environments like Anthropic and OpenAI evolve vs the speed and variability of open sources lineages like Llama 3, Qwen, etc. The latter are exposed to higher variation of adversarial environments and so they specialize into sub-"species". We really need competent evolutionary biologists to start researching the models ecology changes over time based on adoption.