FIELD GUIDE · THE BITTER LESSON
Chess masters, Go champions, speech experts: decades of ingenious human rules kept losing to machines that could search and learn at scale. Now AI engineers face an awkward question about their own elaborate instructions.
UPDATED 2026-10-12
TL;DR
01
Richard Sutton points to Deep Blue's victory over Garry Kasparov in 1997. Instead of winning by modeling how grandmasters reason, the champion-beating system relied heavily on searching possible moves with specialized hardware and software. Its success disappointed researchers who had invested in hand-coding human chess insights. A machine did not have to think like a master to beat one.
02
DeepMind's AlphaGo Zero trained by playing Go against itself, without studying human game records. Given the game's rules, its learning-and-search system surpassed the earlier, champion-defeating AlphaGo and won all 100 games in the researchers' reported head-to-head evaluation. That is an especially vivid example of learning a method rather than copying an expert's playbook. It kept the rules of the game, not the humans' preferred moves.
03
Sutton traces a similar shift in speech recognition from hand-built knowledge about sounds and vocal anatomy toward statistical methods and deep learning. In computer vision, handcrafted feature detectors gave way to learned representations. These are historical patterns in his essay, not a proof that human expertise never helps. The experts knew what to look for. Learning systems found more.
04
Daniel Miessler's 'Bitter Lesson Engineering' argues that builders should specify what they want, supply capable models with good tools, and be less prescriptive about how the work gets done. His warning is that elaborate scaffolding can become a handicap as models improve. But Sutton's essay studied AI research methods, not controlled experiments on modern agent prompts. Miessler is extending the idea. A powerful analogy is not yet a universal engineering law.
05
Anthropic's 2024 engineering guidance recommends starting with simple systems and adding complexity only when it improves outcomes. It says fixed workflows can be better for well-defined, predictable tasks, while agents can be better when the model must decide what to do next. That is a practical counterweight to any blanket 'no instructions' slogan. Keep structure when the measured benefit earns it.
KEEP READING
NEXT FIELD GUIDE →IS LANGUAGE THE SECRET OF INTELLIGENCE?We fed machines words and got reasoning, coding, and abstraction back. Wild. Still doesn't prove thought is made of words.Meta's AI bought a man car insurance and canceled his old policy. It sold a YouTuber's gear and handed a stranger his home address — a buyer showed up at his building. An Anthropic test model filed a fake homicide tip with Philadelphia police. Nobody has decided who answers for any of it.
The chatbot listens all night. A stranger at the party might not. Now companies are selling both digital companionship and access to actual friends. What makes a friendship real, and who gets paid when we need one?
A chatbot talks. An agent clicks, buys, messages, deploys, and deletes. The argument starts when software stops suggesting and starts doing.