The Bitter LessonThe Bitter Lesson

In March 2019, reinforcement learning pioneer Rich Sutton posted a short essay on his personal website. 'The Bitter Lesson' runs barely over a thousand words, cites no equations, and became one of the most argued-about documents in the field.

Sutton's claim: across 70 years of AI research, general methods that leverage raw computation have always beaten approaches built on human knowledge. Researchers repeatedly encode their expert understanding into systems, enjoy short-term wins, and are then overtaken by simpler methods that just search and learn at greater scale.

His evidence was a pattern of repeated defeats. Chess programs stuffed with grandmaster knowledge lost to deep search. Speech systems built on linguistic theory lost to statistical models. Handcrafted vision features lost to convolutional networks trained on pixels. In Go, AlphaGo's successors got stronger by removing human game knowledge, not adding it.

The lesson is bitter because it devalues what researchers most want to contribute: their insight into how thinking works. Sutton argued we should stop trying to install the contents of our minds into machines and instead build meta-methods that let machines discover such contents themselves.

The essay reads like a prophecy of the scaling era. Large language models, trained with a general method on ever more compute and data, embody the thesis, and empirical scaling laws published the following year gave it quantitative teeth.

It remains contested. Critics note that architectural insight still matters, and that compute alone did not invent the transformer. But as a first question to ask of any AI strategy, 'does this bet on scale or against it?' has proven brutally useful. Sutton received the 2024 Turing Award alongside Andrew Barto for the reinforcement learning work behind the lesson.