The Bitter Lesson
After 70 years of AI research, the pattern is clear: hand-coding human knowledge always loses to brute-force computation plus learning—a "bitter" lesson because it makes researchers' domain expertise obsolete.
Read Original Summary used for search
TLDR
• Only two methods scale arbitrarily with computation: search and learning. Everything else plateaus.
• Historical pattern across chess, Go, speech recognition, and vision: human-knowledge approaches dominate initially, then get crushed when computation scales
• The psychological trap: encoding domain knowledge is satisfying and works short-term, but inhibits long-term progress by making systems less suited to leverage computation
• We should build AI agents that discover, not agents that contain what we've discovered—the complexity of minds is "irredeemably complex" and can't be hand-coded
• Moore's law makes this inevitable: what works in a 3-year research project becomes obsolete when computation increases 100x
In Detail
Rich Sutton argues that 70 years of AI research reveals a consistent, uncomfortable pattern: general methods that leverage computation always beat approaches that encode human knowledge, despite the latter working better initially. This happens because researchers optimize for constant computation (typical research project timescales), but Moore's law ensures massively more computation becomes available, making human-knowledge approaches obsolete. The lesson is "bitter" because it devalues researchers' domain expertise in favor of brute-force scaling.
The pattern repeats across domains. In chess, search-based methods crushed human-knowledge approaches in 1997, despite researchers' dismay. In Go, it took 20 years longer, but the same thing happened—human knowledge about the game proved "irrelevant, or worse" once search and self-play learning scaled. Speech recognition saw statistical methods (HMMs, then deep learning) dominate over phoneme-based and vocal-tract knowledge. Computer vision abandoned edges, cylinders, and SIFT features for convolution-based deep learning. Only two methods scale arbitrarily: search and learning.
The meta-lesson: stop trying to encode how we think minds work. Human minds are "tremendously, irredeemably complex"—the complexity is endless and can't be hand-coded. Instead, build meta-methods that can discover and capture arbitrary complexity through search and learning. We want AI agents that can discover like we can, not agents containing what we've discovered. Building in our discoveries obscures the discovery process itself.