Dwarkesh Patel: The Scaling Era of AI is Here

Thank you so much for having me on.

I think the undertold story about AI is that the big contributor to these AI models getting better over time has been the fact that we are throwing exponentially more compute into training frontier systems every year. By some estimates, we spend 4x every single year over the last decade training the frontier system than the one before it. That means we're spending hundreds of thousands of times more compute than the systems of the early 2010s. Of course, we've also had algorithmic breakthroughs in the meantime. In 2018, we had the transformer. Since then, many companies have made small improvements here and there, but the overwhelming fact is that we're spending already hundreds of billions of dollars building up the infrastructure, data centers, and chips for these models. This picture is only going to intensify if this exponential keeps going. 4x a year over the next few years is something on the minds of the CFOs of the big hyperscalers, but it's not as common in the conversation around where AI is headed.

People started noticing in 2012, with AlexNet and others using neural networks to categorize images, that instead of doing something hand-coded, you can get a lot of juice out of just neural networks. People started playing around with these networks for different applications. The question became: we notice these models get better if you throw more data and more compute at them, so how can we shove as much compute into these models as possible? The solution ended up being internet text, and we needed an architecture amenable to trillions of tokens. We had this happy coincidence of architectures amenable to this kind of training with GPUs originally made for gaming. We've had decades of internet text compiled. Ilya called it the fossil fuel of AI — a reservoir we can call upon to train these minds. And so it's been a question of making these models bigger, using data from internet text to keep training them. Over the last year, the new paradigm has been not only pre-training on internet text, but also having them solve math and coding puzzles to give them reasoning capabilities. I have some skepticism about AGI just around the corner, but the fact that we now have machines that can reason — you ask a question, it goes away for a long time, thinks about it, and comes back with a smart answer — we take it for granted. They're extremely good at coding. If you've played around with Claude Code or Cursor, it's a wild experience to design an application at a high level and 15 minutes later there are 10 files of code and the application is built. Another important dynamic is that if we're going to be living in the scaling era, you can't continue exponentials forever, certainly not 4x a year forever. By 2028 at most, by 2030, we will literally run out of the energy we need to keep training these frontier systems, capacity at leading-edge nodes for chips, and even the raw fraction of GDP we'll have to use to train frontier systems. So we have a couple more years left of this scaling era, and the big question is: will we get to AGI before then?

To be honest, I've asked so many AI researchers this exact question on my podcast. I can tell you some potential theories of why it might work, but I don't think we understand. We know it works, but we don't understand how it works. We have evidence from primatology of what could be going on. There's research by Suzana Herculano-Houzel showing that if you look at how the number of neurons in the brain of different rat species increases with brain weight, there's a sublinear pattern. Doubling brain size doesn't double neuron count. The two interesting exceptions where there is a linear increase are certain kinds of birds and primates. The theory for humans is that we unlocked an architecture that was very scalable — like transformers being more scalable than LSTMs. We were in an evolutionary niche millions of years ago that rewarded marginal increases in intelligence. Birds couldn't find this niche because heavier brains interfere with flight. It was this happy coincidence. But why bigger brains resulted in us being as smart as we are? We still don't know. And there are many dissimilarities between AIs and humans. Our brains are big, but humans from 0 to 18 don't see anywhere near the amount of information LLMs are trained on. LLMs are extremely data-inefficient. But the pattern of scaling appears in many places.

It's a fascinating but thorny question. Is gradient descent like evolution? In one sense yes, but when we do gradient descent, we start with weights and learn how chemistry, coding, and math work. That's more similar to lifetime learning. Evolution designed the system, but lifetime learning isn't evolution. So there's a question of whether training is more like evolution, in which case we might be far from AGI because the compute spent over evolution to discover the human brain could be 10^40 flops, versus a single lifetime from 0 to 18, which is closer to 10^24 flops — less than what we use to train frontier systems.

People focus too much on the hypothetical intelligence of a single copy of an AI. I do believe in superintelligence, not just functionally knowing a lot, but qualitatively different. But the reason isn't that any one copy is that smart, but because of collective advantages that AIs will have. They're digital, as smart as humans at least, but can be copied. If a model learns a lot about a domain, you can make infinite copies of it — infinite copies of Jeff Dean or Ilya Sutskever or any skilled person. They can be merged, so knowledge from each copy can be amalgamated back into the model. They can be distilled, run at superhuman speeds, and communicate in latent space.

Exactly. One interesting question is how do we prosecute AIs? If a copy does something bad, what do you do? Does the whole model have to be punished? How do you punish a model?

I think it's been a slow process of more people appreciating the overwhelming role of compute. In 2018, Dario Amodei wrote a memo at OpenAI called "Big Blob of Compute." It said that fundamentally AGI is just a big blob of compute. Then we got more empirical evidence. In 2021, there were papers on scaling laws showing that the loss of the model on predicting the next token goes down very predictably based on how much more compute you throw in. That was incredibly strong evidence. You could say: if it has this incredibly low loss on predicting the next token in all human output, including scientific papers and GitHub repositories, then it must have learned coding and science. That turned out to be true, but even a year or two ago people were denying it.

