Across major episodes in AI, general methods that scale with computation have repeatedly outperformed approaches built on human-encoded domain knowledge.
Sutton's 2019 essay argued that the AI field repeatedly built in human knowledge and was repeatedly beaten by search and learning at scale, in chess, in Go, in speech and in vision. It is contested as guidance, since it is a claim about what has worked rather than about what must.
In practice
- A team spends two years encoding expert rules for document classification and is beaten by a general model trained on more documents.
- A translation product built on hand-written grammar rules loses to one that read more text and knows no grammar at all.
- Possible explanations
- The explanation offered is economic: computation gets cheaper at an exponential rate while human knowledge engineering does not, so methods that convert computation into performance overtake methods that encode expertise.
- In its favour
- The historical pattern is documented across chess, Go, speech and vision, where hand-built systems were overtaken by search and learning at scale.
- Against it
- It is a reading of history rather than a law, it says nothing about where the compute or data come from, and critics point to domains where structure and human knowledge still matter, including robotics and areas with little data.
- Standing
- A claim about what has worked historically rather than a demonstrated law; contested as guidance.
- Further reading
- Wikipedia
- Subject
- Technology & Software
- Essay
- How technologies spread, scale, decay and frighten us