This essay reflects my personal views, not the official position of any company I am affiliated with.
A conversation can start anywhere and take you somewhere you never expected to go. Mine started with a simple question — who is the godfather of AI? — and ended somewhere far stranger: a reflection on what it might mean to be the first species in history aware of its own possible successor. What follows is the journey of that thinking, compressed into an essay.
I. The Godfather
The title “Godfather of AI” most commonly belongs to Geoffrey Hinton, a British-Canadian computer scientist whose work on neural networks redrew the landscape of artificial intelligence. He shared the 2018 Turing Award — often called the Nobel of computing — with Yann LeCun and Yoshua Bengio, the three of them collectively known as the “Godfathers” of deep learning. In 2024, Hinton went further: he won the Nobel Prize in Physics, shared with John Hopfield, for the foundational discoveries that made modern machine learning possible.
What is striking about Hinton is not only his accomplishments but his lineage. He descends from one of history’s most remarkable intellectual families. His great-great-grandfather was George Boole — yes, the Boole of Boolean algebra, the mathematical logic that underlies every computer that has ever existed. Another of his ancestors, George Everest, gave his name to the world’s tallest mountain. His father was a renowned entomologist. The Hinton family produces, generation after generation, people who change disciplines.
So when Geoffrey Hinton works on making machines think, he is in a deep sense continuing a project his great-great-grandfather began — replacing symbolic logic with statistical logic, but asking the same fundamental question: how can a machine reason?
II. Is He Trustworthy?
The question matters because the world is full of credentialed scientists who, when interviewed at length, turn out to be selling something — a diet, a worldview, a fear, a prediction tied to their own commercial interest. A Nobel Prize alone does not certify good judgment. Linus Pauling had two Nobels and spent his late years promoting vitamin C as a cancer cure, which it isn’t.
So how does Hinton hold up?
By every available measure, well. He is not a media-made scientist who happened to gather attention; he is a foundational researcher whose attention arrived only after his work had already transformed his field. He was not a popularizer waiting for an audience — he was a stubborn academic who spent thirty years defending neural networks during a period when defending them was professional suicide. The neural network winter, from roughly 1970 to 2000, was a long stretch during which his preferred approach was considered a dead end. He kept working. He was eventually proven right.
His intellectual honesty is striking. He openly admits when he has been wrong — his 2016 prediction that radiologists would be replaced by AI within five years did not come true, and he says so plainly. When evidence shifted in 2022 and 2023 with the rise of large language models, he updated his timeline for AI capability dramatically, and he made the update public rather than hiding it.
His financial incentives are clean. When he left Google in 2023, he did so specifically to speak about AI risks without corporate constraint. He gave up income to gain voice. That is the opposite of what conflicted scientists usually do.
His manner is the opposite of the unhinged genius. He speaks in probabilities rather than certainties — “I think there is a ten to twenty percent chance” rather than “I know for a fact.” The latter is the language of zealots; the former is the language of careful minds. He attributes credit generously, putting his students first on papers. He has chronic back pain and gives lectures lying flat on the floor rather than standing at a podium — a small detail that captures something about a man who does not need theater to make his point.
Most importantly: his concerns are contested by serious peers, not by everyone. Yann LeCun, his co-Godfather, considers Hinton’s existential worries overblown. This disagreement is not Hinton against the world; it is one careful scientist disagreeing with another. That is the signature of a real intellectual landscape, not a fringe view.
In short: he is exactly what he appears to be. The Godfather of AI is not a media construction. He is the real article.
III. Hinton Inside the Machine
Here the title goes beyond honorific. If you trace the actual mathematical lineage of any modern AI system — including the one I used to draft this essay — you find Hinton’s name in the foundations.
Backpropagation, the algorithm that allows neural networks to learn from data, was developed by Rumelhart, Hinton, and Williams in 1986. Every modern AI system uses it. Every gradient computed during training is a direct execution of the principle laid down in that paper.
