Notes and Thoughts

Discursive working and evergreen notes


The bitter lesson

  • Author:: Rich Sutton
  • Source::

The biggest lesson from AI research is that general methods that leverage computation are the most effective by a large margin. This is because of Moore's law. The background assumption of a lot of AI research was that computation is constant. This led researchers to take actions that are effective in the short term - such as fine tuned, complex models - but that will fall short as computation increases. In the long run this trend will dominate, and simple methods that can use large amounts of compute will work better.

Fine tuned vs general methods are not necessarily opposed, but in practice if you spend time doing one you won't do the other. This same pattern played out in chess, in Go, in computer vision, in speech recognition, etc. Researchers tried to make the systems work the way they thought human minds worked, or by leveraging knowledge of the structure of the problem domain, but in the end massive computation was better. AI Researchers have, frequently, tried to build specific knowledge into their agents, which helps in the short term and is satisfying, but in the long run it plateaus and inhibits further progress. Breakthrough progress has come from scaling learning and search, both of which (appear) to scale to arbitrarily high levels.

Minds are very complex and we should stop trying to think that we can build in the important parts; instead we should build the ability to discover.