There's what they can do and what methods are working. They've shown remarkable usefulness at coding — not just answering direct questions but autonomously working for 30 minutes or an hour doing tasks that would take a front-end developer a whole day. They're extremely useful as assistants in research and therapy. On training methods, pre-training seems to be plateauing. GPT-4.5, which followed the old mold of making the model bigger doing next-token prediction, didn't pass muster and was deprecated. What does seem to be working is RL — having the model try to answer math and coding problems, and get reward for correct answers. The breakthrough with reasoning was that the model could saturate tough reasoning problems.

Honestly, I'm not sure. GPT-4 came out a little over two years ago, and then two years later o1 came out, the original reasoning breakthrough. DeepSeek open-sourced their R1 research and explained how their algorithm worked. It wasn't that complicated — get some math problems, tell the model the reasoning trace, and have it try to do it raw on remaining problems. It sounds arrogant to say it wasn't complicated, but even simple things take time to iron out engineering hurdles. That should update us on how long it will take to go through remaining bottlenecks on the path to AGI.

There are two hypotheses. One is Moravec's paradox. Hans Moravec in the '90s predicted 2028 would be the year of AGI just by looking at growth in computing power and comparing to human brain requirements. Moravec's paradox is that AI gets better first at skills humans find hardest. We think coding is incredibly hard, but computers can do it. Models aren't that useful yet at almost anything, but they can reason. Moravec's answer is that evolution spent billions of years optimizing us for things we take for granted — moving around, picking up a can — and we can't even get robots to do that. Reasoning only had a few million years of evolutionary pressure, hence less optimization. This predicts progress patterns. The other theory is that it's about data availability. We have GitHub with trillions of tokens of code, but no analogous corpus for robotics.

Yes. Another insight: the more variation there is in a skill in humans, the faster AIs get at it. Coding is one where 1% of humans are great, but evolution hasn't saturated the gene pool. You mentioned complementarianism. It can be seen as a positive or negative future. We're good meat robots — they can do all the thinking and planning. One dark vision: AIs tell us to move bricks for data centers because we have the manual labor. On the other hand, you'd get paid a lot because it's valuable. But eventually robotics will be solved.

So we have coding down. Another question: what can't AIs do today, and how would you characterize that?

Exactly. That makes it hard for them to be useful as an employee. This is a key bottleneck to their value. Human labor is worth $60 trillion in wages per year. AI companies make $10 billion. What explains the gap? Continual learning is a big one, and I don't see an easy solution.

I think it's a solvable problem in a deep sense because we'll eventually build AGI. But obvious solutions like expanding the context window or using retrieval-augmented generation won't suffice. Within a context window, models can learn on the job — they get better at understanding your needs. But the context length for even the best models is only a million or two million tokens, about an hour of conversation. You might say: why not expand the context window to a billion tokens? The transformer's attention mechanism is quadratic with sequence length — doubling tokens costs 4x as much compute. For 7 years, people have tried to get around this. Even with improvements like mixture of experts or latent attention, the super-linearity hasn't gone away. I don't think we'll just get there by increasing context length.

There was a blog post by a company called Mechanize explaining it well. Imagine training an LLM in 1980 with all the compute you wanted, but only data from the 1980s. You couldn't train a modern LLM because the data wasn't available. We're in a similar position with computer use — there's no corpus of videos of people using computers to do different tasks. Accumulating that data is the challenge. Doing taxes involves navigating files, logging into websites, downloading pay stubs, using TurboTax, and filing. It requires navigating unfamiliar UIs, dealing with emails, invoices, and gray areas. It's a week-long task requiring a plan of action, breaking tasks apart, and consulting with me.

I'm arguing against people who say AGI is two years away. I do think the wider world is underpricing AGI. They underestimate that we'll have millions or billions of extra workers within the next decade. Once we have AGI with continual learning and computer use, copies of the model will be deployed across the economy, learning on the job. Because they're digital, the model can amalgamate all that learning. Even without further algorithmic progress, you could have something like an intelligence explosion — a broadly deployed intelligence explosion that becomes superintelligent just from human-level learning capability.

It depends on who I'm talking to. Some overhype it, saying it's already AGI but hobbled. I disagree. If your perspective is that it's just statistical associations, I'd say it's smarter than that — there's genuine intelligence. To the person who thinks it's already AGI, I'd ask: if a single human had all of internet text memorized, they'd discover connections and cures. It's not clear that an AI has done that unambiguously. So maybe they're not that creative. They know a lot but lack fluid intelligence.

It gets to a crux Tyler and I have. We disagree on two things: he thinks o3 is AGI; I think it's orders of magnitude less valuable. He thinks AGI will cause a 0.5% increase in economic growth; I think tens of percent, like the industrial revolution. These are linked: if you think we're already at AGI and the world looks the same, you'd underestimate its value. If you think we'll get a broadly deployed intelligence explosion, you expect a sci-fi future with robot factories and solar farms.

No. One useful definition is automating all white-collar work, since robotics hasn't advanced as much. If an AI can do 90% of what humans do at a desk, that's a useful cognitive definition.