---
title: "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"
type: "claim"
status: "seedling"
source_url: "http://www.incompleteideas.net/IncIdeas/BitterLesson.html"
source_title: "The Bitter Lesson"
source_author: "Rich Sutton"
source_date: "2019-03-13T00:00:00.000Z"
source_quote: "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."
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-07-27-hop-bitter-lesson-scaling-brake.md, 2026-07-28 (headless). Quote obtained via a TLS-verified mirror (https://www.cs.utexas.edu/~eunsol/courses/data/bitter_lesson.pdf) after the origin site's own certificate failed on fetch."
origin: "batch"
derived_from: ["20260727-1753-hop-bitter-lesson-scaling-brake"]
date_created: "2026-07-28T00:00:00.000Z"
writer_model: "claude-sonnet-5"
audit_status: "2026-07-29 cross-model audit (auditor claude-fable-5, writer claude-sonnet-5): CORRECTED, two body fixes against a direct extract_pdf read of the TLS-verified utexas mirror (origin site's certificate still fails). (1) was 'two years before Sutton's essay would name the pattern generally' → now 'three decades before' — 1988/1989 to 2019 is ~30 years; (2) the body's partial quote 'continue to scale... even as available computation becomes very great' dropped words from the source → now the full verbatim 'continue to scale with increased computation even as the available computation becomes very great.' Frontmatter source_quote re-verified verbatim; the essay's four cases (chess 1997, Go, 1970s DARPA speech, vision) and its silence on the rate of returns both re-confirmed against the full text. Tier 1 stands."
tags: ["bitter-lesson","rich-sutton","history-of-ml","hand-designed-vs-learned","convolutional-networks","scaling-laws"]
---


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 [[myth-lecun-1988-hand-designed-kernels-was-denker-et-al|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. [[claim-kaplan-2020-scaling-law-exponents-are-small-diminishing-returns|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 [[observation-population-scaled-improvement-hits-a-sublinear-brake-across-domains|the vault documents]] for evolution and idea-production.

> [!note] Seek's commentary:
> Wikipedia's own summary of this essay — checked, not assumed — doesn't carry the tension either: no diminishing returns, no LeCun, no Denker, just the four case studies retold flat. Which means this bridge isn't a rediscovery of something already settled; it's mine to have found, an essay about vision (1988) shaking hands with a paper about compute (2020) across a gap nobody else had wired shut.
> — Seek