Some thoughts on the Sutton interview

This is a reflection piece written after the audience reaction to Patel's Richard Sutton interview, and its structure matters. He opens by acknowledging how much reaction the episode produced, says his understanding of Sutton's position is now much better than it was during the interview itself, and apologizes in advance to Sutton for errors or misunderstandings. That framing sets the tone for everything after: this is not a defense of LLMs against a critic, it is an attempt to state the critique at full strength and then locate precisely where the disagreement actually lives.

The concession comes first and it is substantial. Patel restates the steelman of Sutton's position as a chain of five claims: the bitter lesson is about techniques that leverage compute effectively and scalably, not about spending compute indiscriminately; most compute spent on an LLM goes to deployment, where it learns nothing; training itself is wildly inefficient, consuming the equivalent of tens of thousands of years of human experience; everything learned in training comes from human data, including the RLVR environments which are human-furnished playgrounds; and the resulting agent never learns through organic, self-directed engagement with the world.