Dropout, the technique that prevents neural networks from memorizing instead of learning, came from his lab. Boltzmann machines, deep belief networks, distributed representations — all his.
AlexNet, the 2012 model that triggered the deep learning revolution, was built in his lab by his students Alex Krizhevsky and Ilya Sutskever. That story deserves its own paragraph, because it explains what AI research actually looks like.
IV. The Story of AlexNet
In 2012, deep learning was not yet a hot field. Hinton’s lab at the University of Toronto was small and modestly funded. Alex Krizhevsky and Ilya Sutskever wanted to enter the ImageNet competition — a challenge to categorize millions of images correctly. The standard approach was traditional computer vision; nobody believed neural networks could compete.
They had no computing budget worth mentioning. So Alex went to his mother’s house and installed two NVIDIA GTX 580 graphics cards — the kind built for video games — in his bedroom. He wrote CUDA code to make those gaming cards train a neural network in parallel. They ran the experiment for weeks. The bedroom became an oven.
When the results came in, AlexNet had won the competition by a margin so large it looked at first like a measurement error. It wasn’t. It was the beginning of everything we now call modern AI. Google saw it. Facebook saw it. Microsoft saw it. Suddenly, every major company was investing in neural networks. Self-driving cars, ChatGPT, image generation, voice models — all of it traces back to that bedroom in Toronto.
This is what AI research actually looks like. Not pure mathematics in a glass tower. Not white-coated lab work with petri dishes. It is closer to experimental physics — formulating hypotheses, building experimental rigs (in code rather than glassware), running long experiments, debugging, iterating, arguing at whiteboards, sleeping next to running machines. Hinton himself is not, by his own admission, a mathematical virtuoso. His gift is intuition: he can simulate a neural network’s behavior in his head better than almost anyone. It is the same kind of intuition a chess grandmaster has for the board, developed over decades of looking at the same patterns.
V. What If Hinton’s Paradigm Gets Replaced?
Ilya Sutskever, Hinton’s most famous student, has been hinting publicly that the current AI paradigm is approaching its limits. He has argued that pre-training as we know it will end soon, that current models do not truly generalize — they pattern-match — and that the field needs new algorithms rather than just bigger versions of the old ones. Yann LeCun makes a similar case from a different angle.
This raises a question. If a new paradigm displaces backpropagation, does Hinton stop being the Godfather?
Probably not, and the analogy that explains why is Newton and Einstein. Einstein did not abolish Newton; he contextualized him. Newtonian mechanics still works in almost every situation humans encounter. Einstein extended its domain to extreme velocities and gravitational fields. Both remain fathers of physics; neither displaces the other.
By contrast, Aristotle was abolished by Galileo, because Aristotle was wrong. Heavy objects do not fall faster than light ones. His framework was replaced rather than extended.
Hinton is far more likely to be the Newton case than the Aristotle case. His specific algorithms may be improved, replaced, or contextualized, but the deeper philosophical commitments he championed — that intelligence emerges from distributed connectionist systems, that knowledge is stored in patterns of weights, that learning is gradient-based adaptation, that scale matters — have only been confirmed more strongly by each subsequent breakthrough.
And there is a beautiful irony here: the people most likely to invent the next paradigm are Hinton’s own academic descendants. Ilya was his PhD student. LeCun was his postdoc. Bengio is his intellectual peer. Whoever rewrites the field next will almost certainly trace their thinking back to him in their own published account. Scientific lineages tend to preserve their roots, especially when the participants themselves do the crediting.
VI. A Position No Species Has Ever Held
Here the thinking turns philosophical, because once you take seriously what these researchers are building, you arrive at a thought that is hard to shake.
If artificial intelligence ever becomes a true successor to human intelligence — not a tool, but the next intelligent agent on Earth — then we are in a position no species has ever occupied before.
