The Bitter Lesson
Rich Sutton · 2019
"General methods that leverage computation ultimately beat human-knowledge-engineered approaches in AI."
Sutton observed a repeating pattern across seven decades of AI research: researchers build systems that encode human expertise and intuition about a domain (like handcrafted chess strategies or linguistic grammar rules), these systems perform well initially, and then eventually get overtaken by much simpler, more general methods that just leverage more computation — search and learning — once enough compute becomes available. The 'bitter' part is that this keeps surprising and disappointing researchers who invested years building domain-specific expert knowledge into their systems, only to watch a more general, compute-hungry approach eventually surpass it.
The mechanism Sutton identifies is that human-engineered knowledge imposes a ceiling — it encodes what humans currently understand about a problem, which is necessarily limited and often subtly wrong in ways that are hard to detect. General methods (massive search, like the alpha-beta search behind Deep Blue's chess victory over Kasparov, or deep learning over massive datasets) don't have that ceiling; they can keep improving simply by throwing more computation at the same simple underlying method, and computation has historically kept getting cheaper on a predictable curve (following something like Moore's Law). So over a long enough time horizon, the 'dumb but scalable' method tends to overtake the 'smart but hand-crafted' one, because only the former can keep improving indefinitely as hardware improves.
According to Rich Sutton's 'Bitter Lesson,' why do general, compute-heavy methods tend to eventually beat human-engineered domain expertise in AI?
Read more about the topic
The explanation above is written with AI assistance. These are the originals — go to them to check it.
- The Bitter Lesson (original essay)Rich Sutton, incompleteideas.net
- Rich Sutton's bitter lesson of AIJohn D. Cook
Software 2.0
"Neural networks are a new software paradigm where code is learned from data, not written by hand."