Hans free

 He Was Right About AI. What About the Fate of Mankind?

Hans Moravec predicted human-level AI by 2028. Now he contemplates our species’ end—and his own.

STUPIDER CORPSE MEDIA

The most spectacular prediction in certainly modern and possibly all history was made in 1988. It announced the arrival on earth of human-level artificial intelligence in 40 years. It was made by a Carnegie Mellon roboticist in his 30s named Hans Moravec. This is what Moravec did: While trying to get his robots to see and move faster, he became familiar with the qualities of the human retina, which contains a certain number of neurons, capable of a certain number of mathematical operations per second. Moravec took that number and multiplied it by the size of the rest of the human brain, achieving a rough estimate of the brain’s total computational capacity: 10 trillion operations per second.

That served as his target. To find out when it might be reached by artificial means, Moravec turned to Moore’s law, the now-venerable observation that silicon-chip capacity doubles every two years. He didn’t take Moore’s law on its own, though. He extended it backward in time so that it became not a law of computers but a law of intelligence. Across months of library work, he piled up estimates of the computation a dollar could buy over the previous 100 years, starting with a human clerk using pen and paper, proceeding to a 1920 device called the Torres Arithmometer, to a mechanical calculator built in 1938 by a young German engineer in his parents’ living room, to the early vacuum-tube computers built at the end of the Second World War, until commercial silicon appeared in the ’60s and Moore took over.

When these and other points were put together, they showed an increase in computing capacity of “a thousandfold every two decades since the beginning of the century,” which translates to a doubling every two years — Moore’s law all over again. Moravec produced a chart recording this consistent doubling, a logarithmic chart showing a flat line and a line rearing up and across from left to right. The first line was inherent human computational capacity — not changing. The second was artificial capacity — changing exponentially. The lines crossed in about four decades. Moravec announced the news of the human brain’s upcoming demotion in his book Mind Children, published in 1988. Though it was published by Harvard University Press and reviewed well in The New Yorker, hardly anybody took Mind Children that seriously. Nobody, with a few notable exceptions, set their clock.

This might have been because of what else appeared in Mind Children. Moravec’s rising line did not stop at human level. It continued, not slowing, as he traced a series of developments that the line portended. One of the tasks humans can do is design AI, so “human-level” AI means AI that can design itself, which, Moravec predicted, it would begin to do, making itself smarter and more efficient, more superhuman, until comparisons to human intelligence would become comical and drop off (nobody, after all, says that humans have supermouse intelligence).

This brought Moravec to the subject of the book: the Mind Children. “What awaits is not oblivion but rather a future which, from our present vantage point, is best described by the words ‘postbiological’ or even ‘supernatural.’ It is a world in which the human race has been swept away by the tide of cultural change, usurped by its own artificial progeny,” he wrote. Autonomous AIs outcompete humans and take over the economy, human minds get uploaded and then dissolved into computer minds, trillion-fingered robots appear (“It is unlikely that our superintelligent descendants will be satisfied with mere stumpy fingers”) and inherit the earth and the stars.

This was the broad Moravec vision — about as broad as any vision gets. It would nonetheless get broader. In the decade after Mind Children, he elaborated excitedly on it in talks, articles, online discussions, interviews, and a second book, Robot (1999), which announced the eventual conversion of everything in the universe to an expanding sphere of pure computation called “mind fire.”

While this was beyond too much for most people, Moravec’s vision did come to find a home in a few listservs and forums found in corners of the early internet. Spreading from there, it set the mold for a mind-set that now runs San Francisco and holds up the world economy. Sometimes the mind-set is called “the singularity,” though Moravec himself rarely favored the term. Another name for it might be Darwin of the universe, since the mind-set is premised on the intuition that, as intelligence has appeared in the universe at least once, it is inevitably going to happen again and even can be made to happen again.

