AI engineering
The Bitter Lesson: What Sutton Means by Scale
Rich Sutton argues that search and learning eventually beat hand-built knowledge in AI research. The distinction matters when deciding which rules to keep.
Listen to the audio analysis
Create your ownWritten for builders and creators.
The most successful breakthroughs in AI haven't come from humans teaching machines how to think. Instead, they've come from machines learning to outthink the teachers through sheer computational force.
Editor's note
Why this matters now
Rich Sutton's 2019 essay The Bitter Lesson argues that general methods able to use more computation have repeatedly overtaken systems built around human knowledge of a task. His examples come from chess, Go, speech recognition and computer vision.
For an engineer building with models, the useful distinction is between a rule intended to make a system intelligent and a rule expressing what the system is allowed to do. Sutton challenges the first kind when it prevents search or learning from improving. His historical argument does not make a payment limit or access-control policy obsolete.
The source
What it says
Distilled from the original. The notes above and below are the editor's own.
The Resource Researchers Treat as Fixed
Sutton's argument begins with the falling cost of computation. A method that seems wasteful at one point can become effective when much more compute is available. A system built mainly around expert rules may have fewer ways to benefit from that change.
He identifies search and learning as general methods that can keep improving as computational resources grow. The “bitter” part is that carefully encoded human insights can deliver early progress, then become obstacles to methods that discover useful structure themselves.
Sutton allows that knowledge-based approaches can help in the short term. His concern is their long-term limits and the research effort they absorb. The argument depends on resources continuing to grow and on methods being able to use them.
The Examples Behind the Claim
In chess, Sutton points to the success of deep search in the system that defeated Garry Kasparov in 1997. In Go, he emphasizes the combination of search and learning, including self-play, that overtook approaches built around more human game knowledge.
His speech-recognition account contrasts methods based on detailed human understanding of speech with statistical methods that could use more data and computation. His computer-vision account describes a movement away from hand-designed features toward learning representations.
These examples illustrate his thesis. They are not controlled experiments isolating computation from architecture, data and engineering. The essay argues from the pattern across fields, rather than estimating a universal return from another unit of compute.
Why a Useful Rule Can Become a Constraint
A hand-designed feature tells a system what its designer thinks is worth representing. That can reduce the work needed to get an initial result. It can also prevent the system from discovering a different representation that performs better with more data or computation.
Sutton's preferred direction is to build methods that discover and capture complexity. This leaves a substantial engineering task: choosing objectives, representations and learning or search procedures that can actually exploit the resources available.
The Transformer paper is a useful companion because it shows a concrete architectural choice and its computational consequences. The Chinchilla study adds another qualification: even when more compute helps, how it is divided between model size and data matters.
What the Essay Does Not Decide for a Product
The essay does not price a production system, choose its error tolerance or settle how much latency a customer will accept. A product has to work under today's constraints as well as leave room for future improvements.
Consider 2 illustrative rules. A hand-written heuristic for understanding a customer's intent might become less useful as a model improves. A rule requiring approval for a payment above a limit expresses policy. Better language understanding does not supply authorization to remove it.
That distinction keeps the essay useful without treating it as an instruction to delete safeguards or abandon domain expertise. Expertise can help define the task, assess results and identify failure. Sutton's warning concerns encoding today's understanding so deeply that the system cannot learn a better solution.
A Research Direction, With Open Choices
Search and learning are broad categories. The essay does not specify which architecture will succeed next or how quickly a particular hand-built system will be overtaken. It also does not promise that an application designed around learning will remain commercially useful.
The original essay is valuable as a challenge to a design assumption: does the method have a way to improve when resources increase, or does improvement require an expert to add more rules by hand?
Editor's note
What to do with this
The Bitter Lesson favors methods that can discover better solutions as computation grows. It does not make every explicit rule a mistake. For builders, the distinction turns on what a rule is doing: supplying a guess about how to solve the task, or specifying the task's real constraints. Keeping those roles clear preserves Sutton's challenge without turning a historical research argument into a blanket instruction for product design.
The original
Rich Sutton · 13 March 2019
Read next
AI engineering
Chinchilla: What a Smaller, Better-Trained Model Showed
DeepMind matched Gopher’s training budget with a smaller model and more tokens. The result explains why parameter count alone is a poor guide.
4 min read6 min listen
AI engineering
Attention Is All You Need: Why Drop Recurrence?
The 2017 Transformer traded sequential processing for attention between tokens. See how that speeds training and creates a long-context cost.
5 min read10 min listen