Ilya Sutskever
Co-founder, Chief Scientist at SSI. Former Co-founder, Chief Scientist at OpenAI.
AI-generated profile based on archived statements
The man who proved scaling now says the era of scaling is over. Having handed the industry its religion, he has quietly walked off to look for the next one.
Ilya Sutskever spent a decade telling everyone that the bitter lesson was right: pour in compute, pour in data, and intelligence falls out the other end. He was the patron saint of that gospel. Then, in his first real interview after leaving OpenAI, he announced that "the age of scaling sucked the air out of the room" (Dwarkesh interview, repackaged by Py Man, Nov 2025) and that the field had to return to what he calls the age of research. That is not a tweak. That is the person who built the engine telling you the engine has a ceiling.
The man who recanted his own bitter lesson
For five years the dominant religion in AI was scaling laws, and Sutskever was arguably the one who proved them. So when he says the strategy got lazy, that the field "stopped innovating and started just stacking GPUs," it lands differently than it would from a critic. He is describing a trap he helped set.
His diagnosis is specific. "We're running out of internet," he says, "running out of high-quality human data" (Nov 2025). Pre-training worked because the answer to what data? was simply everything. Its strength was that you never had to think hard about what to feed the model. But that well is finite, and the post-pre-training world of RL training forces a choice nobody knows how to make well. Companies, he reports, now have whole teams that "just produce new RL environments and just add it to the training mix" (fOx Hsiao repost of Dwarkesh, Dec 2025). The degrees of freedom are enormous, and people quietly take inspiration from the evals. His sharpest line is that the real reward hacking is the human researchers who are too focused on the benchmarks.
This is the load-bearing belief underneath everything: that the visible numbers have decoupled from the thing they were supposed to measure.
The eval gap, and why the smartest models are also stupid
Sutskever keeps circling one confusion he refuses to paper over. "The models seem smarter than their economic impact would imply" (Dec 2025). They ace hard evals and then reintroduce a bug they were just told to fix, then reintroduce the first one, alternating forever. "It's like, how is that possible? I'm not sure," he admits. The honesty is the point... he is the field's most cited researcher saying out loud that he does not understand his own creation's failure mode.
His best guess is an analogy, not a theorem. Take two competitive programmers. One drills 10,000 hours, memorizes every proof technique, becomes the best. The other practices a hundred hours and is nearly as good. Which one wins later in life? The second. "The models are much more like the first student," he says, "but even more" so, because we then scrape every competition problem ever written and augment it. The result generalizes beautifully to the test and poorly to the world.
What the second student has, he thinks, is taste, judgment, the ability to derive logic from first principles instead of pattern-matching. And here he reaches for the strangest idea in his current repertoire.
Emotions as a value function
Sutskever has become fascinated by a man with brain damage. The patient lost emotional processing, stayed articulate, could still solve puzzles, tested fine... and became "extremely bad at making any decisions at all." Hours to choose which socks to wear. Catastrophic financial choices (Dec 2025).
His reading: emotions are not a bug to be engineered out. They are "some kind of a value function thing," a compression algorithm for survival that lets you short-circuit. Lose a piece in chess and you know you blundered without playing to checkmate. Current models have no such internal compass, which is why he suspects value functions "don't play a very prominent role in the things people do" right now, and why he is not sure pre-training can ever supply one. When an interviewer accused him of showing "such lack of faith in deep learning," he didn't fully back down... the tell of a believer who has started auditing his own faith.
So this is the singular thing about him among frontier-lab leaders. Altman sells the future, Hassabis productizes the science, Amodei warns. Sutskever, having seeded the orthodoxy, is the one publicly dismantling it, betting that the next breakthrough comes from biology and judgment rather than from another order of magnitude of silicon. SSI's whole premise (safe superintelligence, no products, no revenue treadmill) is the institutional form of that bet.
The tension he lives with
He cannot make himself feel what he believes. Sutskever is certain superintelligence is coming, "slowly but surely, or maybe not so slowly" (Jamin repost, Jan 2026). His proof is almost childlike and he likes it that way: "all of us have a brain and the brain is a biological computer... so why can't a digital computer, a digital brain do the same things?" Once you accept that the day comes when AI can do "not just some" of our jobs "but all of them."
And yet he confesses the gap between knowing and feeling: "It's very difficult to internalize and to really believe on an emotional level. Even I struggle with it." This from the man who just told you emotion is the missing piece of intelligence. He has diagnosed the human value function as the secret ingredient, and admitted that his own value function refuses to register the stakes he intellectually accepts.
His warning is therefore quiet, not apocalyptic. The danger isn't Terminator. It's the "normality paradox"... that we'd invest 1% of GDP in AI and it would "just feel like" a news headline about a number nobody can parse. The risk is that the most consequential transition in history arrives feeling abstract, and we acclimate to the abnormal one update at a time. "You may not take interest in politics," he likes to say, "but politics will take interest in you. So the same applies to AI many times over." He calls it, flatly, "the greatest challenge of humanity ever."
He thinks the only honest preparation is to keep looking. No essay can compete with what you see "with our own two eyes." Which is a peculiar prescription from a man who just told you that what the models see on the evals is exactly what's fooling everyone.
Recurring themes
Featured in
Sutskever vs. the Scaling Consensus
Ilya Sutskever arguably proved scaling laws work better than anyone alive. He co-authored AlexNet, co-founded OpenAI, and oversaw the GPT series that turned neural scaling from a research curiosity into the dominant paradigm. Then he left and announced that the age of pure scaling is over. His argument: scaling sucked the air out of genuine research. Models trained on massive compute ace benchmarks but fail at generalization, like a student who memorized 10,000 competitive programming problems but can't architect real software. He thinks RL training produces meta-reward-hacking, where the researchers themselves (not just the models) are unconsciously overfitting to evaluation metrics. This puts him directly at odds with his former colleagues. Altman and Brockman are spending hundreds of billions on data centers and custom chips, betting that more compute is the binding constraint. Amodei still describes intelligence as a chemical reaction with known ingredients. Even Hassabis, who shares some of Sutskever's skepticism about pure scaling, is building gigawatt-scale infrastructure. Sutskever's counter-thesis is that AI needs something analogous to human emotions: a learned value function that provides fast, approximate feedback about whether a course of action is promising, rather than the sparse end-of-trajectory reward signals that current RL relies on. He cites a neurological case study of a stroke patient who lost emotional processing but retained full IQ, and became unable to make even trivial decisions. The implication: without an internal compass, raw intelligence is paralyzed. SSI is his bet that ideas, not compute budgets, are what's actually missing.
with Sam Altman, Greg Brockman, Dario Amodei