Hitting Walls

I

n the winter of 1943, while the world burned in the fires of the Second World War, two men sat in a Chicago apartment and tried to reduce the human soul to a circuit diagram.

Walter Pitts was a homeless runaway with a genius for logic that frightened his professors. Warren McCulloch was an older, established neurophysiologist who believed that the brain was a machine made of meat. Together, they were an impossible pair—the vagrant and the patrician—united by a single, heretical idea: that thought was computation.

They didn't have computers. The ENIAC was still a classified fever dream in Pennsylvania. They didn't have data. They didn't have funding. All they had was a pencil, paper, and a conviction that the neurons in our heads were nothing more than biological logic gates.

"If a neuron can fire or not fire," Pitts argued, his voice thin and reedy, "then it is a proposition. It is True or False. And if it is True or False, then the brain is not a mystical organ. It is a logic machine."

This was the first wall. The wall of complexity. To scale it, they had to strip the brain of its chemistry, its hormones, its messy biological reality, and flatten it into a math problem. The paper they wrote that winter, A Logical Calculus of the Ideas Immanent in Nervous Activity, was dense, impenetrable, and largely wrong in its biological details.

It was also the spark that would eventually burn down the entire field of psychology and replace it with cognitive science. It was the blueprint for the first artificial neuron. It was the beginning of the end of the idea that the human mind was magic.

We like to tell the history of science as a story of inevitable brilliance. We imagine it as a clean ascent. A straight line. A sequence of geniuses handing each other the same torch. We sanitize the errors, delete the dead ends, and present the final result as the only possible outcome.

It is the classic error of the survivor. We look at the planes that returned from the war, count the bullet holes in the wings, and decide to armor the wings. We forget that the planes hit in the engine never came back to be counted.

In AI, we study the papers that got cited. We ignore the thousand architectures that were hit in the engine.

The public story of Artificial Intelligence allows for no such comfort. It moved like a speeding car in fog. It found progress by impact. It discovered its limits by breaking on them. It is, in its purest form, a biography of wreckage.

This book is about the missing planes.

The field did not “advance” so much as it collided. A model worked in a narrow room—clean digits on paper, a toy maze, a curated dataset that behaves—and then failed at the door. The failure was not a footnote. It was the plot.

Think of the great Gothic cathedrals. The reason they soared so high is that the early ones fell down. The collapsing nave at Beauvais wasn't a disgrace; it was the only way the masons could learn the crushing weight of stone. The collapse revealed the stress.

Each crash left wreckage: half-true theories, discarded architectures, abandoned funding, shuttered labs, conference rooms that stop taking you seriously. Then someone built a ladder from the scrap and climbed anyway.

That is not a metaphor for drama. It is the mechanism of the field.

Every generation of Artificial Intelligence believes it is approaching a final wall.

In the 1960s, neural networks were declared biologically implausible. In the 1970s, learning machines were pronounced mathematically insufficient. In the 1980s, expert systems collapsed under their own brittleness. In the 1990s, statistical learning was dismissed as shallow pattern matching. Each moment arrived with confident proclamations that the path ahead was blocked—that intelligence had revealed its limits.

Not all of these moments were the same kind of limit. Some were winters of funding and fashion. The three collisions this book treats as pillars are the ones where the physics of the architecture itself forced the field to reroute.

Today, that wall has a new vocabulary: large language models, scaling limits, hallucinations, truth, energy, alignment. The words change. The anxiety does not. The field has been here before—many times.

And each time, those proclamations were both right and wrong.

They were right about the wall. They were wrong about what walls mean.

A wall is not evidence that a field is failing. In Artificial Intelligence, it is evidence that the field is alive.

A dead end fails everywhere. A wall is something that works—until you ask it to cross a boundary it was never built to cross.

(Excerpt from the Prologue)