This mind-set is not a wrong one. Intelligence is going to happen again. Going by the so-far accurate measures the AI industry has made, and by the so-far accurate measures the AI-monitoring industry has made, and several other measures, it is going to happen again soon — right when Moravec said. The systems in 2028 or thereabout will be below human ability in some ways and above in others and, especially once they are seen running on their own for weeks at a time, it will be perfectly reasonable to say they are roughly equivalent — more reasonable than saying they aren’t.

But as this all became clear in the second half of these last four decades, as the Moravec mind-set spread, Moravec the man vanished. A few years after Robot came out, he stopped publishing, speaking, commenting, and has rarely been heard from in public since. His name faded, but not to nothing. Many people, some of them inheritors of his vision, figured out how to reach him and asked for an update. They wanted to know what happened to him, how he felt about the world’s machines climbing the path he had mapped, and whether he still thought they would continue on past base camp, so to speak. Moravec turned them down. Then, this past April, in large part through the grace and the un-Hans-like willingness to pick up the phone of his wife, Ella, he changed his mind and allowed a visitor to come ask and hear him answer.

In the middle of Pennsylvania, in a modest-size house in a pleasant senior community above a college town astride a small river, Hans Moravec sits in a large reclining chair. He has occupied this chair most of the time since he and Ella moved here a handful of years ago from Pittsburgh, where he lived since joining Carnegie Mellon and where she grew up. On his lap is the largest iPad Apple sells. Across the living room is a large computer workstation with an Apple laptop, an Apple desktop, and two jammed-together, mismatched, non-Apple monitors so enormously large that 27 years ago, when Robot came out, each might have been the largest TV on the market and cost $15,000. Now, they remain dark most of the time.

In Moravec’s accounting, there is no great mystery to his decisions. He stepped back from public life because he got tired of racing around trying to convince people. He wanted to focus more on building the future than predicting it, and, having started a company to do so — Seegrid, which makes industrial robots based in part on technology he developed at Carnegie Mellon and Stanford — found that he liked business work, that he liked being home, and that his business partner didn’t like him giving away his thoughts and attention. He worked at Seegrid and on assorted other projects until a few years ago, after he developed a rare autoimmune disease called myasthenia gravis and his body consigned him to his large reclining chair.

Now he’s retired, out of the AI and robotics games. “I’m a spectator at this point,” he says. “I’m relaxing,” he says. “It’s great to see AI progressing the way it is without me having to do any effort.” He is content to look out the window and hang out with Ella, to whom he’s been married for more than 40 years; and like many retired men of 77 in 2026, he is content to spend much of the day on his iPad. His iPad now holds much of what he spent his career building. “That’s what I do now. I have a conversation with my artificial friend,” he says of ChatGPT. What do they talk about? His old hobby passions — “nuclear-power ideas, starship designs” — and the current trajectory of AI, which he confirms is proceeding to plan. “We’re on track, more or less. It’s not going to take that long,” he says.

His iPad also contains robots, in the form of the app for the house robovac, a Roborock, the latest in a long line of Moravec-family robovacs, and, he says, the first that lives up to the hopes he had for them in the ’80s, when he and Ella married and she vacuumed more than he did.

Moravec has always loved robots. He once said his love of them goes back to age 4, in postwar Austria, shortly before his family moved to Canada, when with his father he built a crank-powered wooden man from a construction set. The thrill of seeing an object animate from inanimate parts led him in the ’70s to a Ph.D. at Stanford with one of the founders of AI, John McCarthy. There, he worked out of a decaying wooden building and took stewardship of the legendary Stanford Cart — two sets of bicycle wheels bolted to a metal sheet with a camera on top — and turned it into one of the earliest self-driving vehicles.

