The data black hole at the center of AI

So one definition of intelligence is sample efficiency. That is to say, how much data do you need in a given domain to operate fluently and competently? And it's actually not clear that we've made that much progress in training sample efficiency over the last few years.

Reframes intelligence as sample efficiency, raising the provocative question of whether AI progress is actually about better learning or just more data.

It seems more like we've just dramatically widened and improved the data distribution. The main way that AIs have been getting better is from adding more and better data, and scaling the compute required to develop that data in the first place.

Challenges the popular narrative that algorithmic breakthroughs are driving AI progress, attributing gains primarily to data quantity and quality.

Obviously, RL is the main way that this has happened. You can think of RL as basically a kind of synthetic data generation, where you dump a ton of compute against a verifier — or a rubric, if you have an LLM as a judge — in order to find out what the good data is in the first place. And then you train your model to predict these correct rollouts, much in the same way that you might train that model to predict the next word in internet text.

Reconceptualizes reinforcement learning as a compute-intensive data mining process rather than a fundamentally new learning paradigm.

For this process to work, the model must have at least some prior probability of anticipating the correct solution in the first place, which is why you need mind-stretching amounts of human expert trajectories in every single field and skill that you want the model to eventually be competent in.

Exposes a fundamental dependency: RL-based AI improvement is bottlenecked by the prior breadth of human expert demonstrations.

It's hard to overstate how task-specific and bespoke this human expert data is. If you want some intuition, I recommend checking out the job descriptions on Mercor or Surge's websites. There are listings for Word specialists who will convert legacy documents into polished Word files, and legal experts who will write realistic M&A diligence reports or securities filings, and management consultants who will write up template market research.

Makes the abstract concept of domain-specific training data concrete, revealing a sprawling hidden labor industry supporting AI capabilities.

Now imagine if it took a couple decades' worth of courses with hundreds of concurrent professors and millions of practice tasks for you to learn how to polish a Word file. Even the task-count difference here understates the gap, because the models have to grind through their far more numerous tasks, each far harder. Whereas a human student might practice a textbook problem once or twice, with GRPO, these models are generating hundreds to thousands of rollouts per task, and they need to do this to solve the credit assignment problem.

The GRPO analogy vividly illustrates the staggering inefficiency of current AI training compared to human learning, making the sample-efficiency gap viscerally real.

The correct way to think about these models is not like a human who has learned all these different skills that you see the models displaying. It's more like a Frankenstein's monster that has been built out of a billion grafts of carefully constructed examples, all sewn together.

The Frankenstein metaphor reframes AI capability as stitched-together pattern matching rather than unified understanding, with deep implications for reliability.

I think the reason it is relatively easy for open source and previous laggards to catch up to within months of the frontier is that data is the real driver of progress. And data can be easily distilled from public APIs, whereas hyperparameters, training tricks, and architectural optimizations cannot.

Offers a data-centric explanation for why frontier AI leads are so thin, with major implications for competitive dynamics and IP strategy.

It is easy to forget how much data these models are trained on, and how much more it is than what we humans see in our lifetimes. We see these AIs as a galaxy glittering with capabilities. But at their center, invisible to the naked eye, holding all the constellations together, is an unimaginably massive black hole of data.

A striking metaphor that reframes visible AI capability as a surface phenomenon sustained by an invisible, vast data substrate.

If a person sees and hears on average, let's say generously, 2,000 words an hour, then between the time they're born and the time they're an adult, they'll see about 200 million tokens. Now, by contrast, these frontier models are trained on somewhere between tens to hundreds of trillions of tokens. That is close to a millionfold difference.

A concrete, quantified comparison that makes the human-AI sample-efficiency gap impossible to dismiss as merely theoretical.

If you wanted to, you could learn to teleoperate any random humanoid or robot arm within hours. And if we could get AIs to learn just as fast, robotics would be a deca-trillion-dollar industry, and you'd have an endless army of Unitree G1s doing all kinds of useful work in the world. But the reason we can't do this is that our AIs learn much less efficiently than we do, and even with the millions of hours of demonstrations that we've collected, this is not enough to allow them to perform complex, open-ended tasks.

The robotics example grounds the sample-efficiency problem in trillion-dollar economic stakes, showing what's directly blocked by this unsolved challenge.

One thing people will say, and I think Karpathy said this when he came on my podcast, is that for humans, many billions of years of evolution had to go into basically pretraining us. And so we're being unfair when we're comparing how little data we see within our lifetimes to what these cold-started LLMs, which are just starting off with a totally random initialization, have to learn from. I think this is not the right way to think about it. Our genome is only three gigabytes, and only one to two percent of it is protein coding. There is simply not enough space to store the parameters of this network that evolution supposedly pretrained.

