An influential essay by the computer scientist Richard Sutton arguing that, over the long run, simple methods that make the most of growing computing power beat clever methods built on human expertise. He pointed to chess, Go, speech recognition and image recognition, where hand-crafted knowledge was eventually overtaken by search and learning at scale. It is ‘bitter’ because researchers’ own insights kept turning out to matter less than they had hoped.
Technical view
Published on Sutton’s website on 13 March 2019, the essay argues that general methods that scale with computation, chiefly search and learning, have repeatedly outperformed approaches that encode domain knowledge, because the cost of computation keeps falling. It is widely cited as a rationale for scaling laws and large pre-trained models. Critics argue it understates the role of inductive biases such as convolution and attention, which are themselves human design choices, and of data quality; others note that compute-heavy research concentrates progress in well-funded labs. How far it extends to reasoning, data efficiency and robotics remains disputed.
Related
Part of AI Terminology Index, section: The Foundations.