The RunawAI

Detroit, 1935

In 1935, a twelve-year-old boy named Walter Pitts was running through the streets of Detroit, running away from a group of boys, to be exact. He knew where to hide, somewhere they would never think to look and would likely not even be allowed in, and that was the library. So there he went, hiding deep in the stacks until it closed for the evening.

Walter was born in 1923 to a tough, abusive father, so he knew home was no better. His father was a boilermaker who had no trouble using his fists to get his way, and he wanted Walter to drop out of school and go to work. The library became the one place where Walter could be safe, so there he went.

On this particular day, when Walter ran to hide in the library, he stumbled upon something that would leave him spending the next three days inside it: a three-volume, nearly 2,000-page book titled Principia Mathematica. Its big claim was that all of mathematics could be built from simple logic. Walter not only read all of it, but ended up finding several mistakes within it. So he wrote to the authors, a philosopher named Bertrand Russell and his colleague Alfred Whitehead. Russell, impressed by the young man's eye, invited him to study at Cambridge. Walter was twelve. Due to his age and circumstances, Walter knew he couldn't, so he declined the invitation.

Three years later, he heard Russell was visiting the University of Chicago. Walter ran away from home, never saw his family again, and ended up homeless in Chicago, sneaking into lectures.

The Brain

It was there that he met Warren McCulloch, a 42-year-old scientist who was trying to answer a question: how does a brain think?

The brain is made of billions of nerve cells called neurons. These neurons listen to signals from their neighbours. If enough signals arrive, a neuron fires its own electrical pulse. If not, it stays quiet. Each neuron is a bit like a doorbell, which either rings or doesn't. McCulloch saw that this on-or-off behaviour resembled logic. Consider three everyday sentences:

• "I'll go to the beach if it's sunny and I'm free."
• "I'll take an umbrella if it's raining or cloudy."
• "I'll go out if it's not raining."

A neuron that needs two incoming signals before it fires behaves like "and". One that fires when either signal arrives behaves like "or". A signal that blocks firing gives you "not". Russell's book had shown that these few building blocks could make up all of mathematics. Walter and McCulloch argued that neurons wired together in the right way could carry out logical reasoning too.

Walter also solved a problem that had stalled McCulloch. Neurons can loop back and feed on their own signals, which broke McCulloch's way of tracking time. Walter used a kind of maths where numbers wrap around, like the hours on a clock, and the problem went away. A looping signal keeps circulating, which makes it a way of holding on to information. That's where memory comes from.

McCulloch took Walter into his family home near Chicago. When McCulloch's wife and three children went to bed, the two men poured whiskey and worked through the night. In 1943, they published their paper, which was overlooked by most neuroscientists for its simplicity. So, who did notice it?

That same year, Walter was introduced to a famous MIT mathematician, Norbert Wiener. Wiener led him to the blackboard and let him do his thing. By the second blackboard, Wiener was officially impressed. He decided to work more closely with Walter, promising him a PhD even though Walter had never even attended high school.

Another mathematician noticed too, one of the most celebrated of the century. At the time, the most advanced computer in the country filled a room of about 1,800 square feet, and it was closer to a giant electronic calculator than the computers we know today. Giving it a new job meant several operators spending weeks rerouting its wires and switches by hand.

In June 1945, this mathematician wrote the first published description of a better design. His idea was to write the instructions as information, in the same form as the data, and feed them into the machine the same way. The computer could then be given a new task without touching a wire, much like your phone changes jobs when you open a different app. To build it, he borrowed from Walter and McCulloch. Electronic switches, each either on or off, would act as the "and", "or" and "not" gates from their paper. Those instructions had to be held somewhere, so the machine also needed a memory. He proposed a component that stimulates itself, so a signal keeps circulating and the information stays put. That was Walter's looping idea from the brain paper. The entire report cited one paper: theirs.

The unravelling

In 1952, Wiener's wife, who disliked McCulloch, told her husband a damaging story about McCulloch's circle. According to the account I'm drawing on, it was invented. Wiener believed it. He cut Walter off, never spoke to him again and never told him why. For Walter, who had lost his own family at fifteen, this was the loss of a father figure, and it came with no explanation.

Then the science turned against him too. In 1959, Walter's colleagues ran experiments on frogs and found that the eye does a lot of processing before any signal reaches the brain. The brain was not a pure logic machine, which was the idea Walter had built his career on. Walter began drinking heavily and pulled away from his friends. He refused to sign the paperwork for his PhD, then burned his dissertation, all his notes and anything he had ever worked on. He died on 14 May 1969, alone and from complications of liver disease. McCulloch died four months later.

Like most geniuses gone too soon, Walter never had the chance to see what his work would turn out to be. Because while the people studying real brains had mostly set his model aside, and the dissertation he had spent years on was burned along with everything else, there were other people building something based on his research. The unit at the centre of the 1943 paper, a cell that adds up its inputs and fires once the total passes a threshold, had been picked up by people who weren't interested in brains, but machines. That unit contributed to the design of the first programmable and fully electric computer in 1945, and it has since become the basic building block of the neural networks behind the modern AI you use today.

(You can stop here for the cool story about another child prodigy, or continue to see how those building blocks operate in the AI you use.)

So how does AI actually answer your question?

Anyone who has used ChatGPT or Claude has probably wondered what is happening in the few seconds before the answer appears. Most people assume it is searching the internet or pulling a stored answer out of a giant database. Some chatbots can search the web as an extra feature, but the core of the system does neither. What it does is closer to what Walter and McCulloch described in 1943.

A chatbot is a program, a long set of instructions a computer follows, and inside it are huge numbers of copies of Walter's unit. These are now called artificial neurons, and each one is a tiny calculation. It takes in a handful of numbers, adds them up and passes a signal on to the next neuron, with a stronger signal for a bigger total.

Every connection between neurons has a knob that decides how much that signal counts. Turning a knob up makes that connection matter more, turning it down makes it matter less, and turning it below zero makes it push the other way. The technical name for these knobs is weights, or parameters. GPT-3, a model from 2020 whose details were published, stored everything it learned in 175 billion of them, roughly twenty for every person on Earth.

How are the knobs set?

Nobody set those knobs by hand. They start out random and are adjusted during learning, which fits Walter's guess that a brain starts roughly random and is tuned over time.

The learning, called training, is a single game repeated an enormous number of times. The model is shown some text with the next word hidden, such as "the cat sat on the", and it has to guess the word. Early guesses are nonsense. After each guess, it is shown the real word, and every knob is nudged a tiny amount in whichever direction would have made that guess slightly better. The text is fed in as tokens, which are words or pieces of words. GPT-3 played this game across around 300 billion tokens of text from the internet and books. Claude models begin the same way, trained to predict the next word across large amounts of text.

This is the answer to the question. The model mostly does not keep copies of what it read, although it can memorise passages that appeared many times. What it keeps are the knob settings that got it better at guessing. The difference is like two people answering a question about Paris. One reads you the entry from an encyclopaedia, and the other has read about Paris for years and answers from what has stuck. A chatbot is the second person, with an enormous amount of reading behind it and a memory that exists only as billions of numbers.

What happens when you press enter?

Your question is chopped into tokens. The tokens are turned into numbers, and those numbers pass through the neurons in stages called layers, like stations on an assembly line where each one refines what the previous one produced. GPT-3 had 96 of them.

What comes out at the other end is not an answer, but a list of odds for the next token. For example, a 60 per cent chance of "the" and a 15 per cent chance of "a". The chatbot picks one, adds it to the text, and runs the whole process again. The reply is built one small piece at a time, which is why the words appear on your screen as a stream.

Why is it sometimes confidently wrong?

This also explains the most frustrating thing about chatbots. Nothing in that process checks whether a sentence is true. The model produces the continuation that its knob settings rate as likely, and a plausible but false sentence can score well. That is how it can invent a quote or a book that sounds exactly right, and the industry term for this is hallucination.

Teaching it to be an assistant

A model trained only on the guessing game can continue any text, but it has no idea what a good answer looks like. A second stage, called fine-tuning, keeps adjusting the same knobs towards helpful behaviour. The standard method is reinforcement learning from human feedback, where people rate answers and the model is rewarded for producing the kind they prefer, much like training a dog with treats.

Anthropic, the company that makes Claude, added a method it calls Constitutional AI. The model critiques and revises its own answers against a written set of principles, and an AI judge using those same principles picks between answers to guide the reward step, in place of most human ratings. One aim of this stage is to make the model more honest about what it doesn't know, which reduces hallucination but does not eliminate it, because the underlying machine is still predicting and not checking.

Where the knowledge actually is

If you asked where inside Claude the fact "Paris is the capital of France" is stored, nobody could point to it. In Walter's networks, each neuron had a job he designed, so you could say exactly what it did. In a trained model, what it knows is spread across billions of knobs set by the guessing game. Researchers are still developing ways to read them, in a field called interpretability.

So…

That brings the story back to where it started. In 1943, a runaway teenager and a neurologist argued that knowledge could live in the connections between simple units that add up signals and fire, with no single place where any one idea sits. They were trying to explain the brain. Eighty years later, when you ask a chatbot a question, the honest answer to "how does it know that?" is roughly the answer they gave.