The Bitter Lesson, ten years on
A 2026 re-read of Rich Sutton's famous essay, with a decade of scaling results to mark its homework.
Sutton's 2019 'Bitter Lesson' argued, provocatively, that the history of AI is a graveyard of clever hand-built methods, all eventually beaten by general approaches that simply leverage more computation. Contentious then; nearly prophetic now.
A graveyard of clever methods, all beaten by ones that just used more compute.
This re-read does the valuable thing: it tests the claim against everything since — the scaling of language models, the wins from sheer compute and data — and is honest about where the lesson is too blunt, where structure still earns its keep.
It's short, sharp, and a useful corrective to both camps: the 'just add compute' maximalists and the 'we need new ideas' romantics. The truth, as usual, is load-bearing in the footnotes.
The Bitter Lesson is the single most load-bearing idea in how the field spends its effort. A clear-eyed re-test of it is worth fifteen minutes — and it's just genuinely good thinking.