Suppose you want a drone that navigates by echolocation, the way a bat does: emitting sound and locating itself from the echoes. You could train it in a simulated environment. But someone would have to decide which details that environment should contain—and the real drone would still encounter physical conditions, including wear in its motors, that its training world did not capture.
Khurram Javed raises the drone on the Training Data podcast to explain an idea he developed with Rich Sutton: an intelligent agent has limited resources, while its surroundings contain more than it can represent. It must keep discovering what its model leaves out.
They call this the Big World Hypothesis. It leads them into an argument with the hosts about synthetic data and self-driving cars, then into a shared question: how can a machine learn enough about the world to imagine something it has never done?
The idea and where it came from
An agent, in this vocabulary, is a system that senses its surroundings and acts—a robot, a game-playing program or a person. Its model is an internal representation it can use to predict the consequences of actions.
Javed and Sutton's position paper, first presented at the Finding the Frame workshop at the Reinforcement Learning Conference in 2024, starts with the mismatch between an environment's complexity and an agent's limited perception, memory and computation.
On the podcast, Sutton describes the world as infinitely big. The paper's argument does not require mathematical infinity. A world is big relative to the mind trying to understand it. Adding more computing power does not necessarily eliminate that mismatch: the surroundings can include other increasingly capable agents whose behavior is difficult to predict.
The result is a need for selective, useful approximations rather than a complete internal copy of reality. The paper favors a Newtonian description over a quantum one when the simpler account serves the agent's goals efficiently. More detail is not automatically more useful.
The same reasoning makes learning an ongoing activity. As an agent encounters different parts of the world, it needs different knowledge. The paper calls this tracking: giving resources to what is useful now and releasing resources devoted to less relevant information. Its example is a computer's small, fast cache, which holds information likely to be needed soon rather than everything the computer could ever use.
This extends Sutton's 2019 essay, The Bitter Lesson. That essay argues that search and learning have repeatedly benefited from growing computational power more than approaches built around researchers' domain knowledge. The Big World Hypothesis adds a reason the process cannot simply end with a sufficiently comprehensive model: the agent must keep adjusting its approximations to the part of the world it actually inhabits.
The synthetic-data objection: who decides what to simulate?
The podcast exchange begins sharply. A host asks whether synthetic data—training material generated by programs or models—offers a way to keep scaling beyond the finite supply of human text on the internet.
Sutton answers: “That's just a big mistake.”
Javed develops the objection through the echolocating drone. To generate suitable training data, he argues, you may first need experts who understand the relevant physical problem. “But the domain expert has to exist first.” Generating more examples does not, by itself, supply the expertise needed to decide what those examples should contain.
A host pushes back: cannot the training results decide which dataset is better? Javed's reply is that someone still has to choose the data-generation process worth testing. His thought experiment is to send the engineers at the large AI labs on vacation and ask who would generate the right synthetic data then.
Sutton emphasizes a related problem: fidelity. The drone's interactions with the physical world include friction and motor wear. A simulator contains only a selection of those details. Running it for longer cannot uncover something its model has no way to represent.
Their paper makes a similar distinction even in Go, where the board's rules are fully specified. Simulating a legal move is one problem; predicting what a sophisticated opponent will choose is another. Knowing the rules does not mean knowing everything relevant to play.
The self-driving counterexample
A host offers a practical objection: their understanding is that newer self-driving companies have trained largely in simulation, then adapted their systems for the road, with useful results.
Javed asks how many engineers were needed to build those simulators. He imagines a repeated process: train in simulation, discover an unacceptable gap between simulated and real behavior, have people repair the simulator, then train again. His question is whether AI should be limited to problems for which a large team can first construct and maintain such an environment.
He is not proposing that agents give up simulation. “There is obviously value in learning from simulation,” he says. Instead, an agent should learn a predictive model from its own experience and use that model to consider possible actions. When experience contradicts a prediction, the agent should be able to revise the model rather than wait for an engineer to repair it.
That learned model would still be an approximation. The proposed advantage is not that the agent creates a perfect simulator, but that it can participate in correcting its own errors.
This leaves a real question about scope. A bounded simulator need not contain the whole world to be useful for a bounded task—the hypothesis itself favors useful approximations. The disagreement is over how far human-maintained environments can take an agent, what they cost to maintain, and when learning directly from experience becomes the better way to close their gaps.
