Ilya Sutskever

The Architect Bets Against Scale

Ilya Sutskever is the researcher behind three of the results modern AI runs on: AlexNet in 2012, sequence-to-sequence learning, and the scaling philosophy that produced ChatGPT and the first reasoning models at OpenAI. In 2024 he walks away to start Safe Superintelligence, a company with no product, built on the bet that the field has exhausted easy gains from raw scale and now needs new ideas — chiefly, models that generalize the way humans do. For a CTO, the signal is that benchmark scores are drifting apart from business value, and the next capability jump may not arrive on a predictable compute schedule.

In 1991, a five-year-old boy leaves the city of Gorky (GOR-kee) — a Soviet industrial town on the Volga that will later take back its old name, Nizhny Novgorod (NIZH-nee NOV-guh-rod). This is Ilya Sutskever (IL-yuh SOOT-skuh-ver). His family makes aliyah (ah-lee-YAH) — they emigrate to Israel — and settle in Jerusalem (juh-ROO-suh-lem). His father is an engineer, and somewhere around that age, at an exhibition, the boy sees a computer for the first time. As he tells it later, the machine leaves him "utterly enchanted." That word does a lot of work in this story.

He turns out to be the kind of student who runs out ahead of the room. By eighth grade, he is already taking university classes through the Open University of Israel. At sixteen, in 2002, the family moves again, this time to Toronto. He spends about a month in a Canadian high school before the University of Toronto admits him straight out of eleventh grade, into a mathematics program, as effectively a third-year student.

He stays for three degrees. A bachelor's in math in 2005, a master's in computer science in 2007, a PhD in 2013. And early on he does the single most consequential thing of his young career. He walks down the hall and attaches himself to Geoffrey Hinton (JEF-ree HIN-tun) — the British-Canadian scientist who has spent decades on an idea most of the field treats as a dead end: training artificial neural networks. Sutskever becomes Hinton's student and works alongside him for more than a decade.

In 2012, the dead-end idea pays off in public. Sutskever convinces a fellow graduate student, Alex Krizhevsky (KRIZH-ev-skee) — who has a rare gift for wringing performance out of graphics cards — to train a large neural network on ImageNet, a dataset of over a million labeled photographs. Hinton is the principal investigator. The network, which the field will call AlexNet, enters the 2012 ImageNet contest and wins it, and not by a hair. It cuts the top-five error rate from about twenty-six percent to about fifteen.

That is not an improvement. That is a different regime.

The gap is so wide that within a year or two, computer vision reorganizes itself around deep neural networks, and the rest of AI follows. Then events move fast. In 2012, Sutskever spends about two months at Stanford as a postdoc with Andrew Ng (Andrew Ing). He returns to Toronto, where he, Hinton, and Krizhevsky form a tiny company called DNNResearch. In 2013, Google buys it, and Sutskever becomes a research scientist at Google Brain. There, with Oriol Vinyals (OR-ee-ol vin-YALS) and Quoc Le (kwok lay), he builds sequence-to-sequence learning — a method for mapping one sequence to another with deep networks — which becomes the basis for a major upgrade to Google Translate. He is also one of the many co-authors on the AlphaGo paper. By his late twenties, his hands are on three of the decade's landmark results.

And then, at the end of 2015, he leaves Google.

He becomes co-founder and chief scientist of a new nonprofit lab called OpenAI. The conviction he brings — the thing colleagues credit him with pushing harder than anyone — is almost embarrassingly simple. Make the networks bigger. Feed them more data and more compute. They will keep getting better. People start calling it the scaling ethos, and for years it is the house religion at OpenAI, and for years it is right. GPT, then a second version, then a third, then ChatGPT at the end of 2022. He also helps steer the image models, DALL-E and CLIP, and later the research line that becomes the first reasoning models, the ones in the o1 family.

But somewhere in those years, the worry grows alongside the capability. In 2022, he tweets that "it may be that today's large neural networks are slightly conscious," and takes the mockery that comes with it. In July 2023, he co-founds and co-leads a team at OpenAI called Superalignment, with the researcher Jan Leike (yahn LY-kuh). Its job is to figure out how to control AI systems smarter than people — a problem Sutskever thinks could arrive this decade, however far off it sounds.

Then comes November 2023, and the turn that makes him worth an episode.

Sutskever is one of six directors of the nonprofit board that controls OpenAI. On November 17, 2023, that board fires its own chief executive, Sam Altman. The board's public statement says Altman was "not consistently candid in his communications with the board." Sutskever has written a long internal memo — reported to run to dozens of pages — building the case against Altman. In an all-hands meeting right after, he tells staff the firing is the board doing its duty.

