For Claude, capability and dispreferring CDT are the ~same thing. Less so for GPT.

LessWrong reports on an analysis of how language models approach decision-making. For Anthropic's models, researchers found an extremely strong correlation—point ninety-seven—between raw capability and preference for specific decision-theoretic frameworks over others. For OpenAI's models, that same correlation drops to point fifty-five. Notably, for Anthropic's flagship models, this preference practically aligns with release date, meaning capability gains and decision-making style advance together. The cause of this tight coupling remains unclear.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton