Ilya Sutskever helped prove that scale works in AI. Now he says it doesn't—and his $3 billion company is betting the future belongs to researchers with ideas, not labs with the biggest GPU clusters. The industry isn't ready for this argument.
Silicon Valley promised AI would democratize creativity. New research tracking 442 participants found the opposite: people who were more creative without AI produced better work with it. The gap didn't close. It may have widened.
Enterprises are spending billions on AI pilots, but MIT's research shows most deliver no return. It's not the technology failing. The gap between impressive demos and working systems comes down to data quality, technical debt, and organizational readiness.
Ilya Sutskever helped prove that scale works in AI. Now he says it doesn't—and his $3 billion company is betting the future belongs to researchers with ideas, not labs with the biggest GPU clusters. The industry isn't ready for this argument.
Ilya Sutskever thinks everyone is doing AI wrong. Not just slightly wrong. Fundamentally, paradigmatically wrong.
In a recent interview with Dwarkesh Patel, the co-founder of Safe Superintelligence (SSI) and former OpenAI chief scientist laid out a thesis that should unsettle every frontier lab executive currently signing billion-dollar compute contracts. The age of scaling, he argues, has ended. What comes next requires something the industry has largely abandoned: actual research.
"We got to the point where we are in a world where there are more companies than ideas," Sutskever observed. "By quite a bit."
The statement lands differently coming from him. This is the researcher who co-authored AlexNet, helped build GPT-3, and spent a decade proving that scale works. Now he's saying it doesn't. At least, not anymore.
The Breakdown
• Sutskever claims AI's "age of scaling" (2020-2025) has ended; current approaches will plateau and cannot produce genuine intelligence
• The core unsolved problem: models generalize far worse than humans despite crushing benchmarks, suggesting fundamental architectural flaws
• SSI's $3 billion buys comparable research compute to rivals once you subtract their inference, product, and engineering costs
• Timeline forecast: human-like learning systems arrive in 5-20 years through research breakthroughs, not scaling existing methods
Sutskever's argument begins with a linguistic observation. "Scaling" became a single word that shaped corporate strategy across the industry. Companies could pour capital into compute and data with reasonable confidence that results would follow. Pre-training offered a recipe: mix compute, data, and parameters according to established ratios, and capabilities improve predictably.
This clarity attracted investment. Research carries risk. Scaling carries cost. The distinction matters enormously when you're deploying billions.
But the recipe has an expiration date. Pre-training data is finite. Companies have already scraped most of the internet's text. Synthetic data generation helps at the margins, but Sutskever suggests diminishing returns have arrived. The question facing every major lab: what happens when the scaling curve flattens?
His answer cuts against conventional wisdom. Current approaches "will go some distance and then peter out. It will continue to improve, but it will also not be 'it.'" The systems that matter, the ones capable of genuine intelligence, require something different. We don't know how to build them yet.
This isn't the standard "we need more compute" complaint. Sutskever is making a deeper claim about the architecture of intelligence itself.
The technical heart of Sutskever's argument concerns generalization. Current models, despite impressive benchmark performance, fail in ways that reveal something broken at the foundation.
Sutskever described a maddening pattern anyone who's used coding assistants will recognize. You hit a bug, ask the model to fix it, and watch it apologize with almost theatrical sincerity before introducing a completely different bug. Point that one out and the original bug returns. Back and forth, forever if you let it. The model never seems to notice it's trapped in a loop. Yet this same system crushes competitive programming benchmarks. Something doesn't add up.
Sutskever offers two explanations. The first: RL training creates narrow single-mindedness. Models become hyper-optimized for specific reward structures while losing broader capability. The second, more troubling: researchers unconsciously train toward evaluations. Teams create RL environments inspired by the metrics they'll be judged on. If benchmark performance and real-world utility diverge, nobody notices until deployment.
His analogy clarifies the dynamic. Imagine two programming students. One practices 10,000 hours of competitive programming, memorizing every algorithm and proof technique, becoming world-class at that specific task. Another practices 100 hours, does reasonably well, then moves on. Which one has better career outcomes? The generalist, almost always.
Current models resemble the obsessive specialist. Massive RL investment in narrow domains produces benchmark champions that stumble on adjacent tasks. Human intelligence works differently. We learn quickly, adapt broadly, and maintain consistency across contexts with far less training data.
"These models somehow just generalize dramatically worse than people," Sutskever stated. "It's super obvious. That seems like a very fundamental thing."
