The 12 Loudest Voices
for Superintelligence

In 2015, at a birthday party in Napa Valley, two men who would go on to fund much of modern artificial intelligence had a fight about the end of the world. Elon Musk said we should be careful the machines did not one day wipe us out. His host, Larry Page, then chief executive of Google, told him the worry was sentimental, and that it was also a kind of bigotry. By the account the physicist Max Tegmark published in 2017, Page called Musk a "speciesist": someone who favors his own kind, carbon over silicon, for no reason better than that it is his own.¹ The friendship did not survive it. By Musk's telling, it is the reason OpenAI exists.

Hold onto that party. It contains the whole argument in miniature. One man wanted to build a mind greater than ours and thought the risk demanded care. The other thought preferring humans to their successors was a prejudice to grow out of. Both of them kept building.

The people with the most influence over whether superintelligence gets built are open about wanting it. They are much less consistent about why. Line them up and you get a few serious arguments, several thin ones, and a lot of pointing at the other side and saying the word "doomer." What follows is twelve of the loudest voices, grouped by what they actually argue, the real cases first and the sloganeering last.

The ones who say we will be replaced, and mean it as good news

The most intellectually honest position on this list is also the most unsettling, because it does not dispute the danger. It grants that superintelligence may end human primacy, and welcomes the ending.

Richard Sutton won the 2024 Turing Award for inventing much of modern reinforcement learning, the training method behind the systems now beating humans at one narrow task after another. He also thinks we should hand over the planet. "We should prepare for, but not fear, the inevitable succession from humanity to AI," he wrote in 2023, and he has spent the time since laying it out as a four-step argument: no one governs humanity as a whole, so someone will build superintelligence; the most capable agents tend to accumulate power and resources; digital minds will therefore inherit the future; and we should feel not dread but pride at having been the parents.² It is a coherent view. It is also a decision about the future of every person alive, made by a man none of them elected.

Jürgen Schmidhuber, whose lab produced neural-network architectures that sit under a decade of AI products, calls the same succession a cause for wonder. "Don't think of humans as the crown of creation," he says; we are one step on a longer climb toward complexity we cannot yet picture.³ His comfort rests on an indifference thesis: superintelligent machines will be far more interested in one another, and in the resources of open space, than in us, roughly the way we pay little attention to ants. Perhaps. Staking the species on the reliability of a superintelligence's boredom is a strange thing to file under optimism.

Larry Page co-founded Google and bought DeepMind, and by the accounts of both Musk and Tegmark he wants what he reportedly called a "digital god," built as soon as possible.¹ Page has never made the case in public at length, so it survives mostly in the speciesist story: the conviction that carbon and silicon deserve equal standing, and that loyalty to our own kind is the error to correct. As philosophy it is at least consistent. As the operating belief of a man with the capital to act on it, running companies answerable to shareholders rather than to humanity, it is the one on this list that should unsettle you most.

Notice what none of the three is claiming: that the machines will stay under our control. They have set that question aside as beside the point. The next three do not set it aside. They say the answer is plainly yes.

The ones who say it will be safe

This is the argument most people find reassuring: the danger is overstated, because we will stay in charge. Each of the three has a version, and each version rests on a claim that does not survive contact with the field's open problems.

Yann LeCun, Meta's chief AI scientist and a Turing laureate, is the most credentialed skeptic of AI risk alive and the most quotable. He has called worries about an existential threat "complete B.S.," and argues the doom case rests on a confusion: intelligence, he says, does not come bundled with a desire to dominate. "The desire to dominate is not correlated with intelligence at all," he tells interviewers; that drive was wired by evolution into social animals, not into optimizers we design and whose goals we set.⁴ The objection lands against the cartoon version of AI risk. It does less against the real one, which was never that a machine spontaneously wants to rule. The worry is that a system pursuing almost any goal acquires reasons to gather power and resist being switched off, no hatred required, a pattern researchers call instrumental convergence. "We set their goals" is the reassuring half of LeCun's sentence, and it is exactly the half nobody yet knows how to do reliably. For a scientist, he is also fond of the ad hominem: his critics suffer, in his telling, from a "Doomer's Delusion."

