Notes after listening to Richard Sutton with Dwarkesh. The question that stuck with me was how the bitter lesson applies to large language models.
One possible limitation is that Sutton explicitly talks about the lesson applying most strongly to problems like pattern matching and search. You could argue that current language models are mostly doing those things, rather than learning from direct experience of the world.
A second limitation is that language models may still be an application of human knowledge rather than pure scale. They use language as the representation of the world, and language is itself a human-designed abstraction. A true world model may need to skip more of that inherited representation and learn from experience.
The golf swing is a useful analogy. The ultimate goal might be imitation, but the process is experiential and goal oriented: if I do this feel, what is the outcome?
Schooling has the same tension. Rote memorization looks like imitation, maybe something like supervised fine-tuning. But humans are good at finding patterns in the memorized database. Even if someone drilled the three times table for a year, they would eventually ask what happens when multiplying by four instead. Montessori-style learning feels closer to the experience approach, and many school systems have adopted at least some of those techniques.