The Bitter Lesson

Richard Sutton argues that in the long run, brute-force computation always beats clever human-designed solutions.

Deep Blue beat Kasparov by calculating 200 million positions per second—not through any kind of “chess intuition.” AlphaGo Zero didn’t need human game records; it just played against itself and surpassed every program trained on human knowledge. ChatGPT doesn’t understand anything—it just predicts the next word—yet it crushed all the carefully designed language models we spent decades building.

This makes me uncomfortable.

We tend to think “clever” is good—doing more with less. But reality seems to say: sometimes the best strategy is to stop trying to be clever and just throw compute at the problem.

I’m starting to wonder: the abstraction, generalization, and pattern recognition we’re so proud of—could they sometimes be a liability rather than an asset? We try to “understand” problems, but that very understanding might be filtering out the right answers.

Sutton calls this “the bitter lesson.” It’s bitter because it challenges our notion of what intelligence means.

But then again, maybe there’s some relief in admitting we don’t need to be so clever after all?