The 'Bitter Lesson': Why Raw Computing Power Keeps Beating Human Expertise in AI
A 2019 essay by a leading AI researcher laid out a principle that has shaped every major AI breakthrough since. The short version: brute-force scale wins, every time.
Key points
- Richard Sutton published "The Bitter Lesson" on his website in March 2019, arguing that general AI methods powered by more computing consistently beat hand-crafted, expert-driven ones.
- The principle has received hundreds of formal academic citations and was independently echoed in at least two major Google research papers published in 2022.
- Examples spanning chess, the board game Go, speech recognition and computer vision all follow the same pattern: scale beats expertise.
- A 2024 study reviewed 20 years of computer vision research and found strong evidence that the field has followed the bitter lesson throughout.
- As of 2025, one research team argues the lesson still has not been fully absorbed in speech recognition and brain data science.
What is the bitter lesson, exactly?
It is this: when you give a simple, general AI method enough computing power and enough data, it will eventually beat a more clever method built by human experts. Always. The lesson is called "bitter" because many researchers spent careers building sophisticated, human-inspired AI systems, only to watch a simpler scaled-up approach overtake them.
Richard Sutton, a computer scientist at the University of Alberta and one of the founding figures of modern AI, coined the term in a short essay posted to his personal website in March 2019. The idea has since been cited hundreds of times in formal research, according to Google Scholar.
Where has this actually happened?
Sutton lists several clear cases.
In chess, IBM's Deep Blue defeated world champion Garry Kasparov not by mimicking how grandmasters think, but by using a straightforward search algorithm, basically trying millions of possible moves, backed by enormous specialised hardware. Earlier programs that tried to encode chess wisdom directly fell behind.
In Go, a far more complex board game, Google DeepMind's AlphaGo surpassed human players by leaning much less on human expertise than its predecessors. Then AlphaGo Zero went further: it removed human knowledge entirely and learned purely by playing itself, and it was stronger still.
Speech recognition followed the same arc. Systems trained on huge volumes of recorded speech beat the carefully hand-crafted approaches researchers spent the 1970s building. Deep learning, which uses layered mathematical networks loosely inspired by the brain, pushed that lead even further.
Computer vision tells the same story. Algorithms built to mimic how human eyes detect edges and shapes were overtaken by convolutional neural networks, a type of AI that makes far fewer assumptions about how vision works and simply learns patterns from data.
| Field | Old expert-led approach | Scale-based winner |
|---|---|---|
| Chess | Grandmaster knowledge encoded by hand | Deep Blue's move-search algorithm |
| Go | Human strategy rules | AlphaGo Zero (self-play only) |
| Speech | Hand-crafted 1970s phonetics models | General models trained on large datasets |
| Computer vision | Explicit edge and shape detectors | Convolutional neural networks |
Why does this matter for people outside the lab?
It explains why today's AI products look the way they do. ChatGPT, Claude and similar tools, which are built on large language models (software trained on vast amounts of text to predict and generate language), are products of exactly this philosophy. They are not carefully hand-programmed to understand English. They are general systems scaled to enormous size.
For workers and business owners wondering where AI investment is heading: follow the computing budget. Sutton's conclusion is that money spent on scale, more data, more chips, more parameters, tends to pay off more reliably than money spent on clever human-designed shortcuts.
One honest caution: survivorship bias runs through every example here. We remember the systems that won. Plenty of scaled-up approaches also failed quietly. Scale is a strong signal, not a guarantee.
Takeaway: If you are deciding which AI tools to trust or adopt, prefer ones backed by large, well-funded training pipelines over ones selling proprietary "secret sauce" logic. History says the former wins.
Common questions
Does this mean human expertise is worthless in AI?
Not quite. Researchers note that without some reasonable starting assumptions, even large models cannot generalise well. Scale reduces the need for expert input dramatically, but it rarely eliminates it entirely.
Has every field absorbed this lesson yet?
No. A 2025 paper presented at the International Conference on Machine Learning argues that speech recognition and brain data science have still not fully caught up to what Sutton observed six years ago.