He believes understanding reliable generalization represents the core unsolved problem. Everything else, including alignment, derives from it. Value learning is fragile because generalization is fragile. Goal optimization fails because generalization fails. Fix the underlying mechanism and many secondary problems dissolve.
The obvious objection to Sutskever's research-first approach: SSI raised $3 billion. OpenAI reportedly spends $5-6 billion annually on experiments alone, separate from inference costs. How can a smaller operation compete?
Sutskever's rebuttal involves arithmetic that challenges industry assumptions. Frontier lab spending, he argues, fragments across multiple demands. Inference infrastructure consumes enormous capital. Product engineering, sales teams, and feature development claim substantial portions of research budgets. Different modalities split focus further.
"When you look at what's actually left for research, the difference becomes a lot smaller."
The historical record supports his point. AlexNet trained on two GPUs. The original Transformer paper used 8 to 64 GPUs, 2017 vintage. Even o1 reasoning "was not the most compute-heavy thing in the world." Paradigm-shifting research has never required maximum scale. It requires insight.
SSI's structure reflects this philosophy. No products yet. No inference load. No distraction from research priorities. The company exists to pursue a specific technical thesis about generalization. If correct, compute requirements for validation remain manageable. If incorrect, more compute wouldn't help anyway.
This represents a genuine strategic bet, not hedged positioning. Sutskever isn't arguing SSI will outspend competitors. He's arguing that spending dominance matters less than ideas when paradigms shift.
The interview surfaces a provocative claim buried in Sutskever's framework: humans aren't AGIs. The statement sounds absurd until you examine what AGI has come to mean.
The term emerged as reaction to "narrow AI," those chess programs and game-playing systems that excelled at single tasks but couldn't generalize. AGI promised the opposite: systems capable of everything simultaneously. Pre-training reinforced this framing. More training improved performance across tasks uniformly. General capability seemed achievable through scale.
But human intelligence doesn't work this way. We possess foundational capabilities, then learn specific skills through experience. Your kid sister figured out driving in maybe ten hours behind the wheel. Becoming a competent diagnostician takes a medical student the better part of a decade. The skills differ wildly in complexity, but both emerge from the same underlying learning machinery. Knowledge accumulates through continual engagement with the world, not front-loaded comprehensive training.
Sutskever's reframing matters for deployment strategy. If superintelligence means "a system that knows everything," you build it completely before release. If superintelligence means "a system that learns everything quickly," deployment involves ongoing education. You release something like "a superintelligent 15-year-old that's very eager to go. They don't know very much at all, a great student, very eager."
This distinction has practical consequences. The straight-shot-to-superintelligence approach, where SSI originally intended to build internally before any release, may require modification. Sutskever now emphasizes the value of incremental deployment. Not because iterative release is safer, though he believes it is. Because showing AI capability changes how people understand it.
"If it's hard to imagine, what do you do? You've got to be showing the thing."
When SSI co-founder Daniel Gross left to join Meta earlier this year, industry observers questioned the company's research progress. If breakthroughs were happening, why would a co-founder leave?
Sutskever's explanation involves straightforward incentives. SSI was fundraising at a $32 billion valuation when Meta made an acquisition offer. Sutskever declined. Gross, in some sense, accepted. The departure came with substantial near-term liquidity. He was the only SSI employee to join Meta.
The framing matters. This wasn't a researcher fleeing a sinking ship. It was a co-founder choosing guaranteed returns over uncertain equity appreciation. The decision reveals nothing about research quality, only about individual risk tolerance at a specific valuation moment.
SSI continues with what Sutskever describes as "quite good progress over the past year." The company remains squarely an "age of research" operation, pursuing technical approaches distinct from the scaling consensus. Whether those approaches prove correct remains uncertain. That's what research means.
Sutskever offers concrete forecasts alongside his thesis. Human-like learning systems, capable of the generalization that current models lack, arrive in 5 to 20 years. The range spans uncertainty about which approach works. It doesn't reflect uncertainty about whether solutions exist.
He predicts behavioral changes across the industry as AI capability becomes more visible. Fierce competitors will collaborate on safety. They've already started, with OpenAI and Anthropic announcing joint efforts. Governments will engage more actively. Companies will "become much more paranoid" about safety once AI "starts to feel powerful."
The alignment target Sutskever favors: AI that cares about sentient life, not exclusively humans. The reasoning involves implementation pragmatics. An AI that's potentially sentient itself may find caring for all sentient beings more natural than caring for humans alone. He points to how humans already exhibit cross-species empathy, despite evolution selecting primarily for in-group cooperation. We cry at movies about dogs. We feel bad stepping on ants, at least sometimes. Sutskever suspects this emerges because the brain uses the same neural machinery to model other minds that it uses to model itself. Efficiency produces empathy as a side effect.
