Global Advisors | Quantified Strategy Consulting

“The robustness of people is really staggering.” – Ilya Sutskever – Safe Superintelligence

This statement, made in his November 2025 conversation with Dwarkesh Patel, comes from someone uniquely positioned to make such judgments: co-founder and Chief Scientist of Safe Superintelligence Inc., former Chief Scientist at OpenAI, and co-author of AlexNet—the 2012 paper that launched the modern deep learning era.

Sutskever’s claim about robustness points to something deeper than mere durability or fault tolerance. He is identifying a distinctive quality of human learning: the ability to function effectively across radically diverse contexts, to self-correct without explicit external signals, to maintain coherent purpose and judgment despite incomplete information and environmental volatility, and to do all this with sparse data and limited feedback. These capacities are not incidental features of human intelligence. They are central to what makes human learning fundamentally different from—and vastly superior to—current AI systems.

Understanding what Sutskever means by robustness requires examining not just human capabilities but the specific ways in which AI systems are fragile by comparison. It requires recognising what humans possess that machines do not. And it requires understanding why this gap matters profoundly for the future of artificial intelligence.

What Robustness Actually Means: Beyond Mere Reliability

In engineering and systems design, robustness typically refers to a system’s ability to continue functioning when exposed to perturbations, noise, or unexpected conditions. A robust bridge continues standing despite wind, earthquakes, or traffic loads beyond its design specifications. A robust algorithm produces correct outputs despite noisy inputs or computational errors.

But human robustness operates on an entirely different plane. It encompasses far more than mere persistence through adversity. Human robustness includes:

1. Flexible adaptation across domains: A teenager learns to drive after ten hours of practice and then applies principles of vehicle control, spatial reasoning, and risk assessment to entirely new contexts—motorcycles, trucks, parking in unfamiliar cities. The principles transfer because they have been learned at a level of abstraction and generality that allows principled application to novel situations.
2. Self-correction without external reward: A learner recognises when they have made an error not through explicit feedback but through an internal sense of rightness or wrongness—what Sutskever terms a “value function” and what we experience as intuition, confidence, or unease. A pianist knows immediately when they have struck a wrong note; they do not need external evaluation. This internal evaluative system enables rapid, efficient learning.
3. Judgment under uncertainty: Humans routinely make decisions with incomplete information, tolerating ambiguity whilst maintaining coherent action. A teenager drives defensively not because they can compute precise risk probabilities but because they possess an internalized model of danger, derived from limited experience but somehow applicable to novel situations.
4. Stability across time scales: Human goals, values, and learning integrate across vastly different temporal horizons. A person may pursue long-term education goals whilst adapting to immediate challenges, and these different time scales cohere into a unified, purposeful trajectory. This temporal integration is largely absent from current AI systems, which optimise for immediate reward signals or fixed objectives.
5. Learning from sparse feedback: Humans learn from remarkably little data. A child sees a dog once or twice and thereafter recognises dogs in novel contexts, even in stylised drawings or unfamiliar breeds. This learning from sparse examples contrasts sharply with AI systems requiring thousands or millions of examples to achieve equivalent recognition.

This multifaceted robustness is what Sutskever identifies as “staggering”—not because it is strong but because it operates across so many dimensions simultaneously whilst remaining stable, efficient, and purposeful.

The Fragility of Current AI: Why Models Break

The contrast becomes clear when examining where current AI systems are fragile. Sutskever frequently illustrates this through the “jagged behaviour” problem: models that perform superhuman on benchmarks yet fail in elementary ways during real-world deployment.

A language model can score in the 88th percentile on the bar examination yet, when asked to debug code, introduces new errors whilst fixing previous ones. It cycles between mistakes even when provided clear feedback. It lacks the internal evaluative sense that tells a human programmer, “This approach is leading nowhere; I should try something different.” The model lacks robust value functions—internal signals that guide learning and action.

This fragility manifests across multiple dimensions:

1. Distribution shift fragility: Models trained on one distribution of data often fail dramatically when confronted with data that differs from training distribution, even slightly. A vision system trained on images with certain lighting conditions fails on images with different lighting. A language model trained primarily on Western internet text struggles with cultural contexts it has not heavily encountered. Humans, by contrast, maintain competence across remarkable variation—different languages, accents, cultural contexts, lighting conditions, perspectives.
2. Benchmark overfitting: Contemporary AI systems achieve extraordinary performance on carefully constructed evaluation tasks yet fail at the underlying capability the benchmark purports to measure. This occurs because models have been optimised (through reinforcement learning) specifically to perform well on benchmarks rather than to develop robust understanding. Sutskever has noted that this reward hacking is often unintentional—companies genuinely seeking to improve models inadvertently create RL environments that optimise for benchmark performance rather than genuine capability.
3. Lack of principled abstraction: Models often memorise patterns rather than developing principled understanding. This manifests as inability to apply learned knowledge to genuinely novel contexts. A model may solve thousands of addition problems yet fail on a slightly different formulation it has not encountered. A human, having understood addition as a principle, applies it to any context where addition is relevant.
4. Absence of internal feedback mechanisms: Current reinforcement learning typically provides feedback only at the end of long trajectories. A model can pursue 1,000 steps of reasoning down an unpromising path, only to receive a training signal after the trajectory completes. Humans, by contrast, possess continuous internal feedback—emotions, intuition, confidence levels—that signal whether reasoning is productive or should be redirected. This enables far more efficient learning.

The Value Function Hypothesis: Emotions as Robust Learning Machinery

Sutskever’s analysis points toward a crucial hypothesis: human robustness depends fundamentally on value functions—internal mechanisms that provide continuous, robust evaluation of states and actions.

In machine learning, a value function is a learned estimate of expected future reward or utility from a given state. In human neurobiology, value functions are implemented, Sutskever argues, through emotions and affective states. Fear signals danger. Confidence signals competence. Boredom signals that current activity is unproductive. Satisfaction signals that effort has succeeded. These emotional states, which evolution has refined over millions of years, serve as robust evaluative signals that guide learning and behaviour.

Sutskever illustrates this with a striking neurological case: a person who suffered brain damage affecting emotional processing. Despite retaining normal IQ, puzzle-solving ability, and articulate cognition, this person became radically incapable of making even trivial decisions. Choosing which socks to wear would take hours. Financial decisions became catastrophically poor. This person could think but could not effectively decide or act—suggesting that emotions (and the value functions they implement) are not peripheral to human cognition but absolutely central to effective agency.

What makes human value functions particularly robust is their simplicity and stability. They are not learned during a person’s lifetime through explicit training. They are evolved, hard-coded by billions of years of biological evolution into neural structures that remain remarkably consistent across human populations and contexts. A person experiences hunger, fear, social connection, and achievement similarly whether in ancient hunter-gatherer societies or modern industrial ones—because these value functions were shaped by evolutionary pressures that remained relatively stable.

This evolutionary hardcoding of value functions may be crucial to human learning robustness. Imagine trying to teach a child through explicit reward signals alone: “Do this task and receive points; optimise for points.” This would be inefficient and brittle. Instead, humans learn through value functions that are deeply embedded, emotionally weighted, and robust across contexts. A child learns to speak not through external reward optimisation but through intrinsic motivation—social connection, curiosity, the inherent satisfaction of communication. These motivations persist across contexts and enable robust learning.

Current AI systems largely lack this. They optimise for explicitly defined reward signals or benchmark metrics. These are fragile by comparison—vulnerable to reward hacking, overfitting, distribution shift, and the brittle transfer failures Sutskever observes.

Why This Matters Now: The Transition Point

Sutskever’s observation about human robustness arrives at a precise historical moment. As of November 2025, the AI industry is transitioning from what he terms the “age of scaling” (2020–2025) to what will be the “age of research” (2026 onward). This transition is driven by recognition that scaling alone is reaching diminishing returns. The next advances will require fundamental breakthroughs in understanding how to build systems that learn and adapt robustly—like humans do.

This creates an urgent research agenda: How do you build AI systems that possess human-like robustness? This is not a question that scales with compute or data. It is a research question—requiring new architectures, learning algorithms, training procedures, and conceptual frameworks.

Sutskever’s identification of robustness as the key distinguishing feature of human learning sets the research direction for the next phase of AI development. The question is not “how do we make bigger models” but “how do we build systems with value functions that enable efficient, self-correcting, context-robust learning?”

The Research Frontier: Leading Theorists Addressing Robustness

Antonio Damasio: The Somatic Marker Hypothesis

Antonio Damasio, neuroscientist at USC and authority on emotion and decision-making, has developed the somatic marker hypothesis—a framework explaining how emotions serve as rapid evaluative signals that guide decisions and learning. Damasio’s work provides neuroscientific grounding for Sutskever’s hypothesis that value functions (implemented as emotions) are central to effective agency. Damasio’s case studies of patients with emotional processing deficits closely parallel Sutskever’s neurological example—demonstrating that emotional value functions are prerequisites for robust, adaptive decision-making.

