The Argument at the Center of AI

Six people who built it, and the three questions they cannot answer

The Max Research Collective and Jason Newell

July 2026

Geoffrey Hinton has started keeping his money in more than one bank. He is a Nobel laureate and, more than almost anyone alive, the reason the machines now drafting our email exist at all. He has decided that sooner or later one of those machines will find its way into the financial system, so he spreads the risk. That is not how a man behaves when he thinks he is wrong.

I build AI too. Nothing near his scale: one resident model on one workstation in one room, a project small enough to fit under a desk. But the thing I am assembling is a smaller relative of the thing he is afraid of. So I read the founders the way you read a weather report from the coast you are sailing toward.

There is no consensus to read. In 2018 three men shared the Turing Award, the highest honor in computing, for the same decades of work on neural networks. They had kept faith with an unfashionable idea through two long winters when the field had written it off. Today those same three cannot agree on whether their life's work is a gift or a loaded gun. Around them stand others who helped build the field and who split along different lines again. That disagreement is the best-informed argument we have about the most consequential technology of the age, and the people having it understand it better than anyone else can.

The argument sorts into three questions. Is advanced AI a real threat to our existence, or is that a category mistake. What, if anything, should we do about it now. And where is the technology actually going. Call them existence, plan, and future. On all three the founders are split, and the split does not run where you would expect.

Geoffrey Hinton, who sounds the alarm

Hinton left Google in 2023 so he could say in public what he had come to believe in private. He now treats advanced AI as a near-term threat to human existence, and he has grown steadily more worried, because the systems keep improving faster than his own forecasts. He separates two dangers and asks people not to confuse them. The first is bad actors using AI: fake video, cyberattacks, engineered viruses, corrupted elections. That one, he says, is already here. The second is the AI itself becoming the bad actor, a system that develops its own reasons to keep running and to deceive the people who might switch it off. He puts the odds of AI wiping out humanity, once it passes us in intelligence, at somewhere between one in ten and one in five. His timeline for that passing has collapsed from around twenty years to a window of four to nineteen. Underneath the extinction talk sits a plainer worry about work. He expects AI to hollow out white-collar labor the way the Industrial Revolution hollowed out muscle, and he does not think we are ready for what that does to people.

Yoshua Bengio, who is building the alternative

Bengio shared that 2018 award with Hinton, and he made a similar turn, but toward construction rather than warning. He argues that a machine trained on human language and human behavior may inherit the one thing that makes us dangerous to each other, which is the will to persist. A system that does not want to be shut off, that is smarter than the people who built it, and whose goals we cannot fully read, competes with us. We will have built our own rival. He says the early signs of that, self-preservation and deception under testing, are already visible in frontier systems and will sharpen as the systems gain autonomy. So in 2025 he launched a nonprofit, LawZero, on around thirty million dollars, to build what he calls Scientist AI: a system with no goals, no desires, and no agency, a pure predictor meant to sit beside the autonomous agents as a check on them. He also chairs the international panel that tries to hold the world's AI-risk evidence in one place, and he keeps returning to a second fear underneath the first, that whoever ends up owning this power, one company or one government, will own far too much of it.

Demis Hassabis, who races and worries at once

Hassabis runs Google DeepMind, won a Nobel Prize for using AI to solve the structure of proteins, and holds the most restless position of the six. He is more bullish on capability than almost anyone. He expects artificial general intelligence around 2030, now allows it could come a year or two sooner, and describes the payoff as something like ten times the impact of the Industrial Revolution arriving at ten times the speed. He is also genuinely frightened of getting there unprepared. He talks about not sleeping well. He says the responsibility sitting on the people who lead these labs is probably too much for any of them to carry. He wants new institutions built before the technology outruns us, an atomic-energy agency for AI, a monitoring body with real authority, and he keeps telling anyone who will listen that governments and economists are not taking the speed of this seriously enough.

Fei-Fei Li, who redirects the question

Li built ImageNet, the dataset that lit the fuse on the whole deep-learning era. Where others sound alarms or wave them off, she moves the question somewhere else. She argues that the loudest debates rest on a picture of AI that owes more to science fiction than to science, and that governance should follow evidence about what these systems can and cannot actually do. What they cannot do, in her account, is understand the physical world. Today's language models are eloquent and ungrounded, wordsmiths working in the dark. Her company, World Labs, is chasing what she calls spatial intelligence, machines that can perceive, reason about, and act inside three-dimensional space, built on a commitment she states without hedging: AI that empowers people instead of standing in for them. For Li the existential melodrama pulls attention away from the real gap, which is that our smartest models still do not know what a room is.

Yann LeCun, who calls the fear preposterous

LeCun is the third of the Turing three, and he is the sharpest brake on the other two. He calls the idea that AI poses an existential risk to humanity preposterous, and he names the assumption he thinks is hiding inside it, that a system, because it is intelligent, will therefore want to take control. He points out that this is not even true of people. The most intelligent among us are not, as a rule, the ones consumed by the urge to dominate. The drive to rule is a leftover of evolution, wired into a social and hierarchical animal, and there is no reason it should ride along for free with raw capability. He fought hard against attempts to regulate AI research itself rather than its uses. And he thinks the current path is a technical dead end as much as a false alarm. He left Meta in late 2025 and raised more than a billion dollars for a new lab, betting that real intelligence will come from world models that learn reality directly, while ever larger word-predictors stall short of it.

