Anthropic says its most advanced systems may be learning not just to reason, but to reflect internally on how they reason. Why it matters: These introspective capabilities could make the models safer ...