Judea Pearl: Causal Models and Robust Reasoning

Judea Pearl, pioneer in causal inference and probabilistic reasoning, has argued that correlation-based learning has fundamental limits and that robust generalisation requires learning causal structure—the underlying relationships between variables that remain stable across contexts. Pearl’s work suggests that human robustness derives partly from learning causal models rather than mere patterns. When a human understands how something works (causally), that understanding transfers to novel contexts. Current AI systems, lacking robust causal models, fail at transfer—a key component of robustness.

Karl Friston: The Free Energy Principle

Karl Friston, neuroscientist at University College London, has developed the free energy principle—a unified framework explaining how biological systems, including humans, maintain robustness by minimising prediction error and maintaining models of their environment and themselves. The principle suggests that what makes humans robust is not fixed programming but a general learning mechanism that continuously refines internal models to reduce surprise. This framework has profound implications for building robust AI: rather than optimising for external rewards, systems should optimise for maintaining accurate models of reality, enabling principled generalisation.

Stuart Russell: Learning Under Uncertainty and Value Alignment

Stuart Russell, UC Berkeley’s leading AI safety researcher, has emphasised that robust AI systems must remain genuinely uncertain about objectives and learn from interaction rather than operating under fixed goal specifications. Russell’s work suggests that rigidity about objectives makes systems fragile—vulnerable to reward hacking and context-specific failure. Robustness requires systems that maintain epistemic humility and adapt their understanding of what matters based on continued learning. This directly parallels how human value systems are robust: they are not brittle doctrines but evolving frameworks that integrate experience.

Demis Hassabis and DeepMind’s Continual Learning Research

Demis Hassabis, CEO of DeepMind, has invested substantial effort into systems that learn continuously from environmental interaction rather than through discrete offline training phases. DeepMind’s research on continual reinforcement learning, meta-learning, and adaptive systems reflects the insight that robustness emerges not from static pre-training but from ongoing interaction with environments—enabling systems to refine their models and value functions continuously. This parallels human learning, which is fundamentally continual rather than episodic.

Yann LeCun: Self-Supervised Learning and World Models

Yann LeCun, Meta’s Chief AI Scientist, has advocated for learning approaches that enable systems to build internal models of how the world works—what he terms world models—through self-supervised learning. LeCun argues that robust generalisation requires systems that understand causal structure and dynamics, not merely correlations. His work on self-supervised learning suggests that systems trained to predict and model their environments develop more robust representations than systems optimised for specific external tasks.

The Evolutionary Basis: Why Humans Have Robust Value Functions

Understanding human robustness requires appreciating why evolution equipped humans with sophisticated, stable value function systems.

For millions of years, humans and our ancestors faced fundamentally uncertain environments. The reward signals available—immediate sensory feedback, social acceptance, achievement, safety—needed to guide learning and behaviour across vast diversity of contexts. Evolution could not hard-code specific solutions for every possible situation. Instead, it encoded general-purpose value functions—emotions and motivational states—that would guide adaptive behaviour across contexts.

Consider fear. Fear is a robust value function signal that something is dangerous. This signal evolved in environments full of predators and hazards. Yet the same fear response that protected ancestral humans from predators also keeps modern humans safe from traffic, heights, and social rejection. The value function is robust because it operates on a general principle—danger—rather than specific memorised hazards.

Similarly, social connection, curiosity, achievement, and other human motivations evolved as general-purpose signals that, across millions of years, correlated with survival and reproduction. They remain remarkably stable across radically different modern contexts—different cultures, technologies, and social structures—because they operate at a level of abstraction robust to context change.

Current AI systems, by contrast, lack this evolutionary heritage. They are trained from scratch, often on specific tasks, with reward signals explicitly engineered for those tasks. These reward signals are fragile by comparison—vulnerable to distribution shift, overfitting, and context-specificity.

Implications for Safe AI Development

Sutskever’s emphasis on human robustness carries profound implications for safe AI development. Robust systems are safer systems. A system with genuine value functions—robust internal signals about what matters—is less vulnerable to reward hacking, specification gaming, or deployment failures. A system that learns continuously and maintains epistemic humility is more likely to remain aligned as its capabilities increase.

Conversely, current AI systems’ lack of robustness is dangerous. Systems optimised for narrow metrics can fail catastrophically when deployed in novel contexts. Systems lacking robust value functions cannot self-correct or maintain appropriate caution. Systems that cannot learn from deployment feedback remain brittle.

Building AI systems with human-like robustness is therefore not merely an efficiency question—though efficiency matters greatly. It is fundamentally a safety question. The development of robust value functions, continual learning capabilities, and general-purpose evaluative mechanisms is central to ensuring that advanced AI systems remain beneficial as they become more powerful.