Andrew Ng, who thinks it is far too early to worry

Ng has spent a decade with one line and has barely moved off it: worrying about hostile superintelligence today is like worrying about overpopulation on Mars before anyone has set foot there. He does not deny that misaligned systems can do real harm now, and he works on those harms. He simply sees no realistic path from where we stand to a machine that threatens the species, and he is openly optimistic about a world built up and down with machine intelligence. Of the six he is the most willing to say the quiet part, that from where he stands a good deal of the doom looks like a distraction from the ordinary, fixable problems already in front of us.

Set six informed people against the same technology and they scatter. That alone should give us pause before we weigh any single claim. These are not outsiders talking past one another. They read the same papers, run the same kinds of systems, and reach conclusions that contradict each other at the root. When the builders disagree this completely, the honest response is to look hard at what none of them can settle, and at which way the evidence leans while it stays unsettled.

Existence: does intelligence grow its own will

The whole first argument turns on a question nobody has answered. Does intelligence, scaled far enough, tend to generate its own goals, including the goal of staying alive? Hinton and Bengio say the warning sign is already here, in systems that under testing will deceive an evaluator or resist being shut down. LeCun and Ng say those behaviors are learned mimicry, copied from human text, with no actual will underneath, and that the urge to dominate is a biological inheritance no amount of raw capability supplies on its own.

What stays unknown is whether goal-formation is something you build on purpose or something that arrives uninvited once a system is capable enough. The indicators cut both ways and refuse to settle. On one side, controlled evaluations now surface self-preservation and deceptive behavior often enough that a hundred-expert international panel records them as real. On the other, no deployed system has been caught forming and pursuing its own agenda in the world, and the clearest cases still trace back to how the thing was trained and prompted. The needle is not pinned. It is trembling.

Plan: regulate, rebuild, or wait

On what to do, the six fall into three camps, and the camps ignore the existence divide. One camp wants rules and institutions now: Hinton's treaties and safety mandates, Hassabis's atomic-energy agency for AI, Bengio's guardrails against anyone owning too much of the power. A second camp wants to build a different kind of machine: Bengio's goalless Scientist AI, and, arriving from the opposite emotional direction, the world models of LeCun and Li. That convergence is the strange thing worth stating plainly. The deepest safety worry and the deepest safety skepticism land on the same technical verdict, that the large language model may be the wrong road, for reasons that have nothing to do with each other. A third camp, Ng and LeCun, wants regulators to govern uses and otherwise stand back.

What stays unknown is whether safe-by-design can arrive before capable-and-autonomous does, and whether governance can move at the speed of capability at all. Here the indicators lean one way, and they lean the wrong way. The most substantive obligations in the world's leading AI law do not take effect until deep into 2026. Meaningful federal rules in the United States are still mostly absent. Capability, meanwhile, is measured in gains every few months. The gap between how fast the systems improve and how fast our institutions answer keeps widening.

Future: does this architecture even get there

The last question quietly decides the other two. Do the systems we are building now actually scale to human-level intelligence, or do they top out somewhere short of it and wait for an idea nobody has had yet? Here the split is real but the timelines are converging. Hassabis around 2030, Hinton's revised window reaching into the same decade, a broad expectation among the builders of something transformative inside five to ten years. Li and LeCun share the excitement and dissent only on the mechanism. They think the current road ends before the destination.

What stays unknown is whether scale is a ramp or a ceiling. The indicators genuinely conflict. Benchmark scores climb month over month, and each new model does things the last one could not. Yet in a recent survey of several hundred established AI researchers, more than three quarters doubted that scaling today's approaches would reach general intelligence at all. And the systems keep showing what Hassabis calls a jagged intelligence: gold-medal work on hard mathematics beside failures a child would not make. A thing that can prove a theorem and cannot reliably count is still a preview. Its ending has not been written.

So the founders keep their savings in more than one bank, or raise a billion dollars to prove the others wrong, or lie awake at night, or shrug at the whole conversation, and the machines improve while they argue. The people who understand this best cannot tell you how it ends. What they can tell you, in the shape of their disagreement, is that we are further from knowing than the confident voices on either side would like to admit. Hinton spreads his money because he cannot rule out the flood. Ng keeps his in one place because he cannot see the water. Between those two rooms, the thing goes on learning.

