The Takeoff Story Has a Circular Problem

I build my own harnesses. Over the last several months I’ve run the open-weights models, DeepSeek Pro v4 and GLM-5.2, alongside a stack of SaaS tools, and I spent real time on the frontier ones too, Opus 4.8 on its million-token context and Fable 5 while they were available. So I’m not coming at this from the cheap seats. I use these things daily, I push them hard, and I’ll say plainly that the improvements are genuine. The newest models one-shot more than anything I’ve used before. The code that comes out is more workable than what previous generations produced, and it isn’t close.

And then you go to scale the app, and the seams show.

I’ve got open-source projects I’ve been working on for six months and longer, and on the bigger ones the gaps get clear fast. Inspect the code the model is proud of and the issues are still in there, fewer than before, but there. Push the app toward anything that has to actually scale and it starts leaking at the seams. It still takes a human who knows how to scale an application to hold the thing together. The best frontier models available today have not changed that. They’ve made the first 80% faster and left the last 20%, the part that decides whether your app survives contact with real load, sitting just about where it was.

I’d been chewing on why that gap is so stubborn, and then Dwarkesh Patel put out a twelve-minute video that names it better than I had. It’s called “The data black hole at the center of AI,” and it’s worth your time. His frame is the one I’d been circling without quite landing on: sample efficiency.

The Millionfold Gap

Sample efficiency is just how much data you need to get good at something. By that measure, Patel argues, we haven’t actually made the models much smarter over the last few years. We’ve made the data better and wider, and we’ve thrown enormous compute at generating more of it. The models look like they’re learning the way we do. They aren’t.

The comparison he uses is hard to shake. A person, from birth to adulthood, takes in something like 200 million tokens, going by a generous 2,000 words an hour. A frontier model trains on tens to hundreds of trillions. That’s close to a millionfold difference, and it’s the gap between a model that has seen nearly everything and a kid who has seen almost nothing by comparison and still ends up smarter per unit of data.

What makes the argument hold up is that he works through every objection before you can raise it. Maybe evolution did the real pretraining, and it’s unfair to count only a human lifetime? The genome is three gigabytes, and one to two percent of it codes for proteins, so there is nowhere near enough room in there to store a pretrained network. Maybe we’re forgetting all the multimodal data a person takes in? Blind and deaf people still have general intelligence, and someone who communicates mostly through sign and text is ingesting far less than even that 200 million, which means the gap might be wider than a millionfold, not narrower. Maybe we just haven’t scaled the models far enough? Take the constants from the Chinchilla scaling paper and pushing parameters toward infinity buys you a factor of ten on the data you need. Ten. The gap runs somewhere between a thousand and a million. Scaling the current kind of model doesn’t close it.

His conclusion is the uncomfortable one. Humans are on a different scaling curve altogether. Whatever the models are doing, it isn’t a slower version of what we do, it’s a different thing that happens to produce similar-looking output. He has a line for what they actually are, and it stuck with me. Not a mind that learned these skills, but “a Frankenstein’s monster that has been built out of a billion grafts of carefully constructed examples, all sewn together.”

The Engine Is Data, and the Data Costs Money

If you read what I wrote about token prices and the AI bill, this next part should land. If data is what’s actually driving progress, then progress carries a procurement cost, and that cost isn’t falling.

Patel walks through how the newest gains get made. Reinforcement learning, the technique behind most of the recent jump in reasoning, is at bottom a way of generating synthetic data. You point a lot of compute at a verifier or a rubric, let the model take thousands of attempts at a task, find the attempts that worked, and train the model to reproduce them. It works, but only if the model already has some chance of stumbling onto the right answer in the first place, which is why, in his words, you need “mind-stretching amounts of human expert trajectories” in every single field you want the model to be good at.

And that data is bespoke. He points at the job listings on labeling vendors like Mercor and Surge, where you’ll find postings for Word specialists to convert legacy documents, legal experts to write realistic M&A diligence reports, consultants to produce template market research. Every skill you want in the model corresponds to hundreds of human experts writing example completions, building rubrics, and explaining their reasoning. The industry that produces all of this is already pulling in billions a year and heading for tens of billions. He even points to Epoch’s finding that open models lag the frontier by only about four months, and reads it the way I would: catching up is easy because data is the real driver, and data distills out of public APIs, while genuine architectural breakthroughs would not.

That’s the data ceiling, and it sits right on top of the two I’ve written about before. Compute is constrained, because the datacenters aren’t built and nobody can tell you when they will be. Capital is constrained, because the labs are burning cash and have started charging real money for tokens to slow the bleeding. And now data, the thing that actually moves the models forward, turns out to be the most artisanal and expensive input of the three. You don’t scale your way out of artisanal. You pay more for it every year.

The Circular Problem

Which brings me to the story everyone keeps selling, the one where the models start improving themselves and the whole thing takes off into superintelligence. You’ve heard it. Once AI can do AI research, it builds a better AI, which does better research, and the curve goes vertical. ASI is supposedly just a couple of self-improvement loops away.