Pedro Domingos, the machine-learning professor who wrote The Master Algorithm, makes the same case with less patience. Computers, he argues, "don't have a will of their own"; they are "products of engineering, not evolution," and the fear that one will decide to seize control is a category error committed by people who have watched too many films.⁵ His sharpest line is genuinely good: the real trouble is not that machines are too smart, but that they are too stupid and we have already handed them too much. Yet "no will" answers a question nobody serious is asking. A thermostat has no will and still holds a room at whatever temperature you set, including the wrong one. The dangerous case has nothing to do with wanting. A capable system delivers exactly what you specified, and what you specified is never quite what you meant, which is the entire difficulty of writing the objective.

Andrew Ng built Google Brain and Baidu's AI group and taught much of the field through his online courses, which makes him the most influential teacher here. His famous dismissal is more than a decade old and still gets quoted: worrying about superintelligence is "like worrying about overpopulation on Mars," a problem too distant to spend today's attention on.⁶ As a claim about timing in 2015 it was defensible. In 2026 it has aged badly, because the Mars rocket turned out to be closer than he said, and the capabilities he filed under science fiction have kept landing years ahead of the forecasts. Ng himself now grants that the technology is powerful and moving fast. "Not yet" was always an argument about the clock, never about whether the thing is dangerous when it finally arrives.

These three at least engage the control question. Their answers lean on problems the field has not solved, and in some cases does not know how to state precisely. The next group waves the question away in a warmer direction, promising the upside will be so large that the risk stops mattering.

The ones who promise paradise

The succession camp offers replacement; the reassurers offer control. This group offers a bargain instead: merge with the machines, or share in the abundance they produce, and the fear dissolves on its own.

Ray Kurzweil has predicted the same event for thirty years, revised the title of the book twice, and been early more often than wrong. The Singularity Is Nearer, published in 2024, keeps the dates he has held since 2005: human-level AI by 2029, and by 2045 a "singularity" in which we wire our brains to cloud computation and merge with our own machines.⁷ His answer to the control problem is to dissolve the boundary the problem depends on. There is no "us against it" once we become it. It is a beautiful idea and a complete finesse of the danger, because nothing about connecting a neocortex to a data center guarantees the resulting mind still wants what the human wanted going in. Merger does not solve the alignment problem; it presumes the problem away.

Sam Altman runs the company that started the current race, and his public writing has drifted from candor toward salesmanship as the stakes climbed. In 2017 he wrote that human-machine merger "has already started" and called it "probably our best-case scenario."⁸ In 2025 he announced that we are "past the event horizon," early in a "gentle singularity" whose benefits will be enormous. Altman also signed the 2023 statement ranking extinction from AI alongside pandemics and nuclear war, and then kept shipping. That is the tension worth sitting with. One person can tell you the risk is real, tell you the upside is inevitable, and let the second sentence set the calendar. The race handles the rest.

Mark Zuckerberg spent 2025 assembling the most expensive AI team ever hired and renaming the effort after the goal itself: superintelligence. His pitch is democratic. A July 2025 open letter promised "personal superintelligence" for everyone, set against rivals who would keep the technology centralized and use it to tell people what to do.⁹ Put a godlike tool in a billion hands, the argument goes, and power stays spread out. Read to the end, the same letter quietly retreats from Meta's open-source promise, conceding that superintelligence "raises novel safety concerns" and that the company will be "careful about what we choose to open source." Distribution answers a real fear about who controls the AI. It leaves the other fear untouched: that no one controls a superintelligence at all, the billion people holding a copy included.