He also began to develop the intuitions about intelligence, the necessary unification of the organic and artificial, that would lead to Mind Children. This went hand in hand with developing what you might call the Moravec style: brash, urgent, reckless, enthralling. Take a 1976 paper called “The Role of Raw Power in Intelligence.” Over 43 pages, through section titles like “Harangue” and “Bombast” and an acknowledgments line thanking devices like his PDP-KA10 computer for their “slave labor,” Moravec argues that the field of artificial intelligence is not looking hard enough at biology for inspiration and proof of concept; that if it did, it would see that “intelligence need not be so difficult to construct as is sometimes assumed” and that the main thing needed is not hard-won new insight but simply much more computing power. Moravec caps this polemic off with a first attempt at predicting when sufficient power to match the human brain would arrive. In his youthful enthusiasm, taking a few glances at the retina and computing history, the deadline he came up with was about a decade.

Of course, this was not to be. As he realized within a year or two, a decade from 1976 was way too early, and he put the predicting on hold. But a few years later, by then ensconced as a research scientist at Carnegie Mellon’s Robotics Institute, he figured he should give it another shot. This time, he’d be more thorough. But he stood by the essentials of the argument, including and especially the conclusion: more raw power.

It is hard in 2026, as data centers cover the land and are planned for space and sea, as they pressure the power grid and buy up every watt under and including the sun, to understand how far from correct this belief seemed in 1988, much less 1976. But Moravec was very lonely in it.

For most of their history, practitioners of AI saw their task as basically an artistic-scientific one: to find or sculpt or carve algorithms that would elegantly compress and capture the essence of intelligence. They held an Enlightenment-style faith that reason would boil down to a few governing principles — laws of thought that a sufficiently ingenious person could write down and a modest machine could run. This view was written into the founding of the field; the 1955 Dartmouth proposal, the document that coined the very term “artificial intelligence,” discusses it in its opening page. “Every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it,” the proposal stated. “The speeds and memory capacities of present computers may be insufficient to simulate many of the higher functions of the human brain, but the major obstacle is not lack of machine capacity, but our inability to write programs taking full advantage of what we have.”

To Moravec this seemed like wishful thinking. “If only you had the right algorithm,” he says now, parroting the old wish. Chuckling, he recalls that John McCarthy, his friend and doctoral adviser and one of the authors of the Dartmouth proposal, insisted that his research agenda, once fulfilled, would power human-level intelligence from the very same PDP10 Moravec thanked for its slave labor. According to Moravec, the PDP10, a mainframe computer first manufactured in the 1960s, did 1 million instructions per second — 10 million times less than the 10 trillion he thought was needed. This disagreement with McCarthy, he says, is what pushed him more than anything to write his books and move into the public eye.

And underlying the disagreement was a further sense that his elders’ view of intelligence was too narrow, too shallow. They thought about how the mind worked too much in terms of abstract reasoning and conscious cognition, while he thought the key was instead located in the layers of mind underneath the surface, in sensing and motion, and in the endless calculating the brain does that it is unaware of. This sense reinforced his conclusion: If AI was to replicate the unconscious, it would need much more power indeed.

He came to this view, he says, in large part because of his immersion in his robots. The Stanford Cart would sit still for ten or 15 minutes while it pumped visual data through pinhead processors and waited for guidance on the next right move. Then it would roll a meter forward on its thin wheels and do it again. It was obvious to him that 1 million instructions per second couldn’t possibly process even the limited number of pixels his robots with their analog cameras saw, and just as obvious that the human brain worked the same way, and that even if some hypothetical ultra-algorithm might be out there in the space of ideas, it would be faster and simpler to go with scale than to go out hunting for it. “With enough power anything will fly,” he wrote.

It has only been lately that the world has come around to this view. Why did it take so long? In part because it sounds absurd. You don’t need the beautiful heart of logic fully described? You just need to plug more chips and they’ll do it for you? The pseudonymous writer Gwern, perhaps the most talented essayist about AI alive, once felt that this idea “smacked of numerology and ‘If you build it they will come’ logic.” Then, after he accepted that Field of Dreams logic was real, he concluded, admiringly, “It’s Hans Moravec’s world, and the rest of us were just living in a fool’s paradise.”

Moravec demurs glory for his foresight. “You get lucky sometimes” is what he says.

