AI/ML Wiki

The Bitter Lesson

3 min read

The Bitter Lesson

Authors: Richard S. Sutton | Year: 2019

Seventy years of AI research deliver one consistent lesson: general methods that scale with computation always win over clever domain-specific engineering, no matter how satisfying the latter feels in the short term.

Core Ideas

Sutton surveys major AI milestones — chess (Deep Blue → AlphaZero), Go (hand-coded knowledge → AlphaGo/AlphaZero), speech recognition (phoneme models → end-to-end neural nets), NLP (grammar-based parsing → statistical → neural language models) — and identifies a recurring pattern. In each domain, researchers first try to encode human expert knowledge directly into the system. These knowledge-engineering approaches make rapid early progress and feel intellectually satisfying. But they plateau. Eventually, a simpler approach — scale up computation with a general learning or search method — surpasses them decisively.

The lesson is “bitter” because it is unwelcome to researchers who have invested their careers in understanding the domain and encoding that understanding. The bitter lesson is that their knowledge, while real, is not what matters in the long run. What matters is whether the method can continue to improve as compute increases.

Sutton identifies search and learning as the two general methods that have consistently scaled. Search (Monte Carlo methods, tree search, beam search) is the systematic exploration of possibility spaces. Learning (gradient descent, temporal difference learning, statistical methods) is the extraction of regularive structure from experience. Both are substrate-independent: they work on any domain where there is an objective and sufficient compute.

The essay ends with a prescriptive conclusion: researchers should stop trying to build human knowledge into AI systems and should instead invest in scalable representations and architectures that allow search and learning to do their work. The two meta-level capacities that matter are the ability to discover arbitrary structure and the ability to apply it flexibly — not the pre-specification of that structure by humans.

Significance

“The Bitter Lesson” became one of the most widely cited essays in AI, particularly after the success of large language models confirmed its thesis in NLP. It provided the conceptual framing for what Kaplan et al. (2020) would later quantify as scaling laws: performance improves predictably with compute, data, and model size along a power law, regardless of architectural cleverness.

The essay is controversial because its prescriptive conclusion — invest only in general methods — is arguably an overcorrection. Critics note that data quality, inductive biases for sample efficiency, and architectural innovations (attention, residual connections) matter significantly. But as a historical diagnosis, its accuracy is hard to dispute.

  • Kaplan et al. 2020 — empirical quantification of scaling laws, the numerical confirmation of Sutton’s thesis
  • Silver et al. 2016 — AlphaGo, one of the key examples Sutton cites of general methods beating domain knowledge
  • Silver & Sutton 2025 — extends the bitter lesson argument: the next era is experience-based RL, not larger imitation datasets
  • Vaswani et al. 2017 — Transformer architecture, general enough to dominate NLP, vision, and RL simultaneously

Graph View

Backlinks