It comes apart in days. Employees revolt. Something close to the entire company threatens to walk out the door. And within roughly a week, Altman is back as chief executive. Sutskever reverses himself in public, posting that he "deeply regrets" his "participation in the board's actions," and he steps down from the board. For about half a year after that, he is a kind of ghost at the company he helped build — still employed, largely absent.

In May 2024, he makes it final. He leaves OpenAI for a project he describes only as "very personally meaningful." OpenAI names Jakub Pachocki (YAH-koob puh-HOT-skee) as its new chief scientist. Jan Leike resigns within hours, saying safety had taken a back seat to shiny products.

Here is the shape of the turn. A man who spent nearly a decade arguing that scale is the path, and who built the lab that proved it, walks away from the frontier lab at the exact moment it is winning — to go start something that competes on the opposite principle.

In June 2024, with Daniel Gross and Daniel Levy, he launches Safe Superintelligence Incorporated. The pitch is unusually blunt. As Sutskever puts it, the company's "first product will be the safe superintelligence, and it will not do anything else up until then." No chatbot. No API. No revenue on the way to the goal. One objective, pursued straight, out of offices in Palo Alto (PAL-oh AL-toh) and Tel Aviv (tel uh-VEEV).

The money shows up anyway. In September 2024, barely three months in, the company raises a billion dollars at a valuation around five billion, from investors including Andreessen Horowitz (an-druh-SEN HOR-oh-wits) and Sequoia (seh-KOY-uh). In the spring of 2025, it raises two billion more, at a figure reported at thirty-two billion, with Alphabet and Nvidia (en-VID-ee-uh) reported among the backers. If those numbers wash over you, hold on to just one idea: investors are paying frontier-company prices for a company that has shipped nothing.

In June 2025, Safe Superintelligence turns down an offer from Meta to buy it outright. A month later, Daniel Gross leaves for Meta, Sutskever takes over as chief executive, and Daniel Levy becomes president. And in July 2026, the company announces a long-term partnership with Nvidia; Reuters reports that Nvidia plans to invest five billion dollars, though the companies don't disclose the amount, and Safe Superintelligence says it expects to grow its compute by roughly tenfold.

I want to be honest about the limits here, because Sutskever's whole case rests on things outsiders can't yet see. There is no public product. No benchmark. No released model. The valuation is a price investors negotiated, not a measure of progress. So the right posture is to take the record we can inspect — the papers, the departures, the stated mission — seriously, and to hold the private claims lightly.

Which brings us to the part worth slowing down for: the actual technology, and why a man who won by scaling now thinks scaling alone has run out of road.

Start with what AlexNet actually did, mechanically, because it sets the template for everything after. Before 2012, if you wanted a computer to recognize a cat in a photo, a human expert hand-designed the features — edges, corners, textures — and the machine matched against them. AlexNet throws that out. It learns the features directly from raw pixels. You show it a labeled image, it makes a guess, you measure how wrong it is, and you push the error backward through the layers of the network, nudging millions of internal weights a little in the right direction. Do that across a million images, over and over, and the lower layers learn to detect edges, the middle layers learn textures and parts, the top layers learn whole objects. Nobody tells it what an edge is. It discovers that edges are useful.

Two things had kept that idea from working. One was compute — the math is an ocean of matrix multiplications, and that is exactly what graphics cards are built for, so AlexNet runs on a pair of consumer GPUs. The other was a network this big memorizing its training data instead of learning the general pattern, and AlexNet leans on training techniques that deliberately handicap the network during learning so it can't just memorize. Hardware plus data plus a few tricks. That combination is the whole modern recipe in miniature.

Sequence-to-sequence learning, two years later, stretches the idea from images to language. A photo is a fixed-size grid. A sentence is variable length and ordered, and its meaning depends on that order. So the design uses two networks. One reads the input sentence, word by word, and compresses it into a vector — a point in a high-dimensional space that stands for the meaning. The second network takes that vector and generates the output, one word at a time, in another language. Read the context, then predict the next piece. That read-then-predict loop is the beating heart of every large language model that follows.

Scale that loop up and you get GPT. You take a very large network, feed it an enormous slice of the internet, and train it to do one thing: predict the next word. There are no human labels — the text is its own answer key, because the next word is always sitting right there in the data. Sutskever's deeper claim, the one underneath the scaling ethos, is that prediction forces understanding. To reliably predict the next word across physics papers and legal filings and jokes and code, a network has no choice but to build an internal model of the thing producing that text — the world, and the people in it. Compression, in his telling, is comprehension. And from roughly 2020 to 2025, the field runs hard on that bet: bigger models, more data, more compute, and the capabilities keep climbing.

