The Bitterest Lesson: picking the right task beats data, compute, and algorithms
ZeYanjie · x · 2026-09-20
A TypeSafe AI blog post extends Rich Sutton's Bitter Lesson: compute beating algorithms is just the tip of the iceberg. The real hierarchy is:
Doing the right task > data > compute > algorithms
- Sutton's lesson is clearest in games, where the objective is obvious and data can be endlessly self-played — so compute naturally comes next.
- In the real world, ML ultimately means making reward go up or loss go down, and someone still has to choose the objective. Pick the wrong task and the model can be useless despite beautiful loss curves and perfect scaling.
- ML research tends to attack these in reverse order: researchers love algorithms, recently scaling curves, while data is messy and task selection requires leaving ML entirely to study users, products, and organizations.
Core claim: the bitterest lesson is verifying you're solving the right problem before scaling anything.
Related event: The Bitterest Lesson: Picking the Right Task Beats Data and Compute(2 posts)→
More from AGI Musings
- a16z Partner: World Is Massively Underestimating 'a Country of Geniuses in a Datacenter' — saranormous · 2026-09-20
- e/acc founder mocks EA claim that insects matter more than humans in expectation — beffjezos · 2026-09-20
- Researchers clash: PhD problems becoming today's AI benchmarks shows progress, not an ending — GeorgiaChal · 2026-09-20
- Vision researcher pushes back on 'field is solved' claim: robotics is not simply the next modality — GeorgiaChal · 2026-09-20
- Charlie Warzel urges a middle path between AI doomers and skeptics — round · 2026-09-20
- Muse's Spotify connector seen as start of consumer actions moving into LLM chats — anikembhavi · 2026-09-20