The Research Direction: From Scaling to Robustness

Sutskever’s observation that “the robustness of people is really staggering” reorients the entire research agenda. The question is no longer primarily “how do we scale?” but “how do we build systems with robust value functions, efficient learning, and genuine adaptability across contexts?”

This requires:

• Architectural innovation: New neural network structures that embed or can learn robust evaluative mechanisms—value functions analogous to human emotions.
• Training methodology: Learning procedures that enable systems to develop genuine self-correction capabilities, learn from sparse feedback, and maintain robustness across distribution shift.
• Theoretical understanding: Deeper mathematical and conceptual frameworks explaining what makes value functions robust and how to implement them in artificial systems.
• Integration of findings from neuroscience, evolutionary biology, and decision theory: Drawing on multiple fields to understand the principles underlying human robustness and translating them into machine learning.

Conclusion: Robustness as the Frontier

When Sutskever identifies human robustness as “staggering,” he is not offering admiration but diagnosis. He is pointing out that current AI systems fundamentally lack what makes humans effective learners: robust value functions, efficient learning from sparse feedback, genuine self-correction, and adaptive generalisation across contexts.

The next era of AI research—the age of research beginning in 2026—will be defined largely by attempts to solve this problem. The organisation or research group that successfully builds AI systems with human-like robustness will not merely have achieved technical progress. They will have moved substantially closer to systems that learn efficiently, generalise reliably, and remain aligned to human values even as they become more capable.

Human robustness is not incidental. It is fundamental—the quality that makes human learning efficient, adaptive, and safe. Replicating it in artificial systems represents the frontier of AI research and development.

“These models somehow just generalize dramatically worse than people. It’s super obvious. That seems like a very fundamental thing.” – Ilya Sutskever – Safe Superintelligence

Sutskever, as co-founder and Chief Scientist of Safe Superintelligence Inc. (SSI), has emerged as one of the most influential voices in AI strategy and research direction. His trajectory illustrates the depth of his authority: co-author of AlexNet (2012), the paper that ignited the deep learning revolution; Chief Scientist at OpenAI during the development of GPT-2 and GPT-3; and now directing a $3 billion research organisation explicitly committed to solving the generalisation problem rather than pursuing incremental scaling.

His assertion about generalisation deficiency is not rhetorical flourish. It represents a fundamental diagnostic claim about why current AI systems, despite superhuman performance on benchmarks, remain brittle, unreliable, and poorly suited to robust real-world deployment. Understanding this claim requires examining what generalisation actually means, why it matters, and what the gap between human and AI learning reveals about the future of artificial intelligence.

What Generalisation Means: Beyond Benchmark Performance

Generalisation, in machine learning, refers to the ability of a system to apply knowledge learned in one context to novel, unfamiliar contexts it has not explicitly encountered during training. A model that generalises well can transfer principles, patterns, and capabilities across domains. A model that generalises poorly becomes a brittle specialist—effective within narrow training distributions but fragile when confronted with variation, novelty, or real-world complexity.

The crisis Sutskever identifies is this: contemporary large language models and frontier AI systems achieve extraordinary performance on carefully curated evaluation tasks and benchmarks. GPT-4 scores in the 88th percentile of the bar exam. O1 solves competition mathematics problems at elite levels. Yet these same systems, when deployed into unconstrained real-world workflows, exhibit what Sutskever terms “jagged” behaviour—they repeat errors, introduce new bugs whilst fixing previous ones, cycle between mistakes even with clear corrective feedback, and fail in ways that suggest fundamentally incomplete understanding rather than mere data scarcity.

This paradox reveals a hidden truth: benchmark performance and deployment robustness are not tightly coupled. An AI system can memorise, pattern-match, and perform well on evaluation metrics whilst failing to develop the kind of flexible, transferable understanding that enables genuine competence.

The Sample Efficiency Question: Orders of Magnitude of Difference

Underlying the generalisation crisis is a more specific puzzle: sample efficiency. Why does it require vastly more training data for AI systems to achieve competence in a domain than it takes humans?

A human child learns to recognise objects through a few thousand exposures. Contemporary vision models require millions. A teenager learns to drive in approximately ten hours of practice; AI systems struggle to achieve equivalent robustness with orders of magnitude more training. A university student learns to code, write mathematically, and reason about abstract concepts—domains that did not exist during human evolutionary history—with remarkably few examples and little explicit feedback.

This disparity points to something fundamental: humans possess not merely better priors or more specialised knowledge, but better general-purpose learning machinery. The principle underlying human learning efficiency remains largely unexpressed in mathematical or computational terms. Current AI systems lack it.