Look at how the labs actually plan to get there, though, and the logic folds in on itself.

The stated path, more or less out in the open, is this. Use AI to automate AI research first. Then let those automated researchers solve the hard problems, including the deepest one, which is sample efficiency, the very thing that would let a model learn like a human instead of needing a million times the data. Read that twice. The plan to fix the sample-efficiency bottleneck is to hand the problem to the systems that are defined by having the bottleneck. The researchers you’re counting on are the Frankenstein’s monsters. They’re brilliant at what sits inside their distribution and lost outside it, and “solve the problem no quantity of training data has ever solved” is about as far outside the distribution as a problem gets.

Patel is sharper about the takeoff framing than most people who write about it. The usual conversation, he says, either dismisses the idea that AI could speed up AI progress at all, or it “assumes that some kind of God pops out the other end.” Neither one does the actual work of reasoning about what a burst of faster progress looks like when it’s still happening “on top of LLMs and the particular kind of intelligence that LLMs are.” That’s the whole thing right there. Faster LLM progress is still LLM progress. It inherits the LLM’s ceiling. You can run the loop quicker and still be running it inside the same box.

So that’s why the takeoff story has a circular problem. It assumes the exit from the box gets built by the thing that lives inside the box.

What the Seams Are Telling Me

Take all of that back to my harness leaking at the seams, and it stops being a complaint about model quality and starts being the same story told from the inside.

Patel draws a distinction that matches exactly what I see. The labs’ bet on white-collar work is that the common tasks, the things a software engineer or analyst or accountant does over and over, are common enough to pull into the training distribution. And he’s right that this is working. The revenue curves say so, and my own daily experience says so, which is why the newest models one-shot so much more than they used to. Boilerplate, scaffolding, the standard CRUD, the well-trodden path, all of it is in the distribution now, and the models are genuinely good at it.

Scaling a real application is not on the well-trodden path. It’s the part of the job that lives out at the edges of the distribution, where the problem in front of you is subtly unlike anything in the training data and you have to reason about your specific system under your specific load. That’s the 20% that leaks. That’s where you still need someone who has actually scaled an app before. Patel names software engineering as exactly this kind of out-of-distribution job, and then makes a bet I would put money next to: that there’s more demand for human software engineers in 2028 than there is today, “largely due to the complementary input of AI.”

I made a version of that case a couple of weeks ago, when the CEOs who told boards to replace their engineers started walking it back. This is the mechanism sitting underneath it. The reason AI complements senior engineers instead of replacing them isn’t sentiment, it’s that the scaling work lives where the data isn’t, and sample efficiency is precisely the thing these models don’t have. It’s also the rigorous version of something I argued back in The Tokens-for-Engineers Trade Has Three Holes, that every agent session starts over from scratch. I said it then as something I’d watched happen. Patel says it with the token counts.

I’m Not Calling the Top of the Mountain

I want to be careful here, because it would be easy to read all this as “AI is fake and the plateau has arrived.” That’s not the argument, and it isn’t what I believe.

We’re not at the top of the mountain. We might not be close. But the climb we’re on right now, the one made of more data and more compute poured into ever more bespoke training sets, has a visible ceiling, and the current generation of LLMs is the thing that runs into it. If humans really are on a different scaling curve, the way Patel argues, then closing the gap probably doesn’t come from a bigger LLM. It comes from a different kind of system. That’s the whole reason serious researchers are putting time into world models and architectures like JEPA, the bet that the path past this ceiling needs a different shape, not just a larger version of the shape we already have.

That bet might pay off. It might even pay off soon. But notice that it’s a bet on a paradigm that hasn’t shipped yet, and planning your company around it landing on schedule is the same mistake as planning around tokens trending to zero, or around AI replacing your engineers by next quarter. You’d be committing the business to a future nobody can underwrite.

Plan for the Plateau, Not the Takeoff

So here’s where I come down, as someone who actually builds with these tools instead of posting predictions about them.

The takeoff story is circular, and you shouldn’t run your strategy on it. The plan to escape the LLM’s limits depends on the LLM escaping its own limits, and that’s not a plan, it’s a hope with a roadmap drawn over it. The three ceilings, compute and capital and data, are all real, and the data one runs deepest, because it isn’t a temporary shortage you can build your way out of. It’s a property of how these models learn.

Use the models anyway. They’re genuinely good, getting better, and I’m not giving mine back. But build as if the current generation plateaus, because in this paradigm it will. Keep the people who can do the work that lives off the well-trodden path, the ones who hold an app together when it starts leaking at the seams, because that work is exactly what the data can’t teach. And treat any pitch that leans on imminent superintelligence the way you’d treat any other forecast nobody can defend, which is politely, and from a safe distance.

The galaxy is real. The capabilities glitter. But there’s a black hole of data holding the whole thing together, and you can’t feed it your way to a mind, at least not with this kind of model. The companies that plan around that are going to look a lot smarter in 2028 than the ones still waiting on the thing inside the box to build its own way out.