Rich Sutton
Related people
Episodes
Notes
Training Data · David Silver and Lasse Espeholt of Ineffable Intelligence: A…
-
Why Will Simulation Build Superintelligence Faster?
Sonya Huang asked where the ceiling on simulation sits and when an agent should enter the real world. Sonya Huang said Rich Sutton is very dismissive of simulation. One panelist says the right question is whether a superintelligence can be built in simulation. If a superintelligence can be built in simulation, one panelist believes it will happen 10 times or a hundred times faster, because simulation can be infinitely parallelized.
MetNet is the example. A weather forecasting model from the Google Brain Amsterdam office, built together with Google Research, used machine learning to predict short-term weather forecasts. MetNet was way better than the physics-based simulations. The physics si
Training Data · Rich Sutton and Khurram Javed: Why AI Models Stop Learning, and…
-
Why Are Big AI Labs Stuck in a Local Minimum?
Khurram Javed says the big labs are stuck in a local minimum. A move toward the new algorithms will almost certainly look worse before it looks better. A different paradigm will not hit state-of-the-art performance on the first day. The labs are locked into a product, and the current paradigm still lets them keep scaling, so they cannot take a path where the results get worse first.
The new paradigm is a bet. It depends on hard technical problems that people who have worked on them for many years already consider doable, and those problems take a long time to think through. He also says he does not see research groups that even believe a much more efficient system is possible. A group that
-
What Is Oak Lab's Biggest Ambition?
Rich Sutton says Oak Lab's biggest ambition is a mind that stays coherent while it keeps changing. A large language model stays correct because people do a lot of post-training and then freeze it. A human mind is always revising itself, and something still pulls it back into a good place instead of letting it drift. He wants that self-maintaining quality: a mind that can keep training itself and remain consistent.
He says the aim is within reach. In 2026 the computers are fast enough that he sees a vision and inklings of how the steps can be done, not a plan that is out of reach. The picture includes a trillion-parameter mind running on 20 watts. Khurram Javed says that is impossible with t
-
What Is Rich Sutton's Alberta Plan?
Rich Sutton says the Alberta Plan, which he laid out in 2022, turns broad ideas about AI into smaller chunks. The 12 steps are those chunks. Step two is continual deep learning, and he thinks that step is almost the most important because it unlocks the rest. Continual deep learning would let a system keep updating its model of the world.
The later steps are about abstractions. He does not mean one correct set of abstractions, because no one can say what those are. They depend on the world the agent is in, and the agent has to learn them. He offers those two jobs, finding abstractions and continual deep learning, as maybe the key pair, not as a settled list. Khurram Javed says elite athlete
-
Why Do LLM Assistants Stop Learning After Deployment?
Rich Sutton says the popular language-model assistants are not learning from experience. Their weights never change. A memory of the user, or extra context, can change the state the model conditions on, and the next word can use that state. The model itself is not learning. The structuring and the new concepts that built the model were weight learning, and that should not happen only once.
He says the field does too much pre-training and post-training before launch, then does not let the model learn while someone is using it. Pre-training and post-training are fine. The disagreement is the stop. Khurram Javed says a few products, including Cursor's tab autocomplete, do update weights from a
-
Why Does Rich Sutton Call Synthetic Data a Big Mistake?
Rich Sutton says synthetic data is not a general method that uses computation. He calls it a big mistake, and maybe the next big lesson. The idea has floated around Alberta for five or ten years. He calls it the big world hypothesis, which Khurram Javed later wrote up: the world is incredibly big, there are infinitely many things to learn, and the world is massively more complex than any agent because it contains many other agents.
Humans still decide which synthetic data is good, so the method stays bottlenecked by experts. A short program can only describe a small world. It will not match the real world, and it cannot tell you what is happening in another person's mind or how a drone's mo
-
Are Large Language Models an Example of Rich Sutton's Bitter Lesson?
Rich Sutton says large language models are both a positive and a negative example of The Bitter Lesson. They scaled with computation by taking in the internet, and that produced a much more capable system. The internet is finite, the world is massively bigger than everything stored there, and leaning on that stock of human knowledge eventually holds the systems back.
The essay is not a ban on fancy algorithms. He wants algorithms that scale with computation, such as search and learning, rather than methods that scale with human input. He wrote it from decades of watching the field, including the symbolic AI era, get distracted by putting human knowledge in. He had given versions of the talk
-
What Is Continual Backprop?
Rich Sutton says Continual Backprop is ordinary backprop that also keeps planting new units with fresh random weights. Backprop starts with random weights, and that variety gets used up. Continual Backprop keeps injecting a little randomness so the network can keep learning. He and Khurram Javed published it in Nature a couple of years ago.
A naive update on one sample changes the whole model and hurts older knowledge. Khurram Javed says Cursor avoids that by training on a large batch from many users, which works only when that data exists. Most uses have a single stream, and the naive update then destroys prior knowledge. Catastrophic forgetting is totally curable with the right algorithm.
-
Why Don't Animals Learn by Supervised Learning?
Rich Sutton says no animal learns by supervised learning, because nobody supplies examples of how its muscles should twitch, and muscle movement is the animal's output. School is a tiny fraction of what people learn, and even in school nobody says how to move the mouth to pronounce a word. Squirrels do not go to school. He treats school as a recent human invention, not the essence of intelligence, and he takes animal learning as inspiration rather than a requirement that an AI copy babies.
Khurram Javed gives the speech test. If he moved his mouth, tongue, and vocal cords exactly the way Rich Sutton does when he says Paris, a different sound would come out. No one else knows the right movem
-
Why Does Rich Sutton Say Prior Knowledge and Learning Should Be Friends?
Rich Sutton says prior knowledge and learning should be friends. The Bitter Lesson opens by saying there is no need for a conflict between them. You can put some prior in and then start learning. There is no reason in principle the two have to be opposed. Both are about knowledge. Life is gaining knowledge and having knowledge. Prior learning is what you already have, and then you learn more. Once you have gotten more, that becomes the prior. In practice the two have been enemies, because people who favored existing human knowledge wanted that knowledge to win and wanted learning dismissed.
AlphaGo and AlphaZero made him happy and made him feel vindicated. It could have gone either way, bec