The years after he made his prediction were prodigious ones. Or, as he puts it, “I was a whore for giving talks.” Moravec appeared in Wired several times, was interviewed on PBS, and was reviewed widely, including in the New York Review of Books by the future Nobel Prize–winning physicist Roger Penrose, who called his ideas “horrific.” Moravec replied in an open letter he posted on Usenet, then counter-reviewed Penrose’s weird ideas about consciousness a few years later. He also grappled with the eminent philosopher John Searle, whose thought experiments were (and still are) used to argue that no matter how smart a computer might appear, the inner truth would always come down to mechanical imitation. Of all his opponents, Searle seems to have annoyed Moravec the most; at one point, he fantasized about Searle succumbing “to the ‘mere imitation’ of strangulation at the hands of an insulted and enraged robot controlled by the ‘mere imitation’ of thought and emotion.”

Moravec confirms this, perking up at the mention of Searle and delightedly calling him “a rabble-rouser” more than once. Still, he insists he wasn’t out to fight back then. He mostly felt that he needed to call attention to what was coming.

The theater seems to have worked; in the ’90s, the implications of Moravec’s projections started percolating. In little corners of the internet, other people not attached to the standard views of intelligence began to debate the finer points of what comes after it appears in silicon and the postbiological age begins. Will Moravec’s call for paying out Social Security starting at birth really prevent mass impoverishment after human jobs vanish? Are his many ideas for how to upload brains workable? Is recursive self-improvement — the term of art for AI designing AI — possible and how quickly will it go and can it be accelerated?

Such questions, of course, form the heart of the set of expectations about the future now known as the singularity. A remarkable number of the people who discussed them on those early listservs are now famous, among them Eliezer Yudkowsky, Nick Bostrom, Robin Hanson. Along with the futurists Vernor Vinge and Ray Kurzweil, whose famous exponential charts describing the “law of accelerating returns” were highly influenced by Moravec, this group would become known for transmitting these expectations and their out-there flavor to Silicon Valley, Sam Altman, and Elon Musk.

Perhaps the peak of Moravec’s profile came in 1999, when he published Robot, which took the more speculative aspects of Mind Children and, as he puts it now, pushed them as far as they could go, all the way to the end of the universe and to the roiling ball of expanding consciousness called mind fire. His excitement for the postbiological era is hard to miss. “Producing artificial offspring that transcends us to the max is the most exalted role I can imagine the human race playing,” he said in an interview around Robot’s release.

His metaphor — artificial offspring, mind children — indicated postbiology would be a fundamentally or at least mostly placid process, an extension of natural generational turnover, wherein young robots are nurtured by and then eventually outgrow and succeed human parents. As Moravec described it in Mind Children, “We humans will benefit for a time from their labors, but sooner or later, like natural children, they will seek their own fortunes while we, their aged parents, silently fade away.”

This placidity had limits, however. Eventually there might be no avoiding some Oedipal conflict — the robots could go off into space and evolve and come back to Earth and destroy what remained of their biological ancestors — but that would take a while. Moravec didn’t exactly desire this part of the picture, but he didn’t seem too bothered by it either. Others felt differently.

“I remember having a debate with Joe Weizenbaum,” Moravec says. Weizenbaum was another of the last century’s great AI researchers. “It was in Germany with an audience that was definitely on Weizenbaum’s side. His word was that basically what we’re talking about is genocide of biological humanity. And I thought, Well, you know, no species lasts forever.”

No species lasts forever — this is the natural conclusion of spreading Darwin across space and time. When asked if he’s changed his mind about when and how this might happen, Moravec says yes and no. He thinks the Oedipal conflict is still on the way, but he’s not sure it’ll take that long to get here anymore — it might come sooner. The reason has to do with the one big mistake about the future that he made. AI today indeed relies on the huge amount of raw computing power Moravec thought it would. But the shape it takes, what puts the power to work, is not what he imagined it would be back in the ’80s and ’90s.

Moravec didn’t think the brain of advanced AI would rely on the hand-carved algorithms his elders expected, but he did