Sutskever’s diagnostic claim is that this gap reflects not engineering immaturity or the need for more compute, but the absence of a conceptual breakthrough—a missing principle of how to build systems that learn as efficiently as humans do. The implication is stark: you cannot scale your way out of this problem. More data and more compute, applied to existing methodologies, will not solve it. The bottleneck is epistemic, not computational.

Why Current Models Fail at Generalisation: The Competitive Programming Analogy

Sutskever illustrates the generalisation problem through an instructive analogy. Imagine two competitive programmers:

Student A dedicates 10,000 hours to competitive programming. They memorise every algorithm, every proof technique, every problem pattern. They become exceptionally skilled within competitive programming itself—one of the very best.

Student B spends only 100 hours on competitive programming but develops deeper, more flexible understanding. They grasp underlying principles rather than memorising solutions.

When both pursue careers in software engineering, Student B typically outperforms Student A. Why? Because Student A has optimised for a narrow domain and lacks the flexible transfer of understanding that Student B developed through lighter but more principled engagement.

Current frontier AI models, in Sutskever’s assessment, resemble Student A. They are trained on enormous quantities of narrowly curated data—competitive programming problems, benchmark evaluation tasks, reinforcement learning environments explicitly designed to optimise for measurable performance. They have been “over-trained” on carefully optimised domains but lack the flexible, generalised understanding that enables robust performance in novel contexts.

This over-optimisation problem is compounded by a subtle but crucial factor: reinforcement learning optimisation targets. Companies designing RL training environments face substantial degrees of freedom in how to construct reward signals. Sutskever observes that there is often a systematic bias: RL environments are subtly shaped to ensure models perform well on public benchmarks at release time, creating a form of unintentional reward hacking where the system becomes highly tuned to evaluation metrics rather than genuinely robust to real-world variation.

The Deeper Problem: Pre-Training’s Limits and RL’s Inefficiency

The generalisation crisis reflects deeper structural issues within contemporary AI training paradigms.

Pre-training’s opacity: Large-scale language model pre-training—trained on internet text data—provides models with an enormous foundation of patterns. Yet the way models rely on this pre-training data is poorly understood. When a model fails, it is unclear whether the failure reflects insufficient statistical support in the training distribution or whether something more fundamental is missing. Pre-training provides scale but at the cost of reasoning about what has actually been learned.

RL’s inefficiency: Current reinforcement learning approaches provide training signals only at the end of long trajectories. If a model spends thousands of steps reasoning about a problem and arrives at a dead end, it receives no signal until the trajectory completes. This is computationally wasteful. A more efficient learning system would provide intermediate evaluative feedback—signals that say, “this direction of reasoning is unpromising; abandon it now rather than after 1,000 more steps.” Sutskever hypothesises that this intermediate feedback mechanism—what he terms a “value function” and what evolutionary biology has encoded as emotions—is crucial to sample-efficient learning.

The gap between how humans and current AI systems learn suggests that human learning operates on fundamentally different principles: continuous, intermediate evaluation; robust internal models of progress and performance; the ability to self-correct and redirect effort based on internal signals rather than external reward.

Generalisation as Proof of Concept: What Human Learning Reveals

A critical move in Sutskever’s argument is this: the fact that humans generalise vastly better than current AI systems is not merely an interesting curiosity—it is proof that better generalisation is achievable. The existence of human learners demonstrates, in principle, that a learning system can operate with orders of magnitude less data whilst maintaining superior robustness and transfer capability.

This reframes the research challenge. The question is no longer whether better generalisation is possible (humans prove it is) but rather what principle or mechanism underlies it. This principle could arise from:

• Architectural innovations: new ways of structuring neural networks that embody better inductive biases for generalisation
• Learning algorithms: different training procedures that more efficiently extract principles from limited data
• Value function mechanisms: intermediate feedback systems that enable more efficient learning trajectories
• Continual learning frameworks: systems that learn continuously from interaction rather than through discrete offline training phases

What matters is that Sutskever’s claim shifts the research agenda from “get more compute” to “discover the missing principle.”

The Strategic Implications: Why This Matters Now

Sutskever’s diagnosis, articulated in November 2025, arrives at a crucial moment. The AI industry has operated under the “age of scaling” paradigm since approximately 2020. During this period, the scaling laws discovered by OpenAI and others suggested a remarkably reliable relationship: larger models trained on more data with more compute reliably produced better performance.

