Quick answer
The Bitter Lesson is AI researcher Richard Sutton’s 2019 observation that general methods which improve with more computing power and data keep beating systems built around carefully hand-coded human expertise. For anyone building with AI, the practical lesson is: don’t spend weeks building complicated workarounds for today’s model limits. Keep your setup simple, put your effort into your own knowledge, data and checks, and expect the next general model to do more.
Key takeaways
- Sutton’s essay is short, and the “bitter” part is that hand-built expertise gets overtaken.
- It has played out in chess, Go, speech recognition, computer vision and now, possibly, robotics.
- For your business: build light, replaceable AI workflows, not elaborate patches.
- Invest in what lasts: your own data, examples, instructions and quality checks.
This guide supports our main article, A Coding Agent Ran a Robot With No Extra Training: Stanford’s VIA and the Bitter Lesson, where Nate B Jones uses the Bitter Lesson to explain why a general coding agent could run a robot with no extra training. Here’s what the idea means for the AI tools and workflows you build.
📘 Main article: See the robot study that sparked this: A Coding Agent Ran a Robot With No Extra Training: Stanford’s VIA and the Bitter Lesson.
What is the Bitter Lesson?
Richard Sutton is one of the founders of reinforcement learning. In his 2019 essay The Bitter Lesson, he argued that AI researchers keep repeating a pattern:
- They work out what makes a problem hard and build that understanding into a system.
- It works, for a while.
- A more general method that improves with more computing power, learning and search comes along.
- Eventually it overtakes the carefully designed system.
The lesson is “bitter” because years of clever, hand-built expertise end up being beaten by a simpler approach that just keeps scaling.
Where has the Bitter Lesson shown up?
| Field | Hand-built approach | What won over time |
|---|---|---|
| Chess | Programmed human chess knowledge | Massive search with more computing power |
| Go | Expert rules and patterns | Self-play learning at scale |
| Speech recognition | Hand-designed models of human speech | Statistical and learned methods |
| Computer vision | Hand-designed image features | Deep learning on large datasets |
| Robotics (emerging) | Specialised robot training for each task | General agents with a good interface (early evidence) |
📌 Note: The robotics row is still an open question. Studies like Stanford’s VIA are early evidence, not a settled result.
Why does it matter if you use AI tools?
Small businesses face a small-scale version of the same choice. When a model can’t quite do something, it’s tempting to build a big workaround: long chains of prompts, complicated automations or custom code to patch a weakness.
Sometimes that’s worth it. But if the next model release fixes the weakness, your workaround becomes extra maintenance, or even gets in the way.
How do you apply it to what you build?
- Ask “will a better model fix this?” If yes, use a simple stopgap rather than a big build.
- Keep workflows modular. Make it easy to swap the model behind a workflow, so you can test a new one in minutes.
- Invest in your own material. Clear instructions, good examples, your data and brand rules stay useful whatever model you use.
- Build checks, not crutches. Ways to check output (tests, checklists, review steps) stay valuable as models improve.
- Re-test when new models arrive. Re-run your hardest task on new releases and remove workarounds you no longer need.
💡 Tip: Need a quick way to judge whether a new model is worth trying? Use The Founder Litmus Test for New AI Models.
When is building your own still worth it?
The Bitter Lesson isn’t a reason to build nothing. Custom work still makes sense when:
- It captures your knowledge, processes or customer data, which no general model has.
- It handles a legal, safety or quality requirement you must meet today.
- The time saved now clearly outweighs the cost of rebuilding later.
Your next step: list your AI workarounds and mark each one “keeps value” (your knowledge or checks) or “patches a model weakness”. Re-test the second group with the latest model.
Frequently asked questions
Who wrote The Bitter Lesson?
Richard Sutton, a leading reinforcement learning researcher, published the short essay in March 2019.
Is the Bitter Lesson always true?
It’s an observation about the history of AI, not a law. It has held in many fields, but how far it applies to areas like robotics is still being tested.
Does it mean expertise doesn’t matter?
No. Your knowledge still matters, especially in the instructions, examples and checks you give AI. The lesson is about not hard-coding workarounds a general model will soon handle.
How often should I re-test my AI workflows?
A good habit is to re-run your hardest task when a major new model is released, or every few months.
Related posts
- A Coding Agent Ran a Robot With No Extra Training: Stanford’s VIA and the Bitter Lesson
- The Founder Litmus Test for New AI Models
- Long-Running AI Agents: Why Yours Still Needs Everything Explained Again
- GPT-6.2 Astra Review: Cheaper and Efficient, but Is It Worth Switching?