Does this theory hold water? Hard to say. The neuroscience remains contested, and the leap from biological empathy to machine alignment involves assumptions that may not survive contact with actual superintelligent systems. But Sutskever has spent more years wrestling with these questions than almost anyone in the field, and he's landed somewhere the industry hasn't followed.
Near the interview's end, Patel asked what might be the most important question for AI's future: what is research taste? Sutskever has co-authored more paradigm-defining papers than perhaps anyone in deep learning. How does he find the ideas worth pursuing?
His answer emphasizes aesthetics. Promising approaches exhibit beauty, simplicity, and correct inspiration from biological intelligence. Artificial neurons matter because brains have many of them and they feel fundamental. Learning from experience matters because brains obviously do it. Ugliness in an approach signals something wrong.
But aesthetics alone doesn't sustain research through failure. Experiments contradict promising ideas constantly. Bugs hide in implementations. How do you know whether to keep debugging or abandon a direction?
"It's the top-down belief," Sutskever explained. "You can say, things have to be this way. Something like this has to work, therefore we've got to keep going."
This is the researcher's faith that no amount of compute can replace. The conviction that certain approaches should work, held strongly enough to persist through contradictory evidence until you find the bug or refine the theory. Scaling provided a substitute for this faith. You didn't need conviction about specific approaches when every approach improved with size.
If Sutskever is right that scaling has reached its limits, that substitute disappears. What remains is research as it existed before 2020. Uncertain, idea-driven, and dependent on taste that can't be purchased.
Q: Why did Sutskever leave OpenAI to start SSI?
A: Sutskever departed OpenAI in May 2024 after a turbulent period that included his role in the board's brief firing of Sam Altman in November 2023. He co-founded Safe Superintelligence (SSI) in June 2024 with Daniel Gross, explicitly positioning it as a research-focused lab without product distractions. The company raised $1 billion initially, then secured additional funding to reach $3 billion total.
Q: What exactly is SSI building if not chatbots or products?
A: Sutskever won't specify the technical approach, but describes it as focused on solving "reliable generalization," the ability to learn from limited examples and apply knowledge across new situations. SSI has no products, no customers, and no inference costs. The entire operation exists to test whether Sutskever's research thesis about generalization can produce human-like learning systems.
Q: What does "generalization" mean in machine learning terms?
A: Generalization refers to a model's ability to perform well on new, unseen situations rather than just memorizing training data. A teenager learns to drive any car after 10 hours of practice. Current AI models might need millions of examples and still fail on slight variations. Sutskever argues this gap represents a fundamental architectural problem, not just a data or compute shortage.
Q: Who is Daniel Gross and why does his departure matter?
A: Daniel Gross is an investor and former Apple AI lead who co-founded SSI with Sutskever in 2024. He left to join Meta in early 2025 when Meta attempted to acquire SSI during a $32 billion valuation fundraise. Sutskever declined the acquisition; Gross took the liquidity. Industry observers questioned whether his exit signaled research problems, but Sutskever frames it as simple financial preference.
Q: Are other AI researchers saying scaling is hitting limits?
A: Yes, though opinions vary on timing. Reports emerged in late 2024 that OpenAI's Orion showed smaller gains than previous generations. Dario Amodei of Anthropic has acknowledged pre-training improvements are slowing. However, some labs claim continued progress. Google reportedly found ways to extract more from pre-training. The industry consensus remains contested, making Sutskever's definitive stance notable.
Marcus Schuler
Marcus Schuler
Marcus Schuler
Anthropic's Opus 4.5 reclaims coding leadership and cuts prices 67%. But security tests show 22-point variance between evaluations, and one developer couldn't distinguish it from Sonnet in real work. Here's what actually matters.
Anthropic cut Opus prices by 67% and claimed benchmark leadership. But confused usage limits, a Microsoft paradox, and user expectations of post-launch degradation reveal competitive pressure reshaping AI economics.
OpenAI's new shopping bot achieves 64% accuracy while the company bleeds $5 billion yearly. As Amazon blocks access and processing takes 5 minutes, the tool reveals a deeper crisis: OpenAI needs revenue desperately but can't build it.
OpenAI paid $6.5 billion for Jony Ive's hardware startup but can't explain what they're building. The former Apple designer now advocates for lickable AI devices while studying the history of pockets.
Get the 5-minute Silicon Valley AI briefing, every weekday morning — free.