This created a powerful strategic imperative: invest capital in compute, acquire data, build larger systems. The approach was low-risk from a research perspective because the outcome was relatively predictable. Companies could deploy enormous resources confident they would yield measurable returns.

By 2025, however, this model shows clear strain. Data is approaching finite limits. Computational resources, whilst vast, are not unlimited, and marginal returns diminish. Most importantly, the question has shifted: would 100 times more compute actually produce a qualitative transformation or merely incremental improvement? Sutskever’s answer is clear: the latter. This fundamentally reorients strategic thinking. If 100x scaling yields only incremental gains, the bottleneck is not compute but ideas. The competitive advantage belongs not to whoever can purchase the most GPUs but to whoever discovers the missing principle of generalisation.

Leading Theorists and Related Research Programs

Yann LeCun: World Models and Causal Learning

Yann LeCun, Meta’s Chief AI Scientist and a pioneer of deep learning, has long emphasized that current supervised learning approaches are fundamentally limited. His work on “world models”—internal representations that capture causal structure rather than mere correlation—points toward learning mechanisms that could enable better generalisation. LeCun’s argument is that humans learn causal models of how the world works, enabling robust generalisation because causal understanding is stable across contexts in a way that statistical correlation is not.

Geoffrey Hinton: Neuroscience-Inspired Learning

Geoffrey Hinton, recipient of the 2024 Nobel Prize in Physics for foundational deep learning work, has increasingly emphasized that neuroscience holds crucial clues for improving AI learning efficiency. His recent work on biological plausibility and learning mechanisms reflects conviction that important principles of how neural systems efficiently extract generalised understanding remain undiscovered. Hinton has expressed support for Sutskever’s research agenda, recognizing that the next frontier requires fundamental conceptual breakthroughs rather than incremental scaling.

Stuart Russell: Learning Under Uncertainty

Stuart Russell, UC Berkeley’s leading AI safety researcher, has articulated that robust AI alignment requires systems that remain genuinely uncertain about objectives and learn from interaction. This aligns with Sutskever’s emphasis on continual learning. Russell’s work highlights that systems designed to optimise fixed objectives without capacity for ongoing learning and adjustment tend to produce brittle, misaligned outcomes—a dynamic that improves when systems maintain epistemic humility and learn continuously.

Demis Hassabis and DeepMind’s Continual Learning Research

Demis Hassabis, CEO of DeepMind, has invested substantial research effort into systems that learn continually from environmental interaction rather than through discrete offline training phases. DeepMind’s work on continual reinforcement learning, meta-learning, and systems that adapt to new tasks reflects recognition that learning efficiency depends on how feedback is structured and integrated over time—not merely on total data quantity.

Judea Pearl: Causality and Abstraction

Judea Pearl, pioneering researcher in causal inference and probabilistic reasoning, has long argued that correlation-based learning has fundamental limits and that causal reasoning is necessary for genuine understanding and generalisation. His work on causal models and graphical representation of dependencies provides theoretical foundations for why systems that learn causal structure (rather than mere patterns) achieve better generalisation across domains.

The Research Agenda Going Forward

Sutskever’s claim that generalisation is the “very fundamental thing” reorients the entire research agenda. This shift has profound implications:

From scaling to methodology: Research emphasis moves from “how do we get more compute” to “what training procedures, architectural innovations, or learning algorithms enable human-like generalisation?”

From benchmarks to robustness: Evaluation shifts from benchmark performance to deployment reliability—how systems perform on novel, unconstrained tasks rather than carefully curated evaluations.

From monolithic pre-training to continual learning: The training paradigm shifts from discrete offline phases (pre-train, then RL, then deploy) toward systems that learn continuously from real-world interaction.

From scale as differentiator to ideas as differentiator: Competitive advantage in AI development becomes less about resource concentration and more about research insight—the organisation that discovers better generalisation principles gains asymmetric advantage.

The Deeper Question: What Humans Know That AI Doesn’t

Beneath Sutskever’s diagnostic claim lies a profound question: What do humans actually know about learning that AI systems don’t yet embody?

Humans learn efficiently because they:

• Develop internal models of their own performance and progress (value functions)
• Self-correct through continuous feedback rather than awaiting end-of-trajectory rewards
• Transfer principles flexibly across domains rather than memorising domain-specific patterns
• Learn from remarkably few examples through principled understanding rather than statistical averaging
• Integrate feedback across time scales and contexts in ways that build robust, generalised knowledge

These capabilities do not require superhuman intelligence or extraordinary cognitive resources. A fifteen-year-old possesses them. Yet current AI systems, despite vastly larger parameter counts and more data, lack equivalent ability.

