“OpenAI has not observed large-scale evidence of its models' chain-of-thought reasoning degrading into "neural ease," a non-human-interpretable internal language.”