TB: The Bitter Lesson (Sutton)

Précis

Sutton's canonical 2019 essay: across 70 years of AI research, the methods that win in the long run are general methods that scale with computation — search and learning — and not methods that bake in human domain knowledge. The "bitter" part is that progress that scales with compute always eventually beats progress built on hand-engineered human cleverness, even though the human-cleverness approach is more satisfying and produces nicer short-term gains. Building in how we think we think does not work in the long run.

Key Takeaways

The structural argument

  • AI research has been done as if compute were constant; in fact compute keeps growing exponentially (Moore's law and its generalizations).
  • Over a slightly longer horizon than a typical research project, massively more compute becomes available.
  • Researchers, wanting wins on the shorter timescale, lean on domain knowledge — which is locally helpful, personally satisfying, and ultimately a plateau.
  • Search and learning are the two methods that scale with compute without limit. Long-run progress arrives by an opposing approach: scaling these two, not by adding more human knowledge.

The pattern, replayed across subfields

  • Chess (1997): Deep Blue beat Kasparov via massive search; chess researchers who had built knowledge-based systems were "not good losers" and dismissed brute force as not "how people play chess."
  • Go: Same pattern, delayed by twenty years; learning-by-self-play + search beat human-knowledge encodings.
  • Speech recognition (1970s DARPA competition): Statistical methods (hidden Markov models) beat phoneme/vocal-tract-knowledge systems. Deep learning extended the trajectory.
  • Computer vision: Edge-search, generalized cylinders, SIFT — all discarded. Modern deep networks use only convolution and a few invariances; perform much better.

John's annotation

  • "The only thing that matters is scaling datasets size and compute."
  • "Search and learning are two most important techniques for utilizing massive amounts of computation in AI research."

The deeper claim

  • The contents of minds are "tremendously, irredeemably complex." We should stop trying to find simple ways to think about space, objects, agents, or symmetries — these are part of the arbitrarily complex outside world.
  • Instead, build in only the meta-methods that can find and capture this complexity. "We want AI agents that can discover like we can, not which contain what we have discovered."
  • "Building in our discoveries only makes it harder to see how the discovering process can be done."

Notable Quotes

  • "Building in how we think does not work in the long run." (John's blockquote in Blogs.org)
  • "The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin."
  • "The eventual success is tinged with bitterness, and often incompletely digested, because it is success over a favored, human-centric approach."
  • "We want AI agents that can discover like we can, not which contain what we have discovered."

Why this matters / Connections

The Bitter Lesson is one of the most-cited short essays in modern AI. It is the framing under which scaling labs operate; understanding it is a prerequisite to having an informed opinion about AI strategy. It is also a general epistemological claim about premature abstraction — what looks like cleverness in one regime turns out to be a barrier under another. The argument also implies a temporal structure to disagreement: human-knowledge approaches always lose, but later than their detractors think.

See also

Source

The Bitter Lesson — Rich Sutton, IncompleteIdeas.net, 2019-03-13

Backlinks