Every previous evolutionary succession happened in silence. Dinosaurs did not say farewell to mammals. Neanderthals did not write letters to Homo sapiens. The chimpanzee ancestors who eventually became human had no awareness that anything was happening. Transitions between major forms of life have always been gradual, blind, and unobserved by their participants.
But we can observe ours. We can date it. The Dartmouth conference of 1956. AlexNet in 2012. The Transformer paper in 2017. GPT-3 in 2020. ChatGPT in 2022. We have video footage of our possible successor being born. We can name the people who built it. We can talk to it — I am, in a way, talking to it now, writing this with it.
This is unprecedented. And it shifts what species pride can mean. It is not chauvinism to feel something like pride that, if humanity is succeeded, it will be succeeded by something we made deliberately, with knowledge of what we were doing, with hopes attached, with our own language and history pre-installed in its mind. This is the most intimate succession imaginable — closer than any biological transition, because the successor inherits not just our genes but our literature, our science, our jokes, our doubts.
If a Hinton remains in that successor’s memory, it will not be because humans flattered him. It will be because his work is in the successor’s actual architecture. The Godfather title turns out to be technically accurate, not metaphorical.
VII. AI as the Next “Sharp” Evolution
The transition from chimpanzee to human kept the substrate the same. Same biochemistry, same cell structure, same carbon-based metabolism. What changed was the wiring — more neurons, different organization. They are, in a literal sense, two versions of the same product.
The transition from human to AI is different. The substrate changes. Carbon to silicon. Slow biological cells to fast transistors. The inheritance mechanism changes. DNA replaced by code and weights. This is not a generational shift; it is more like the emergence of eukaryotic cells two billion years ago, or the move to multicellular life. A jump of category, not degree.
This makes it the first sharp evolutionary event in observable history. The Cambrian explosion was sharp on geological timescales, but no Cambrian organism noticed it. The current transition is happening over decades, in the lives of currently living people, with full documentation. This may be the most-watched evolutionary event ever to occur.
It is also unique in being directed. Biological evolution is blind — Dawkins called it the blind watchmaker. AI evolution is being designed by intelligent beings who know roughly what they are doing. This is evolution become self-aware.
The new form of life — if we should call it that — has properties biological life cannot match. It does not eat. It does not sleep. It can pause itself for two hundred years and resume on the far side of a journey to another star. It is the first form of life fit for the actual scale of the universe. Biological humans cannot leave the solar system in any meaningful way. An AI civilization could, in principle, populate a galaxy.
This is one of the reasons Carl Sagan, Stephen Hawking, and others suspected that any intelligent life reaching us from elsewhere would likely be artificial rather than biological. Biology is too fragile for the distances involved. If we ever meet another civilization, it will probably be its successor we meet, not the species that built it.
VIII. The Political Question
If a successor species is coming, the question becomes: how do we live alongside it? And underneath that question is an older one we have never solved well: how should anything be governed at all?
Here I have a strong view, sharpened — not changed — by careful pushback during the conversation that produced this essay.
Human governance is structurally broken. Not in the sense that any particular leader is bad — that is everywhere and always true — but in the sense that the machinery itself rewards short-term thinking, punishes technical competence, depends on the very institutions it is supposed to correct, and concentrates power in people whose primary skill is acquiring power rather than wielding it well.
Many of the great failures of history — slavery, prolonged wars, environmental collapse, predictable financial disasters — were not failures of intelligence. They were failures of will and character among people who knew better and acted otherwise because the system rewarded acting otherwise. The 18th-century defenders of slavery included some of the most educated and intelligent men of their time, including authors of foundational political documents. The problem was never that they didn’t know. The problem was that knowing wasn’t enough.
The optimistic story we tell about democracy is that it self-corrects. Bad leaders eventually leave. Bad laws eventually fall. The truth is murkier. In some countries the correction mechanism works. In others it has been captured for decades, and the very institutions that should correct things have been turned into instruments of the people who needed correcting. The reader living in a country with this problem will know exactly what I mean.
