General methods that leverage computation, such as search and learning, ultimately outperform human-knowledge-based approaches in AI, a pattern known as the bitter lesson.
- AI research shows that general methods leveraging computation consistently beat human-knowledge-based approaches in the long run.
- The Moore's law trend of exponentially cheaper computation makes scalable methods increasingly powerful.
- Researchers often favor human-knowledge approaches for short-term gains, but these plateau and inhibit progress.
- In computer chess, Go, speech recognition, and computer vision, human-knowledge methods were eventually overtaken by search and learning scaled with computation.
- AI should focus on meta-methods that discover complexity rather than building in human preconceptions about the world.
The Central Lesson
The bitter lesson from 70 years of AI research is that general methods that leverage computation are ultimately the most effective. This is driven by Moore's law, which makes computation exponentially cheaper over time.
Most AI research neglects this, treating computation as constant and thus relying on human-knowledge-based methods. However, over longer timescales, massive computation becomes available and makes general methods dominant.
Historical Examples
In computer chess, early methods leveraging human chess knowledge were surpassed by massive search in Deep Blue (1997). Similarly, in Go, human-knowledge approaches delayed progress until scaled search and learning succeeded.
In speech recognition, statistical methods like hidden Markov models outperformed knowledge-driven ones, leading to the dominance of statistics and deep learning. Computer vision followed a similar arc, abandoning engineered features for deep learning.
Why It Bites
The lesson is bitter because researchers are psychologically committed to human-knowledge approaches, which provide short-term satisfaction and progress. Yet these approaches plateau, and the ultimate success of general methods feels like defeat.
The lesson teaches that general methods like search and learning scale with computation, while the contents of minds are too complex to build in directly. AI should build meta-methods that discover complexity rather than embedding human discoveries.
Read this at any depth.
Install Depth and pick your level — Glance for a sentence, Summary for the gist, Read for the full take. Free daily quota, no signup needed.
Add to Chrome