Prior knowledge is not the enemy
Sutton also resists the idea that The Bitter Lesson forbids giving an AI useful knowledge at the start.
“There's nothing wrong with prior knowledge,” he says. Knowledge acquired earlier becomes the starting point for learning something new. His complaint is about designs and research habits that let existing human knowledge crowd out further learning, not about an inherent conflict between the two.
The hosts raise DeepMind's game-playing systems as an example of removing human priors. The history has several distinct steps. Earlier AlphaGo systems began with human game records. AlphaGo Zero instead began with random play and generated its training experience through games against itself. Its network guided a search through possible moves; the resulting games supplied material for improving the network, which then guided later searches.
AlphaZero took a shared learning algorithm into chess, shogi and Go, but trained it separately for each game. Each training process started with that game's rules. This was one general design acquiring different expertise—not one trained network simultaneously mastering all three games, and not learning without any supplied structure.
For future robots, Sutton imagines copying an already trained agent and letting the copy continue learning. Each new machine would not have to rediscover everything from scratch. Prior knowledge would provide a beginning rather than a finishing line.
Rockets and planning beyond the next millisecond
The hosts' next challenge is imagination. Learning through interaction can explain many animal and childhood skills. But building a rocket requires thinking ahead before the launch, drawing on accumulated knowledge and reasoning about something not yet attempted.
Sutton agrees: “You have to be able to plan. You have to be able to imagine.”
Javed's account makes room for learning from culture and schooling. It also asks how someone goes beyond inherited knowledge. In his telling, scientific advances involve constructing new abstractions from accumulated experience, using them to make predictions and plans, and discovering where they work better than the old account.
An abstraction keeps the distinctions useful for a particular problem while leaving other details aside. For planning, that includes thinking in extended activities rather than only tiny movements. A rocket plan cannot practically enumerate every millisecond of every component's behavior.
Sutton contrasts this challenge with settings where the rules for moving between states are supplied. Go provides legal moves. A proof assistant such as Lean provides a state of outstanding goals and assumptions; its tactics can solve a goal or turn it into subgoals, while its kernel checks the resulting proof. These give a search process explicit steps to work with.
The harder ambition is for an agent to discover useful abstractions, learn what happens when it acts through them, and use those predictions to plan.
The Alberta Plan, a twelve-step research roadmap Sutton wrote with Michael Bowling and Patrick Pilarski, breaks that ambition into smaller problems. It was first posted in 2022, with an available revision from 2023. Early steps address continual learning; later ones develop models, planning and abstraction.
One building block is an option: a way of acting, called a policy, together with a condition for stopping. Its learned model predicts where the activity will end and how much reward—the system's measure of success—it will accumulate along the way. A planner can then compare extended activities without simulating every intervening movement. Which activities deserve to become options is itself part of what the agent is supposed to learn.
What supports the proposal—and what remains open
The Big World Hypothesis is a position, not an experimental demonstration that one learning architecture will succeed. Its supporting arguments concern resource limits, useful approximation and the need to adapt as experience changes. The authors propose tests that explicitly constrain an agent's computation, memory or energy; different constraints may favor different algorithms.
Nor does their argument mean learned world models are absent from AI research. Meta's V-JEPA 2, for example, is a video-trained predictive model with robot-control demonstrations. Sutton's concern in the conversation is more specific: learning models and planning with self-discovered abstractions. He acknowledges that there may be counterexamples to his assessment of the field.
The Alberta Plan depends on making that process reliable. How should an agent discover an abstraction worth keeping? How can it tell that an imagined outcome is trustworthy enough to act on? When should it revise the model, and when should it seek more experience? Those are research problems, not capabilities established by the hypothesis. The same applies to the practical balance between a useful human-built simulator and a model the agent maintains itself.
The proposed destination nevertheless has a distinct shape. Asked whether success means a single mind that can swing from branches and design spaceships, Sutton answers: “It's a single design.” There would be many minds, learning different things from different experiences.
A copied robot could begin with another robot's knowledge, then adapt to its own surroundings and its own wearing motors. Under the Big World Hypothesis, that divergence is not a failure to build the one model that knows everything. It is what a finite learner needs to do.