# The big world hypothesis and bounded rationality > The Big World Hypothesis seems to be a concept that redescribes the concept of bounded rationality in the context of artificial intelligence. If so, a well-made RL agent may sometimes commit bizarre errors (Heuristics and biases; Daniel Kahneman), but in the big picture it will be far more rational than one might expect (Ecological rationality; Gerd Gigerenzer), and will often display expert-level decision-making ability (Naturalistic decision making; Gary Klein). [The Big World Hypothesis](https://wiki.g15e.com/pages/The%20big%20world%20hypothesis.txt) seems to be a concept that redescribes the concept of in the context of [artificial intelligence](https://wiki.g15e.com/pages/Artificial%20intelligence.txt). If so, a well-made [RL](https://wiki.g15e.com/pages/Reinforcement%20learning.txt) agent may sometimes commit bizarre errors (; ), but in the big picture it will be far more rational than one might expect (; ), and will often display expert-level decision-making ability (; ). ## Are LLMs a dead end? On , I listened to a conversation between and (). Sutton's argument can be summarized in two main points. - We should not use human-made data (text written by humans, labels attached by humans). - [Reinforcement learning](https://wiki.g15e.com/pages/Reinforcement%20learning.txt), in which the agent finds the answer for itself, is the answer. This was an argument I had already encountered while reading Sutton's essay released last April, "[Welcome to the Era of Experience](https://wiki.g15e.com/pages/Welcome%20to%20the%20Era%20of%20Experience.txt)," but after listening to the conversation, it struck me as even more hard-line than it had when I read it in writing. For example, Dwarkesh asked, "Couldn't you take an [LLM](https://wiki.g15e.com/pages/Large%20language%20model.txt) as a base and stack [RL](https://wiki.g15e.com/pages/Reinforcement%20learning.txt) on top of it to do online continual learning?" and Sutton flatly said no even to that. His point was that if you feed it human-made data, you get results right away, so it creates a tendency to cling to human data unnecessarily for too long; we need to get away from human data as quickly as possible. Come to think of it, 's [Is human data enough?](https://wiki.g15e.com/pages/Is%20human%20data%20enough.txt)—Silver being both someone who recently co-wrote [Welcome to the Era of Experience](https://wiki.g15e.com/pages/Welcome%20to%20the%20Era%20of%20Experience.txt) with Sutton and an early member of —also mentions cases where AI performance improves as human intervention is reduced, and cases where algorithms obtained through meta [RL](https://wiki.g15e.com/pages/Reinforcement%20learning.txt) (making the RL algorithm itself with RL) outperform algorithms designed by humans. ## The big world hypothesis During the conversation, Sutton mentioned something called the "[Big World Hypothesis](https://wiki.g15e.com/pages/The%20big%20world%20hypothesis.txt)," and since I didn't know what that was, I looked around a bit and ended up reading a paper Sutton coauthored ([The Big World Hypothesis and its Ramifications for Artificial Intelligence](https://wiki.g15e.com/pages/The%20Big%20World%20Hypothesis%20and%20its%20Ramifications%20for%20Artificial%20Intelligence.txt); June 2024). Thanks to that, I was able to understand Sutton's position a little better. The big world hypothesis is the hypothesis that the world is too big and too complex, while the agent is too small and too limited, so the agent cannot be fully equipped with solutions to every problem it will encounter in the world. The paper goes further and makes an additional claim here. According to the paper, the big world hypothesis is not a temporary problem that exists only because current computers are too limited; it is something that will inevitably continue to exist no matter how good computers become. If one accepts these two claims—that an agent cannot solve all the world's problems in the textbook way, and that this problem will continue to exist in the future—then the current [LLM](https://wiki.g15e.com/pages/Large%20language%20model.txt) approach has a fundamental limitation. No matter how much you scale up the model or pour in data, there will still be problems that cannot be solved (because the world is too big). Therefore, instead of the approach of finishing training and then releasing it into the world (inference), the solution is an approach in which it continuously learns and adapts in the world according to the situation it finds itself in and the problems immediately facing it (online continual learning). ## Bounded rationality But put another way, the "big world hypothesis" could also be called the "limited agent hypothesis." This suddenly brings 's concept of to mind. (Perhaps the association also happened because Sutton mentioned Simon in another context during the conversation above.) Bounded rationality is the idea that humans are not agents with unlimited cognitive capacity who maximize utility, but rather agents who must make workable decisions within limited cognitive capacity, limited information, and limited time. In other words, human rationality is bounded by cognitive capacity, information, and time. Seen this way, humans are already agents evolutionarily designed with Sutton's big world hypothesis in mind. There is no need to say only humans are like this, either. Almost all life (perhaps all life) is an agent that acts according to bounded rationality within its own ecosystem. ## HB, NDM, ecological rationality Once you think about bounded rationality, there are three concepts that come to mind almost automatically. is the leading figure of the school. Because humans are inevitably satisficers rather than maximizers due to their limitations, they sometimes make decisions that are not optimal (biases, mistakes, errors, and so on), and the HB school studies these problems that arise from bounded rationality. deals with the concept of . Humans use various heuristics in order to make fast and frugal decisions under limited conditions (limitations of cognitive capacity, information, and time), and the claim is that although these heuristics may look illogical or irrational when viewed narrowly, in the larger frame they are often in fact more rational. is the leading figure of the school. He studies cases in which highly trained experts make correct judgments in an instant under urgent circumstances, and the principles behind them (). ## Bigger models vs. better training environments Now let us return to the [big world hypothesis](https://wiki.g15e.com/pages/The%20big%20world%20hypothesis.txt) and connect the three concepts above to the context of RL agents. Most life, including humans, learned how to survive in a "big world" with only bounded rationality through the search algorithm called [evolution](https://wiki.g15e.com/pages/Evolution.txt), so RL agents should naturally be able to do so as well. In this context, what the agent needs is not more parameters, but a training environment that enables it to discover more efficient solutions (fast and frugal heuristics). A well-made [RL](https://wiki.g15e.com/pages/Reinforcement%20learning.txt) agent may sometimes commit bizarre errors (; ), but in the big picture it will operate far more rationally than one might expect (; ), and will often display expert-level decision-making ability (; ). If one accepts Sutton's assumptions, this is probably a reasonable guess. ## Evolution If the big world hypothesis is correct, and if we count living things on Earth as RL agents too, then it seems one also arrives at the conclusion that the countless heuristics used by living beings may not be things they adopted unavoidably because selection pressure was insufficient or because there was not enough time for adaptation, but rather that for quite a large number of problems, finding that sort of heuristic may have been the correct answer from the start.