Sutton's 'The Bitter Lesson' (2019) names hand-designed-loses-to-learned-at-scale as a 70-year AI pattern but never addresses the rate of returns to scaling

Rich Sutton's essay "The Bitter Lesson" (2019) argues that across roughly 70 years of AI research, general methods that leverage increasing computation have repeatedly beaten approaches built on hand-engineered human knowledge — citing chess (1997), Go, 1970s DARPA speech recognition, and computer vision as the four cases. The vision case is stated directly: "Early methods conceived of vision as searching for edges, or generalized cylinders, or in terms of SIFT features. But today all this is discarded. Modern deep-learning neural networks use only the notions of convolution and certain kinds of invariances, and perform much better." That transition is the same one the vault's myth ledger documents at Bell Labs — hand-designed feature detectors in 1988, learned convolutional kernels by 1989 — three decades before Sutton's essay would name the pattern generally.

What the essay does not do is quantify the win. It states that the great power of general-purpose methods lies in their ability to "continue to scale with increased computation even as the available computation becomes very great," but nowhere addresses how fast returns accrue. Kaplan et al.'s (2020) neural-scaling-law exponents supply that missing rate, and it is small — the arbitrary scaling Sutton celebrates turns out to be governed by the same sub-linear brake the vault documents for evolution and idea-production.