Reasoning models are the next layer on top. Instead of answering immediately, the model is trained to think first — to generate a long chain of intermediate steps before committing to an answer. You train that behavior with reinforcement learning against problems where the answer can be checked, like math and code: let the model try, reward the attempts that land on the correct result, and it gradually learns to reason in a way that pays off. In the November 2025 conversation with Dwarkesh Patel (DWAR-kesh puh-TELL), Sutskever talks about one of the tools that makes this more efficient, a value function — roughly, a learned sense of whether you're on a promising track partway through, so you don't have to run every attempt to the end before you know it's bad. But he's careful to downgrade it. Anything a value function buys you, he says, you could get without one, just more slowly. It is an accelerator, not the core.

So what, in his view, is the core problem? This is the pivot. In that same November 2025 interview — the one titled "we're moving from the age of scaling to the age of research" — he lays out a puzzle. The models now ace genuinely hard evaluations. They score like experts. And yet their impact out in the economy lags well behind what those scores would predict. He calls that jaggedness: dazzling on the test, strangely brittle in the messy real task.

His diagnosis cuts deep. As he puts it in that conversation, "these models somehow just generalize dramatically worse than people." That's the whole thing in one sentence. A human teenager learns to drive in something like ten hours. A model needs a quantity of data no person could consume in a thousand lifetimes, and it still stumbles when the situation drifts a little outside what it has seen. The gap isn't knowledge. It's generalization — the ability to take a little experience and extend it correctly to a genuinely new case.

And that reframes his whole career into a timeline. From 2012 to 2020, he argues, the field is in an age of research — new ideas, new architectures. From 2020 to 2025, it's an age of scaling — one word, one knob, turn it up. But the scale is now so enormous that turning the knob further buys less and less. So, in his telling, the field circles back to where it started: an age of research, where the next real jump comes from a new idea about generalization, not another data center. When Dwarkesh asks how long until a system can learn as efficiently as a human and then surpass us, Sutskever gives a wide band — five to twenty years.

Hold on to that range, because the honesty of it matters. Not "two years." Not "it's basically here." Five to twenty, with his own uncertainty built in.

So let's pull back, because this is where it gets practical for anyone who builds, hires, and budgets.

The first thing to take is the eval-versus-value gap, and you can act on it immediately. If the single most scaled-up AI researcher alive is telling you that benchmark scores and economic usefulness have come apart, then a model's leaderboard number is not a purchase order. The only test that counts is the model running on your actual workflow, with your messy data and your edge cases. Budget for that evaluation work as a permanent line item, not a one-time bake-off. The jaggedness is real, and it shows up precisely in the gap between the demo and the deployment.

The second thing is the shape of the bet itself. Sutskever is wagering that the curve bends — that pure scale is hitting diminishing returns and the next leap needs an insight nobody has published yet. His former colleagues at the big labs are wagering the opposite: that scale has more room, and steady engineering keeps delivering. You do not have to pick the winner. But you should notice that two of the most serious people in this field are now pointing in different directions, and that tells you the near future is genuinely uncertain. For a CTO, the move is to avoid roadmaps that assume capability improves on a smooth, predictable schedule. Plan for a world where the next big jump might land next year, and also for one where it stalls for a while. Both are live.

The third thing is the dependency underneath all of it. Notice that even Sutskever — the safety contrarian, the man building the deliberately slow company — ties himself to Nvidia for compute and talks about growing it tenfold. If the person trying hardest to step outside the race still can't step outside the hardware, that is the real structural fact of this industry. Compute is both the moat and the chokepoint. When you plan capacity, treat concentration on a single hardware and software ecosystem as a risk to price in, not a detail.

And the fourth thing is the quietest and maybe the most useful. Safe Superintelligence is, in effect, a thirty-billion-dollar argument that the asset worth building is the model itself, not a product wrapped around it. That is a bet most companies should not copy, because most companies are integrators, not frontier labs — your edge is in the workflow, the data, and the distribution you already own, not in training a base model from scratch. But watch what the lab publishes over time. The open question hanging over it is whether the word "safe" ever acquires an operational, testable meaning — threat models, evaluations, failure reports — or stays a name on the door. How that resolves will tell you a lot about whether "safe AI" is becoming an engineering discipline you can buy and audit, or remains a promise you have to take on faith.

The through-line of this whole story is a man who was right about something enormous — that scale works — and who is now, with real money and real conviction, betting that the thing he was right about has stopped being enough. He could be early. He could be wrong. But the gap he's pointing at, between a model that aces the test and a model that actually generalizes to your problem, is not a philosophical abstraction. It is sitting in your evaluation results right now. He just said it out loud first.