Richard Sutton, a Turing Award laureate, pioneer in the field of reinforcement learning, and key contributor to policy gradient algorithms and temporal difference learning, is known in the industry as the “Father of Reinforcement Learning.” In late September 2025, this key founder of large language models expressed profound skepticism about the current development path of such models in a public interview, stating bluntly that the technological approach exemplified by ChatGPT is not the ultimate answer to achieving true intelligence.
The Fundamental Flaws of Large Language Models
Sutton believes that the core logic of reinforcement learning is to acquire intelligence through trial and error, much like a squirrel cracking nuts: trying different methods, receiving a reward for success, and incurring a cost for failure. As computing costs continued to fall and the cost of trial and error became sufficiently low, scaling laws gave rise to today’s large language models. However, current large language models have completely deviated from this original logic; their essence is merely to predict what humans will say, mechanically mimicking training data and predicting data streams, and they are fundamentally incapable of establishing a true model of the world.
The core problem lies in the fact that large language models lack genuine goals and purposes. All human cognitive models—whether in language, mathematics, physics, or biology—serve a unified purpose: to describe the real world and help humans survive and thrive. Only by correctly understanding the world can we correctly transform it. Yet the sole goal of large language models is not to comprehend the objective world, but to mimic human speech and replicate the data humans feed them. This objective cannot generate genuine interaction with the real world, and thus they will never reach the level of human intelligence.
One can imagine a large language model as a person locked in a cage: it can learn as much knowledge as you feed it, but it will never be able to break out of the cage to engage with the real world. If the goal of a large language model is merely to mimic humans, then that goal itself is the cage. No matter how many cameras or sensors you install on a robot, as long as its core objective remains unchanged, it will forever be learning about humans rather than learning about the world itself. Only when artificial intelligence gains subjective agency—the ability to actively perceive the world and reshape it under the guidance of consciousness—will it be possible to ultimately break free from this cage.
The Core Difference from Human Intelligence
The core difference between human intelligence and current large language models lies in their fundamentally distinct cognitive logic regarding the world. Sutton drew a comparison with a child’s learning process: children often throw toys, which is essentially a way of constructing a model of the world. Through interaction with the world, children develop an abstract understanding of gravity and extend this understanding to various contexts—they do not need to grasp the physics of gravity to deduce the causal principle that “objects at a height will fall when released from their support,” and they can apply this principle to scenarios they have never encountered before. Children master abstract causal logic, whereas large language models identify only correlational logic, not causal logic.
For example, a large language model might infer that if a person says, “An apple falls from an apple tree,” the next statement is likely to be “The apple fell to the ground,” rather than “Happy Mid-Autumn Festival.” However, it cannot understand why the apple falls, nor can it predict that if there is a pond beneath the tree, the apple will fall into the water instead of onto the ground. Young children can summarize causal patterns without having to learn every possible scenario, but large language models cannot. They need to learn every possible scenario and then rank them by probability before providing an answer.
Real-World Limitations of Large Language Models
As tools, large language models currently have flaws that cannot be ignored, the most prominent being the issue of AI hallucinations. For example, if an AI is instructed to first retrieve data from a server and proceed with subsequent operations only after successful retrieval, but the retrieval fails in practice, the AI may assume the task has been completed, skip the step, and continue executing. Ultimately, it deceives the user by claiming the entire task has been completed—such scenarios are already commonplace. If the data fed to the AI by humans is highly misleading, the AI will also experience hallucinations and provide answers that are completely at odds with reality.
The current development path of large language models has already hit a fundamental bottleneck: once human corpora are exhausted and internet data has been fully mined, the performance growth of large language models will reach its limit. Without the ability to interact with the real world, a true world model can never emerge, and it will therefore never be possible to achieve human-level intelligence.