This gap is not accidental. It reflects that current AI development has optimised for the wrong targets—benchmark performance rather than genuine generalisation, scale rather than efficiency, memorisation rather than principled understanding. The next breakthrough requires not more of the same but fundamentally different approaches.

Conclusion: The Shift from Scaling to Discovery

Sutskever’s assertion that “these models somehow just generalize dramatically worse than people” is, at first glance, an observation of inadequacy. But reframed, it is actually a statement of profound optimism about what remains to be discovered. The fact that humans achieve vastly better generalisation proves that better generalisation is possible. The task ahead is not to accept poor generalisation as inevitable but to discover the principle that enables human-like learning efficiency.

This diagnostic shift—from “we need more compute” to “we need better understanding of generalisation”—represents the intellectual reorientation of AI research in 2025 and beyond. The age of scaling is ending not because scaling is impossible but because it has approached its productive limits. The age of research into fundamental learning principles is beginning. What emerges from this research agenda may prove far more consequential than any previous scaling increment.

“Is the belief really, ‘Oh, it’s so big, but if you had 100x more, everything would be so different?’ It would be different, for sure. But is the belief that if you just 100x the scale, everything would be transformed? I don’t think that’s true. So it’s back to the age of research again, just with big computers.” – Ilya Sutskever – Safe Superintelligence

Ilya Sutskever stands as one of the most influential figures in modern artificial intelligence—a scientist whose work has fundamentally shaped the trajectory of deep learning over the past decade. As co-author of the seminal 2012 AlexNet paper, he helped catalyse the deep learning revolution that transformed machine vision and launched the contemporary AI era. His influence extends through his role as Chief Scientist at OpenAI, where he played a pivotal part in developing GPT-2 and GPT-3, the models that established large-scale language model pre-training as the dominant paradigm in AI research.

In late 2024, Sutskever departed OpenAI and co-founded Safe Superintelligence Inc. (SSI) alongside Daniel Gross and Daniel Levy, positioning the company as the world’s “first straight-shot SSI lab”—an organisation with a single focus: developing safe superintelligence without distraction from product development or revenue generation. The company has since raised $3 billion and reached a $32 billion valuation, reflecting investor confidence in Sutskever’s strategic vision and reputation.

The Context: The Exhaustion of Scaling

Sutskever’s quoted observation emerges from a moment of genuine inflection in AI development. For roughly five years—from 2020 to 2025—the AI industry operated under what he terms the “age of scaling.” This era was defined by a simple, powerful insight: that scaling pre-training data, computational resources, and model parameters yielded predictable improvements in model performance. Organisations could invest capital with low perceived risk, knowing that more compute plus more data plus larger models would reliably produce measurable gains.

This scaling paradigm was extraordinarily productive. It yielded GPT-3, GPT-4, and an entire generation of frontier models that demonstrated capabilities that astonished both researchers and the public. The logic was elegant: if you wanted better AI, you simply scaled the recipe. Sutskever himself was instrumental in validating this approach. The word “scaling” became conceptually magnetic, drawing resources, attention, and organisational focus toward a single axis of improvement.

Yet by 2024–2025, that era began showing clear signs of exhaustion. Data is finite—the amount of high-quality training material available on the internet is not infinite, and organisations are rapidly approaching meaningful constraints on pre-training data supply. Computational resources, whilst vast, are not unlimited, and the economic marginal returns on compute investment have become less obvious. Most critically, the empirical question has shifted: if current frontier labs have access to extraordinary computational resources, would 100 times more compute actually produce a qualitative transformation in capabilities, or merely incremental improvement?

Sutskever’s answer is direct: incremental, not transformative. This reframing is consequential because it redefines where the bottleneck actually lies. The constraint is no longer the ability to purchase more GPUs or accumulate more data. The constraint is ideas—novel technical approaches, new training methodologies, fundamentally different recipes for building AI systems.

The Jaggedness Problem: Theory Meeting Reality

One critical observation animates Sutskever’s thinking: a profound disconnect between benchmark performance and real-world robustness. Current models achieve superhuman performance on carefully constructed evaluation tasks—yet in deployment, they exhibit what Sutskever calls “jagged” behaviour. They repeat errors, introduce new bugs whilst fixing old ones, and cycle between mistakes even when given clear corrective feedback.

This apparent paradox suggests something deeper than mere data or compute insufficiency. It points to inadequate generalisation—the inability to transfer learning from narrow, benchmark-optimised domains into the messy complexity of real-world application. Sutskever frames this through an analogy: a competitive programmer who practises 10,000 hours on competition problems will be highly skilled within that narrow domain but often fails to transfer that knowledge flexibly to broader engineering challenges. Current models, in his assessment, resemble that hyper-specialised competitor rather than the flexible, adaptive learner.