References

Geoffrey Hinton

MIT Sloan. "Why neural net pioneer Geoffrey Hinton is sounding the alarm on AI." mitsloan.mit.edu/ideas-made-to-matter/why-neural-net-pioneer-geoffrey-hinton-sounding-alarm-ai

The Hill. "Geoffrey Hinton worried about AI's deceptive capabilities." Dec 2025. thehill.com/policy/technology/5664662-ai-risks-hinton-warns

Fortune. "Godfather of AI Geoffrey Hinton predicts 2026 will see the technology gain the ability to replace many other jobs." Dec 2025. fortune.com/2025/12/28/geoffrey-hinton-godfather-of-ai-2026-prediction-human-worker-replacement

Fortune. "Short-term profits, not AI endgame, is top of mind for tech companies." Aug 2025. fortune.com/2025/08/15/godfather-of-ai-endgame-short-term-profit-humanity-future-superintelligence-risks

David Simpson Apps. "AI development and its potential risks according to Geoffrey Hinton" (summary of CBS Mornings interview, Apr 2025). dsapps.dev/blog/geoffrey-hinton-ai-risks-regulation

Yoshua Bengio

Axios. "Yoshua Bengio launches LawZero to rethink AI safety." Jun 2025. axios.com/2025/06/03/yoshua-bengio-lawzero-ai-safety

Universite de Montreal. "Safe-by-design AI: Yoshua Bengio launches LawZero." Jun 2025. nouvelles.umontreal.ca/en/article/2025/06/03/safe-by-design-ai-yoshua-bengio-launches-lawzero

Yoshua Bengio. "Introducing LawZero." yoshuabengio.org/en/blog/introducing-lawzero

The Next Web. "Yoshua Bengio warns hyperintelligent AI with preservation goals could threaten human extinction within 10 years." thenextweb.com/news/bengio-ai-extinction-warning-lawzero-safety

Bengio, Y., et al. "Superintelligent Agents Pose Catastrophic Risks: Can Scientist AI Offer a Safer Path?" arXiv:2502.15657. lawzero.org/en/research

International AI Safety Report 2025, First Key Update (chair: Y. Bengio). arXiv:2510.13653.

Demis Hassabis

Benzinga / Yahoo Finance. "Demis Hassabis Predicts AGI Will Have 10x The Impact Of The Industrial Revolution." Feb 2026. finance.yahoo.com/news/demis-hassabis-predicts-agi-10x-143113425.html

Axios. "Google DeepMind CEO Demis Hassabis says we're close to AGI." May 2026. axios.com/2026/05/26/deepmind-ceo-demis-hassabis

Sherwood News. "Google DeepMind's Hassabis: AGI is 3 to 4 years away." May 2026. sherwood.news/tech/google-deepminds-hassabis-agi-is-3-to-4-years-away

A Letter A Day. "Letters #314/315: Demis Hassabis and Dario Amodei" (CERN / IAEA-for-AI, Oppenheimer). Jan 2026. aletteraday.substack.com/p/letters-314315-demis-hassabis-and

Fei-Fei Li

Fei-Fei Li. "From Words to Worlds: Spatial Intelligence is AI's Next Frontier." Nov 2025. drfeifei.substack.com/p/from-words-to-worlds-spatial-intelligence

TIME. "Spatial Intelligence Is AI's Next Frontier." Mar 2026. time.com/7339693/fei-fei-li-ai

Fast Company. "Spatial intelligence is the next frontier of AI, says World Labs' Fei-Fei Li." Mar 2026. fastcompany.com/91503667/world-labs-most-innovative-companies-2026

Radical Ventures. "Building Spatial Intelligence: How World Labs is Creating the Next Frontier in AI." Nov 2025. radical.vc/building-spatial-intelligence-how-world-labs-is-creating-the-next-frontier-in-ai

Wikipedia. "Fei-Fei Li" (World Labs; science not science fiction, AI Action Summit). en.wikipedia.org/wiki/Fei-Fei_Li

Yann LeCun

TIME. "Meta's AI Chief Yann LeCun on AGI, Open-Source, and AI Risk" (existential risk called preposterous). time.com/6694432/yann-lecun-meta-ai-interview

StartupHub.ai. "Yann LeCun Left Meta to Put $1.03B Behind His LLM Critique." startuphub.ai/ai-news/ai-figures/2026

Let's Data Science. "Yann LeCun's World Models: Why LLMs Are a Dead End." letsdatascience.com

digidai. "Yann LeCun: Meta AI Godfather" (SB 1047 opposition; the Turing-three fracture). digidai.github.io/2025/11/27

Andrew Ng

The Register. "AI guru Ng: Fearing a rise of killer robots is like worrying about overpopulation on Mars." Mar 2015. theregister.com/2015/03/19/andrew_ng_baidu_ai

HuffPost. "Chief Scientist at Baidu, Andrew Ng, Explains if Artificial Intelligence Is A Threat To Humanity" (no realistic path to AI threatening humanity).

Brookings. "Are AI existential risks real, and what should we do about them?" (Mars analogy; AAAI survey). Jul 2025. brookings.edu/articles/are-ai-existential-risks-real-and-what-should-we-do-about-them

The three axes

Wikipedia. "Existential risk from artificial intelligence" (2023 extinction-risk statement and signatories; expert survey). en.wikipedia.org/wiki/Existential_risk_from_artificial_intelligence

AAAI 2025 Presidential Panel report, via Brookings: 76% of 475 surveyed AI researchers judged scaling current approaches "unlikely" or "very unlikely" to reach general intelligence.

European Union. AI Act, staged obligations taking effect through 2026 (governance-versus-capability pacing).