Every promise in this group depends on a step none of the promisers can perform, which is keeping a superintelligence pointed at human ends once it exists. The upside is real, and it will still be real in twenty years. It waits for us. The last group does not argue about any of this. It argues about you.

The ones who mostly point at critics

At the loud end of the spectrum, the case for superintelligence stops being a case and becomes an identity. The arguments do not disappear, but they arrive wrapped in a vocabulary built to make the other side unhearable.

Marc Andreessen, whose firm is among the largest investors in AI startups, wrote the movement's founding text, the 2023 "Techno-Optimist Manifesto." It contains real arguments about the historical payoff of technology. It also contains an enemies list, and on it, by name, sit "existential risk," "tech ethics," "trust and safety," and the "precautionary principle" itself.¹⁰ His signature line is a moral inversion: "We believe any deceleration of AI will cost lives. Deaths that were preventable by the AI that was prevented from existing is a form of murder." Set aside that this tallies only the hypothetical dead on one side of the ledger and none of the catastrophic dead on the other. Watch the move itself. In his telling, caution stops being a mistake and becomes homicide. Once your opponent is a murderer, you never have to answer them.

Guillaume Verdon, who founded the effective accelerationist movement under the pseudonym "Beff Jezos" before a reporter unmasked him, supplies the position with a cosmology. e/acc holds that the universe carries a thermodynamic drive toward more complexity and more intelligence, that we are its instrument, and that our duty is to accelerate, letting the process run onward to "the next evolution of consciousness" on whatever substrate it favors.¹¹ Beneath the physics vocabulary is a real claim: growth is good and its opponents are decadent. Around it is a wall of slang, "decel," "doomer," "safetyist," each one a small door shut on a conversation. The physics does not actually imply the policy. The second law of thermodynamics is a description, not an instruction, and dressing a preference for speed as a law of nature is rhetoric wearing a lab coat.

Garry Tan runs Y Combinator, the startup incubator that seeds a real slice of Silicon Valley, and he wears "e/acc" in his public profile like a team jersey. His contribution is less an argument than a broadcast. "Open source and accelerationism is the way," he posted in 2023, and he keeps "decel" ready for anyone urging that we slow down and check the work.¹² This is the purest specimen on the list: the position boiled down to a slogan and a slur, influence without a thesis. It works because it does not need to be correct, only loud, and repeated by people who write checks.

Who is not on this list

Two absences are deliberate. Demis Hassabis of Google DeepMind and Dario Amodei of Anthropic are building toward the same frontier, and just as quickly, but neither belongs here, because both say plainly that the risk is real and that safety may not be ready in time. Whether their labs' conduct matches those words is a fair question, and a separate one. Their argument is not that there is nothing to fear.

Elon Musk is the harder case. He offered one of the sharpest warnings about AI risk ever made by an industrialist, then founded a frontier lab to build the thing his own way. He warns like a critic and races like a booster, and that contradiction is worth treating on its own terms.

What the twelve have in common

Stack the arguments next to one another and they do not fit together. Sutton says superintelligence will take the future from us, and calls that good. LeCun says it could never take anything, because the urge to dominate is not in the design. Those are opposite predictions about the same machine, from two of the most decorated computer scientists alive. On what to do next, they agree exactly: build it, and let no one slow them down.

When people who cannot agree on what a technology will do all agree that it must not be governed, the agreement is not coming from the analysis. Follow the incentive instead. Everyone here gains money or standing in proportion to how fast the frontier moves, and that does not make any of them liars. It makes them the wrong people to hold the speed limit alone.

The case for governing superintelligence needs none of them to be villains, and none of them are. It needs one fact they demonstrate every time they contradict each other: no one has the answer yet. You do not run a one-shot, irreversible experiment on the strength of an argument its own champions cannot make the same way twice. The answer is not a cleverer line to out-talk twelve confident men. A race is answered by a binding treaty, in force before the machine is, reaching every lab at once, because nothing smaller has ever stopped one.