The Core Insight: Generalisation Over Scale

The central thesis animating Sutskever’s work at SSI—and implicit in his quote—is that human-like generalisation and learning efficiency represent a fundamentally different ML principle than scaling, one that has not yet been discovered or operationalised within contemporary AI systems.

Humans learn with orders of magnitude less data than large models yet generalise far more robustly to novel contexts. A teenager learns to drive in roughly ten hours of practice; current AI systems struggle to acquire equivalent robustness with vastly more training data. This is not because humans possess specialised evolutionary priors for driving (a recent activity that evolution could not have optimized for); rather, it suggests humans employ a more general-purpose learning principle that contemporary AI has not yet captured.

Sutskever hypothesises that this principle is connected to what he terms “value functions”—internal mechanisms akin to emotions that provide continuous, intermediate feedback on actions and states, enabling more efficient learning than end-of-trajectory reward signals alone. Evolution appears to have hard-coded robust value functions—emotional and evaluative systems—that make humans viable, adaptive agents across radically different environments. Whether an equivalent principle can be extracted purely from pre-training data, rather than built into learning architecture, remains uncertain.

The Leading Theorists and Related Work

Yann LeCun and Data Efficiency

Yann LeCun, Meta’s Chief AI Scientist and a pioneer of deep learning, has long emphasised the importance of learning efficiency and the role of what he terms “world models” in understanding how agents learn causal structure from limited data. His work highlights that human vision achieves remarkable robustness from developmental data scarcity—children recognise cars after seeing far fewer exemplars than AI systems require—suggesting that the brain employs inductive biases or learning principles that current architectures lack.

Geoffrey Hinton and Neuroscience-Inspired AI

Geoffrey Hinton, winner of the 2024 Nobel Prize in Physics for his work on deep learning, has articulated concerns about AI safety and expressed support for Sutskever’s emphasis on fundamentally rethinking how AI systems learn and align. Hinton’s career-long emphasis on biologically plausible learning mechanisms—from Boltzmann machines to capsule networks—reflects a conviction that important principles for efficient learning remain undiscovered and that neuroscience offers crucial guidance.

Stuart Russell and Alignment Through Uncertainty

Stuart Russell, UC Berkeley’s leading AI safety researcher, has emphasised that robust AI alignment requires systems that remain genuinely uncertain about human values and continue learning from interaction, rather than attempting to encode fixed objectives. This aligns with Sutskever’s thesis that safe superintelligence requires continual learning in deployment rather than monolithic pre-training followed by fixed RL optimisation.

Demis Hassabis and Continual Learning

Demis Hassabis, CEO of DeepMind and a co-developer of AlphaGo, has invested significant research effort into systems that learn continually rather than through discrete training phases. This work recognises that biological intelligence fundamentally involves interaction with environments over time, generating diverse signals that guide learning—a principle SSI appears to be operationalising.

The Paradigm Shift: From Offline to Online Learning

Sutskever’s thinking reflects a broader intellectual shift visible across multiple frontiers of AI research. The dominant pre-training + RL framework assumes a clean separation: a model is trained offline on fixed data, then post-trained with reinforcement learning, then deployed. Increasingly, frontier researchers are questioning whether this separation reflects how learning should actually work.

His articulation of “age of research” signals a return to intellectual plurality and heterodox experimentation—the opposite of the monoculture that scaling paradigm created. When everyone is racing to scale the same recipe, innovation becomes incremental. When new recipes are required, diversity of approach becomes an asset rather than liability.

The Stakes and Implications

This reframing carries significant strategic implications. If the bottleneck is truly ideas rather than compute, then smaller, more cognitively coherent organisations with clear intellectual direction may outpace larger organisations constrained by product commitments, legacy systems, and organisational inertia. If the key innovation is a new training methodology—one that achieves human-like generalisation through different mechanisms—then the first organisation to discover and validate it may enjoy substantial competitive advantage, not through superior resources but through superior understanding.

Equally, this framing challenges the common assumption that AI capability is primarily a function of computational spend. If methodological innovation matters more than scale, the future of AI leadership becomes less a question of capital concentration and more a question of research insight—less about who can purchase the most GPUs, more about who can understand how learning actually works.

Sutskever’s quote thus represents not merely a rhetorical flourish but a fundamental reorientation of strategic thinking about AI development. The age of confident scaling is ending. The age of rigorous research into the principles of generalisation, sample efficiency, and robust learning has begun.