The genome-as-storage argument is a sharp rebuttal to the 'evolution pretrained us' defense, forcing a rethink of what biological learning actually inherits.

I think the closer analogy is that evolution found the right hyperparameters and the right loss functions, and that within our lifetime, we are still building up the connectome in our brain from scratch. That is to say, the thing analogous to the weights and parameters of the neural net itself. And even if you granted this comparison and said, 'Yes, the hundreds of trillions of tokens these models see to get pretrained is similar to just catching up to evolution,' that still doesn't explain why any new marginal capability that you want to give these models takes so much data.

The hyperparameters-vs-weights analogy reframes the evolution debate productively, but the killer point is the marginal capability problem that persists even after pretraining.

My response to this objection is simply that blind or deaf people, who are cut off from parts of this sensory stream, still have general intelligence. That suggests to me that all these billions of sensory tokens are not really the thing that is making humans smart. In fact, deaf people who communicate through sign language and reading, and not through hearing, are probably ingesting far less than the 200 million language tokens that we ballparked earlier, which suggests that even the millionfold difference that we calculated earlier might be an understatement.

The deaf/blind argument is a clever natural experiment that isolates the source of human intelligence and suggests the sample-efficiency gap may be even larger than calculated.

If you look at the way the scaling-law equations work, they tell you that the parameter and data terms are added to the loss independently. Suppose you have a model, and you've trained it compute-optimally, and you say, 'I want to be sample efficient. I want to use as little data as possible, and I'll throw in as many parameters as necessary to make that happen.' Take the constants from the Chinchilla scaling-law paper. Even if you increased the number of parameters by infinity, that would only decrease by a factor of ten the amount of data that you need in order to keep the same loss. Humans are somewhere between thousands to millions of times more sample efficient than these models. So scaling the size of current models simply can't make up for that discrepancy, and this really does suggest that humans are on a different scaling curve altogether.

A mathematically grounded argument that scaling model size cannot bridge the human-AI sample-efficiency gap, challenging a core assumption of the scaling hypothesis.

Okay, all these nerdy comparisons aside, you might ask: why do we even care about sample efficiency? Is this actually necessary for the labs to achieve the two overarching objectives they have, which are, one, to automate white-collar work, and two, to automate AI research itself? The bet that the labs are making with white-collar work is that the common tasks that a software engineer or analyst or accountant needs to do are common, and as a result, you can bring them into the training distribution quite easily.

Sharpens the practical stakes by connecting the abstract sample-efficiency debate to the concrete economic bets AI labs are making on automation.

It might be more inefficient to train AIs to do these kinds of tasks than it is to train humans, but so what? Human lifespan simply does not allow for the quantity and the breadth of training that these models experience. If you, as a human, had some weird learning disability where you needed to read through every public repository on GitHub before you could be a competent software engineer, then it would simply not make sense to train you up. You'd be on Social Security by the early stages of your education. But AIs can learn these skills by firehosing gigawatts of training at a time, and what they learn can be amortized across billions of sessions at once.

Flips the framing: AI's sample inefficiency doesn't matter for economic competition because parallelism and amortization make inefficient training commercially viable.

Some jobs are so mechanical and predictable that we were able to automate them long before the modern era of AI, for example, bank tellers or travel agents. But there are other jobs that require dealing on a daily basis with problems that are quite distant from the data distribution. I think software engineering is probably one such job. This is the job that AIs are supposed to take first, but I would be willing to bet that there's overall more demand for human software engineers in 2028 than there is right now, largely due to the complementary input of AI.

A specific, falsifiable prediction that AI will increase rather than decrease demand for software engineers — a contrarian bet worth tracking.

The labs' plan for this latter category of jobs is first to automate AI research and then have the automated AI researchers solve the sample-efficiency problem. So then the question is: can AIs, which do not have human-level sample efficiency, nonetheless solve the remaining research problems that stand in the way of human-like intelligence and learning? I think that the way people currently think about an intelligence explosion is very clumsy, because either people dismiss the possibility of AIs speeding up AI progress altogether, or they assume that some kind of God pops out the other end. They don't reason carefully about what it looks like to have a period where AI progress is much faster than usual, but to have that happen on top of LLMs and the particular kind of intelligence that LLMs are.

Identifies a critical gap in intelligence-explosion reasoning: most discourse skips the messy intermediate phase where AI accelerates research but remains LLM-constrained.