From Computers and Brains to More Human-Like AI

A computer handles long sequences of exact calculations with ease. A person doing the same work is prone to distraction and mistakes. Yet while talking, you can reach out to catch a slipping cup without ever considering how to break the movement into steps. Tasks we can readily express as rules may strain the brain, while tasks whose steps we struggle to explain can feel effortless.

In The Computer and the Brain, von Neumann approaches this contrast through the properties of the components. Comparing the electronic computers of his time with nervous systems, he observes that neurons operate more slowly and their signals offer limited precision, yet the brain performs complex activities quickly. His reasoning directs attention to organization. Slow components call for extensive parallel work; limited precision calls for suitable representations. The book’s discussion of speed, precision, and logical depth turns the vague impression of a fast-thinking brain into tractable questions.

A different notation permits a different computation

Digital arithmetic depends on definite place values: a 1 in the units position means something very different from a 1 in the hundreds position. A small positional change can produce a large numerical change. The nervous system leads von Neumann toward statistical properties of pulses, such as their frequency over an interval. An individual pulse has the discrete appearance of occurring or not occurring, while the processes producing it and the information carried by a pulse train are more complex. He thus turns toward computation combining continuous processes, discrete events, and statistical representation.

A simple analogy helps. If a message rests on one digit at a particular position, misreading it can change the message. If the message resides in a distribution formed by many observations, the remaining evidence can offset disturbances in a few of them. Reliability can arise from relationships across the whole. We can recognize a somewhat blurry face without first recovering every pixel exactly. A representation determines which details a system must preserve and which changes it can tolerate.

Consider catching the cup again. While it is still slipping, vision, arm movement, and touch are already contributing, and the action keeps changing with fresh feedback. Imagining that the brain first solves an exact mechanics problem and then executes the answer leaves out crucial processes. This is what I take from von Neumann’s comparison: speed, precision, parallelism, and feedback can be combined in different ways. A capability need not have only one computational organization capable of producing it.

Replace the neurons one by one

In his interview with Andrew Marr, Geoffrey Hinton discusses a thought experiment: replace a neuron with a machine that receives the same signals, produces the same responses, and preserves its interactions with other neurons. After one replacement, your capabilities are unchanged; after another, they remain unchanged. Continue until the brain consists entirely of artificial components. At which step should intelligence suddenly disappear? This gradual-replacement scenario focuses attention on functional organization: the material changes while the relationships producing the capabilities are preserved.

This makes the expressive capacity of neural networks particularly attractive. A universal approximation theorem tells us that a network satisfying the relevant conditions can approximate a continuous function on a compact domain to arbitrary precision. Treating changes in brain state as mappings to learn helps explain the hope of simulating them with networks. The next difficulty is finding the mapping. Providing enough representational capacity and providing the training experience needed to learn the desired behavior are two separate tasks.

A person’s inputs are much richer than the words they read and speak. Withdrawing a hand from a hot cup, finding a new route after a wrong turn, and revising trust after someone’s response all shape later behavior. These experiences share a structure: actions change the next inputs, while states left by the past influence the next action. Text preserves a great deal of knowledge from this process, but captures only part of living through it. Learning a complete cognitive system means engaging with this continuing chain of causes and consequences.

Language is one capability within a larger system

The brain also has distinguishable functional specializations. A striking example is the separation of language from other forms of thinking: some people with severe aphasia can still perform mathematical, reasoning, and planning tasks. The neuroscientific evidence discussed by Fedorenko and colleagues shows that the language network does not perform all this work. The voice that seems to keep talking inside our heads can lead us to overestimate language’s place in cognition as a whole.

Following this division of labor, I prefer to think of a language model as a part of a cognitive system that has already become very capable. Through language, it encounters knowledge, reasoning, and social experience; it can also learn to process images, sound, and actions. These capabilities need not develop together. Studying linguistic competence separately from the ability to use language in the world suggests more concrete directions: what the model already does, what needs new training experiences, and what is better served by another mechanism.

Return to the slipping cup. Perception supplies its position and movement; motor control keeps correcting your reach; memory tells you it was a gift from a friend; concern about losing it makes you interrupt the conversation. Each function has its own problem to solve, and each can change the others’ work. Important things attract attention, attended things are more likely to enter memory, and past experience changes what matters next time. Adding these capabilities also means making them influence one another within the same event.

Try changing the temperature of the cup below and its significance to you. It is still slipping, but new sensory information and memories alter priorities, which in turn affect how to reach for it.

Emotion is especially useful to consider this way. Fear raises the priority of certain risks, curiosity makes exploration worth the effort, and attachment gives particular objects or people a different weight in our choices. To make AI more human-like, these functions offer directions for design and training: what deserves sustained attention, which experiences should change future choices, and which goals should persist beyond one conversation. Giving an answer a happy or sad tone does not by itself establish these connections.

In engineering terms, this suggests persistent memory, a continually updated state of the environment, actions with real consequences, and mechanisms for deciding which outcomes to pursue. Crucial training happens after these are connected: the system must learn when to recall, observe, or act, and let the results change its future behavior. Functional specialization gives us places to begin; experience coordinating those functions develops the parts into a whole.

Von Neumann helps us see how computation can be organized differently from pencil-and-paper arithmetic. Following that question into the present, I want to ask how such computations jointly sustain an agent’s relationship with the world. As the cup slips, what does the system see, what does it remember, why does it reach out, and what will the experience leave behind? A more human-like AI takes shape in the connections among those questions.