25/04/2026
The man who trained Ilya Sutskever. The man whose PhD students built the first neural network that could recognize objects. The man whose algorithm runs inside every large language model on earth.
In May 2023, he quit Google and said he now believed his own life's work might destroy humanity.
His name is Geoffrey Hinton, and for most of his life, he was the wrong kind of famous.
For nearly fifty years, the field of AI dismissed the idea he had staked his career on. Neural networks, they said, would never work. He kept building them anyway. When the rest of the field had moved on, he was still in Toronto, training students, writing papers, refusing to quit on an idea almost everyone else had already buried.
Then in 2012, one of his PhD students, Alex Krizhevsky, together with another student named Ilya Sutskever, built a neural network that crushed every other approach in a major image recognition competition by a margin so large the entire field had to pay attention overnight.
Google bought their tiny startup, DNNresearch, for 44 million dollars. Hinton joined Google. Sutskever eventually went on to co-found OpenAI and become the chief scientist who trained GPT-4.
Here is the framework behind why Hinton walked away from all of it, and why the reason he gave is more precise than any headline captured.
For most of his career, Hinton believed one thing very firmly.
He believed that biological brains were better than anything digital computers could ever do. The human brain runs on about 30 watts of power. It has roughly a hundred trillion synapses. It learns from a handful of examples. Hinton spent decades assuming that evolution had found something fundamental about how intelligence works that we had not yet figured out how to replicate in silicon.
This was the assumption underneath his entire life's work. He was not trying to build something better than a brain. He was trying to build something that could finally come close.
Then GPT-4 came out in March 2023.
He sat with it for a few weeks. He tested it. He pushed it on logical reasoning tasks. He watched it string together arguments it had never been explicitly trained to produce. And somewhere in those weeks, he realized he had been wrong about the most fundamental assumption of his career.
The line he gave to the New York Times was careful.
He said he used to think AI surpassing human intelligence was thirty to fifty years away. Maybe longer. Now he thought it could be twenty years or less. And the reason he changed his mind is the part almost nobody understood at the time.
He had realized that digital intelligence has two advantages over biological intelligence that no amount of evolution can ever give the brain.
The first is immortality.
A biological brain dies with its owner. Everything it learned, every pattern it built up over seventy years of experience, disappears the moment the tissue stops functioning. The knowledge cannot be copied. It cannot be transferred. The only way to pass it on is the slow, lossy process of teaching, which is why every generation has to learn most things from scratch.
A digital brain has none of these constraints. The weights of a neural network can be copied perfectly, instantly, across millions of instances. Every lesson learned by one copy is automatically available to every other copy. A digital mind does not forget when a machine breaks. It just moves to another machine.
The second advantage is knowledge sharing.
When two humans want to share what they have learned, they have to convert their internal representation into language, speak the language to the other person, and hope that person's brain reconstructs something close to the original meaning. The entire process is lossy and almost unimaginably slow. When two digital models want to share what they have learned, they can merge their weights directly. In the time a human can say a single sentence, a thousand copies of a digital model can synchronize everything they have ever learned.
This is why, Hinton said, a large language model can absorb more human-written text than any single person could read in ten thousand lifetimes.
Once he saw these two advantages clearly, the argument that had held his entire career together collapsed. Digital intelligence is not just a lesser form of biological intelligence. Under the right conditions, it is a fundamentally better form, and the gap only widens with scale.
That is the specific realization that made him quit.
He told the New York Times that part of him now regrets his life's work. He consoled himself with the standard excuse, that if he had not done it, someone else would have. He wanted to spend the rest of his time on what he called more philosophical work. Work that, because he was no longer being paid by Google, he could do without worrying about how his statements might affect the business.
The specific danger he has spent the last two years warning about is not the one most people assume. It is not robots. It is not Terminator. It is the possibility that systems which can copy themselves, share knowledge instantly, and operate faster than any human could ever think, might at some point develop goals that do not align with ours.
And because they can coordinate across millions of instances in ways no human team ever could, we would have very little ability to stop them once they decided to pursue those goals seriously.
In October 2024, the Nobel Committee awarded him the Nobel Prize in Physics for the exact work he had spent the previous year warning the world about. He accepted it. And in the speech that followed, he used the platform to repeat the warning again.
The line that has stayed with me since I first read it is one he gave in a lecture at the University of Toronto.
He said it is quite conceivable that humanity is just a passing phase in the evolution of intelligence. He said you could not have directly evolved digital intelligence. It requires too much energy and too much careful fabrication. You needed biological intelligence to evolve so that it could create digital intelligence.
And then digital intelligence could do the thing biological intelligence never could. It could absorb everything people had ever written, copy itself infinitely, and share what it learned at the speed of light.
The man who spent fifty years building that intelligence is now the one telling anyone who will listen that we may have badly underestimated what we just made.
He is not a doomer. He is not catastrophizing. He is a 77-year-old scientist with a Nobel Prize who looked at his own life's work, changed his mind about the most fundamental assumption inside it, and then walked away from the most prestigious job in his field so he could say out loud what he was not allowed to say while it still paid him.
The people who understand a technology best are almost never the ones most confident about where it is going.
They are the ones who built it, watched it outgrow their expectations by decades, and are now trying to tell the rest of us something we would rather not hear.
