digest · rich sutton · 2019

The Bitter Lesson

A breakdown of Rich Sutton's influential 2019 essay on why general computation-leveraging methods (search & learning) consistently dismantle human-engineered heuristics in AI research.

the bitter premise

"The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin. Seeking an improvement that makes a human-like difference in the short term inevitably leads to a plateau."

01The core thesis

Rich Sutton observed a repeated 70-year historical cycle across every domain of AI: computer chess, Go, speech recognition, computer vision, and natural language processing.

Domain experts hardcode human heuristics → Short-term victory on current hardware → Hits complexity wall / plateau → Compute scales + General search/learning → Handcrafted heuristics permanently crushed

The lesson is called "bitter" because it wounds human ego: researchers naturally want to believe that our conscious introspections about how we see, speak, and reason are essential for building intelligent machines.

02Historical case studies

Domain	Handcrafted human approach (Defeated)	Compute-scaling approach (Winner)
Computer Chess	Deep human opening books, master heuristics, and hand-tuned piece evaluation functions.	Brute-force massive tree search (Deep Blue) and self-play AlphaZero.
Computer Go	Complex shape heuristics, stone territory patterns, and expert rules of thumb.	Monte Carlo Tree Search + deep neural network value/policy networks (AlphaGo).
Speech Recognition	Acoustic phoneme templates, vocal tract physical models, and handcrafted phonetic grammars.	End-to-end deep neural networks and sequence-to-sequence Transformers over raw audio.
Computer Vision	SIFT, HOG features, edge detectors, and hand-engineered spatial primitives.	End-to-end convolutional and vision Transformer architectures trained on raw pixels.
NLP / Translation	Chomskyan parse trees, morphological dictionaries, and rule-based linguistic grammars.	Autoregressive next-token prediction over massive raw text corpora (Transformers).

03The two grand pillars: search & learning

Sutton emphasizes that the only two algorithmic primitives that scale monotonically with exponential compute are Search and Learning:

1. Search (Inference compute)

Exploring candidate trajectories and solution spaces (e.g., MCTS, tree search, test-time compute, reasoning token expansion). Search leverages raw FLOPS to verify hypotheses without needing human intuition.

2. Learning (Training compute)

Optimization algorithms (e.g., Backprop, SGD) that fit high-dimensional statistical representations directly from raw perceptual data. Learning discovers richer, more subtle representations than any human expert could ever articulate.

04The psychological trap of domain heuristics

Why do AI researchers repeatedly make the same mistake? Sutton diagnoses two psychological causes:

Introspective illusion

We assume that because we consciously experience vision as edges and surfaces, or language as nouns and verbs, the machine must also explicitly represent these symbols.

Short-term benchmark games

Adding a hand-tuned heuristic often delivers an immediate +2% gain on a tiny current dataset. But it creates a ceiling that prevents the model from absorbing 100× more data and compute later.

05Implications for modern ML engineering

How does the Bitter Lesson apply to today's frontier work in LLMs, multimodal models, speech, and robotics?

Rely on General Primitives: Prefer generic matrix multiplication, self-attention, and simple loss functions (cross-entropy, flow matching) over specialized domain modules.
Design for Scaling: Always ask: "If compute becomes 10× cheaper next year, will my method get 10× better automatically, or will my hardcoded assumptions become an impediment?"
Raw Signals over Pre-Processed Abstractions: Feed raw audio waveforms, raw pixel streams, and raw token streams directly into the architecture instead of pre-extracting intermediate feature representations.

06The Bitter Lesson evaluation checklist

Design choice	Fails Bitter Lesson	Passes Bitter Lesson
Representation	Manually designed feature vectors or symbolic grammar rules	Dense embeddings learned end-to-end via gradient descent
Inductive bias	Rigid, narrow architectural constraints tailored to human intuition	Minimal structural inductive bias (e.g., permutation invariance, temporal causality)
Improvement vector	Writing more complex edge-case code	Scaling parameter count, training tokens, and search compute

← Previous: Richard Hamming Back to Research Taste Guide Next: Michael Nielsen →