Skip to main content
Zubnet AI › Learn › Wiki › The Bitter Lesson
Fundamentals

The Bitter Lesson

Also known as: Bitter Lesson, Sutton's Bitter Lesson
Rich Sutton's 2019 argument that general methods which scale with computation eventually beat methods built on human expertise.

Why it matters

The essay is probably the most cited piece of informal writing in modern AI, and it shaped how the big labs think. The bet behind large language models is the bitter lesson in action: take a simple, general architecture, give it enormous data and compute, and don't hand-code linguistic rules. When people say an idea is or isn't "bitter-lesson-pilled", they're asking whether it will be swept away by the next round of scale.

For builders, it carries a practical warning: clever workarounds for a model's current weaknesses often become obsolete when the next model arrives.

Deep Dive

"The Bitter Lesson" is a short essay published in March 2019 by Rich Sutton, a founder of reinforcement learning and later a Turing Award winner. Looking back over 70 years of AI research, he saw the same story repeat. Researchers try to build their own understanding of a problem into a system: chess strategy, the rules of grammar, how to detect edges in an image. That helps at first. Then a simpler, more general method that just uses more computation, through search and learning, overtakes it and keeps improving.

The lesson is bitter because it means the knowledge researchers are proudest of tends not to matter in the long run.

Sutton's Examples

  • Chess: Deep Blue beat Garry Kasparov in 1997 mainly through massive search, to the dismay of researchers who had tried to encode human chess understanding.
  • Go: hand-crafted approaches stalled for decades. AlphaGo combined search with learning, and AlphaZero then dropped human game data entirely and got stronger.
  • Speech recognition: systems built on linguists' knowledge of phonemes lost to statistical methods, and then to deep learning.
  • Computer vision: hand-designed feature detectors were replaced by convolutional networks that learned their own features.

What It Doesn't Say

The essay is often flattened into "just add compute". Sutton's actual claim is narrower: favor methods that can use more computation as it becomes available. Structure still matters. The transformer is itself a human design, but a general one that scales well. Data isn't free either: learning needs something to learn from, and for many problems collecting the right data is the hard part.

Applying It to Building with LLMs

Elaborate prompt chains, hand-written parsers to fix model output, and rigid multi-step pipelines often get replaced by a single call to a better model. Many teams building agents report the same pattern: each model generation lets them delete scaffolding. A useful habit is to ask of any workaround, "will a stronger model make this unnecessary?" If so, keep it cheap and easy to remove, and invest in what lasts: your data, your evaluations and your tools.

Criticism and Sutton's Own Twist

Critics note that compute is costly, that efficiency and good priors matter when data is scarce (as in robotics or medicine), and that the largest gains have often come from combining scale with new ideas. Sutton himself has argued that today's LLMs don't fully follow his lesson, because they learn from human-written text rather than from their own experience of the world. In his view, the truly scalable approach is still systems that learn continually by acting and observing.

← All Terms
ESC