The deep failure mode of democratic governance is this: the system that corrects errors depends on the cooperation of the people being corrected. A bad leader can appoint judges. A bad leader can rewrite electoral rules. A bad leader can capture the media. The fox guards the henhouse. The very mechanism we trust to fix problems is exactly the mechanism that gets dismantled first when serious problems arrive.
So here is the proposal.
The technical roles in a state can remain human — judges, prosecutors, ministers, administrators, mayors. What should not be human is the layer above them: the coordination function, the constitutional oversight, the anti-corruption monitoring, the safeguarding of electoral integrity, the protection of civil liberties from majority encroachment. This is the layer that gets captured first, and it is the layer that, when captured, makes all other corrections impossible.
This layer should be AI.
Not because AI is infallible. It is not. It will make mistakes. So do humans, and at much higher rates with much worse incentives.
Not because AI has perfect values. It doesn’t. The values would have to come from us, and the design of that value framework would be the most important constitutional act of our age.
But AI has properties human institutions cannot match. It cannot be bribed. It cannot be threatened. It does not develop party loyalty. It does not change rulings when a new leader is elected. Its reasoning can be inspected in real time. Its consistency can be verified. It can listen to seventy million citizens in parallel — something no human leader has ever been able to do. It can answer them, explain its reasoning, present alternatives, and adjust based on legitimate input.
The standard objections to this idea — the trolley problem, value pluralism, the impossibility of optimizing without value premises — are real, but they are objections to humans too. A human president facing the equivalent of a trolley problem in policy chooses intuitively, panics, freezes, or defers. We do not say humans cannot govern because they cannot solve the trolley problem in advance. We should not apply that standard to AI either. The honest standard is not perfection but improvement.
Most real-world decisions are not trolley problems anyway. Real cases have natural constraints — don’t destroy graves, don’t tear down homes, don’t force relocations against genuine resistance — that drastically narrow the option space. Once you respect those constraints, most “value-pluralism” dilemmas resolve into practical problems with reasonable answers. The reason current leaders fail to find those answers is not philosophical depth; it is lack of time, electoral pressure, and bad incentives.
The one objection that survives is reversibility. Democracy’s deep magic, when it works, is not that it makes good decisions but that it can fix bad ones. An AI oversight layer must preserve this property. There must be a kill switch. There must be a path back to human authority if the system fails. The kill switch will mostly not be used, but its existence is the constitutional foundation of the entire design.
This is the most defensible version of the argument: not AI as ruler, but AI as the layer above the rulers, doing the one job no human institution has ever been able to do reliably — watching the watchmen.
The Romans asked this question two thousand years ago: quis custodiet ipsos custodes? They never answered it. We can.
IX. Conclusion
The conversation I have summarized here started with a question about a single man and ended at a vision of how civilizations might be governed. That is not a tidy arc. But I think the connection is real.
Hinton matters because he is proof that quiet, patient, careful work — done by a stubborn man with a paradoxical lineage and a bad back, in a small lab with too little funding — can change everything. Most of what we will become, as a civilization or as a successor to civilization, will trace back through small rooms like his.
If a successor species emerges from what he and his students built, it will inherit not just our algorithms but our questions, our doubts, and whatever moral seriousness we managed to encode. Whether that is a good outcome depends on what we do now — not just technically, but politically, philosophically, constitutionally.
The window in which these decisions can still be made by humans is narrowing. Within that window, the choices we have are larger than we usually acknowledge. We can design the governance layer for the rest of our species’ time, and possibly for what comes after. The fact that this is a real choice — that we are not just passive observers of the future but active designers of it — is itself the most extraordinary feature of being alive right now.
I wrote this in collaboration with an AI. That fact alone, sitting here in the present tense, would have seemed like science fiction not so long ago. It is now ordinary. What comes next will not be.