Why AI Is So Much Less Sample-Efficient Than Humans

AI models need somewhere between thousands to a million times more data than humans to learn the same skills — and that staggering gap is one of the most underappreciated facts about the current state of artificial intelligence. While frontier models dazzle us with their capabilities, the sheer volume of data required to produce those capabilities tells a very different story about how close we actually are to human-like intelligence. Understanding why AI models need so much more data than humans isn't just an academic exercise — it shapes everything from which jobs AI will automate first to whether an intelligence explosion is even possible.

What Is Sample Efficiency and Why Does It Matter for AI?

One useful definition of intelligence is sample efficiency: how much data does an agent need in a given domain to operate fluently and competently? By this measure, humans are extraordinarily intelligent. A teenager can learn to drive with roughly 20 hours of practice. A person can pick up a new programming language after already knowing one in days or weeks. We generalize aggressively from very few examples.

AI models, by contrast, are remarkably sample-inefficient. Despite years of breathtaking progress on benchmarks, it's not clear that we've made meaningful gains in training sample efficiency. What has actually improved is the data distribution — its breadth, its quality, and the sheer scale of compute used to generate it. The models are getting better primarily because we're feeding them more and better data, not because they've become more efficient learners.

How Much Data Are Frontier AI Models Actually Trained On?

The numbers here are almost incomprehensible. Consider: if a person absorbs roughly 2,000 words per hour from birth to adulthood, they'll encounter around 200 million tokens in their lifetime. Frontier AI models are trained on tens to hundreds of trillions of tokens. That's a gap of close to a millionfold.

And it's not just pre-training data. The expert human-generated data used to fine-tune and align these models is extraordinarily domain-specific and labor-intensive. Job boards like Mercor and Surge list roles for Word document specialists, M&A legal experts, and management consultants — all generating painstakingly crafted example completions and rubrics. Each skill a model learns corresponds to hundreds of human experts producing training examples. The data labeling industry generating these expert labels is already worth billions annually, and growing fast.

The correct mental model for these AI systems isn't a brilliant student who has absorbed many subjects. It's closer to a Frankenstein's monster stitched together from a billion carefully constructed grafts of human expertise. Impressive? Absolutely. But fundamentally different from how human learning works.

How Does Reinforcement Learning Actually Make AI Better?

Much of the recent leap in AI capability — particularly in reasoning and coding — has come from reinforcement learning (RL). But it's worth being precise about what RL is actually doing here. You can think of RL as a form of synthetic data generation: you dump enormous compute against a verifier or an LLM-as-judge to identify which model outputs are correct, then train the model to predict those correct rollouts — much the same way it was trained to predict the next word in internet text.

For this to work at all, the model must already have some prior probability of stumbling onto the correct solution. That's why the pre-training corpus of human expert data is so critical — without it, RL has nothing to build on. And the compute cost is staggering: techniques like GRPO require the model to generate hundreds to thousands of rollouts per task just to solve the credit assignment problem. A human student might practice a textbook problem once or twice. These models grind through the same territory at massive scale.

The implication is significant: open-source models can catch up to frontier models within months largely because data is the primary driver of progress, and data can be distilled from public APIs. If architectural tricks and hyperparameter optimizations were the main factor, catching up would be far harder than we observe.

Can Scaling Bigger Models Match Human-Level Learning?

A common response to sample-efficiency comparisons is that we simply haven't scaled enough. The human brain has roughly 100 trillion synapses; frontier models sit around 5 trillion parameters. Maybe scaling another order of magnitude or two would close the gap?

Here's why that argument doesn't hold up. If you look at how scaling-law equations actually work — specifically the Chinchilla scaling-law formulation — the parameter count and data terms contribute independently to the loss. If you ran a thought experiment where you increased model parameters to infinity, you'd only reduce the required data by roughly a factor of ten. Humans appear to be thousands to millions of times more sample-efficient than current models. No amount of parameter scaling within the current paradigm can close a gap of that magnitude. This strongly suggests that humans and current LLMs are on fundamentally different scaling curves.

What About Evolution as AI's "Pre-training"?

A popular counter-argument — one Andrej Karpathy has made — is that billions of years of evolution effectively pre-trained the human brain, so the comparison is unfair. But this doesn't quite hold either. The human genome is only about 3 gigabytes, with 1-2% protein-coding. There simply isn't enough storage there to encode the equivalent of a pre-trained neural network's weights. What evolution more plausibly did was find the right hyperparameters and loss functions — the architecture and learning rules — while the actual connectome still builds up from scratch within a lifetime. And even granting the evolutionary pre-training argument fully, it still doesn't explain why every new marginal capability requires so much additional data. Once you're educated, you don't need a hundred professors to learn a new programming language. Post-pretrained AI models still require enormous data for each new skill.

Why Do AI Models Need So Much More Data Than Humans?

The honest answer is: we don't fully know, but the evidence points clearly to a fundamental architectural difference in how humans and current LLMs learn. It isn't about sensory data volume — blind and deaf people still develop general intelligence with far less sensory input than sighted, hearing peers, suggesting that raw token count isn't what makes human learning efficient. Something about the way humans build internal world models, generalize from sparse examples, and transfer knowledge across domains is qualitatively different from what current neural networks do, regardless of scale.

Robotics offers a particularly stark illustration. A human can learn to teleoperate a robot arm in hours. If AI could match that, robotics would already be a multi-trillion-dollar industry. Even with millions of hours of collected demonstrations, AI systems still struggle with complex, open-ended physical tasks. The gap is not closing fast.

Will AI Really Automate White-Collar Jobs?

Despite the sample-efficiency problem, the labs' bet on white-collar automation is not obviously wrong — it just depends on the nature of the work. The key insight is that common tasks can be brought into the training distribution. If most of what a software engineer, analyst, or accountant does on a daily basis is predictable and well-represented in training data, you don't need human-level sample efficiency to automate it. You can be ludicrously inefficient in training and still win economically, because AI can amortize its learned skills across billions of simultaneous sessions, something no human can do.

The harder question is how much truly out-of-distribution thinking these jobs require. Some roles — bank tellers, travel agents — were automatable long before modern AI precisely because they were so routine. Others, arguably including software engineering at higher levels of complexity, require daily navigation of novel problems. The prediction here is that there may actually be more demand for human software engineers by 2028 than today, driven by AI as a complementary input rather than a replacement.

What Would a Real AI Intelligence Explosion Actually Look Like?

The labs' plan for tackling the sample-efficiency problem itself is to first automate AI research, and then have automated AI researchers solve it. This is the crux of the intelligence explosion hypothesis. But the current discourse around intelligence explosions tends to be unhelpfully binary — either dismissing the possibility entirely, or assuming some godlike superintelligence emerges suddenly.

The more interesting and careful question is: what does faster-than-usual AI progress look like when it's built on top of LLMs and their particular kind of intelligence? An AI that can meaningfully accelerate research doesn't have to be generally superhuman — it just has to be good enough at the specific tasks involved in doing AI research. Whether today's sample-inefficient models can bootstrap their way to solving their own fundamental limitations is one of the most consequential open questions in technology. The answer will determine not just the pace of automation, but the entire trajectory of what comes next.