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 We had this happy coincidence of architectures amenable to this kind of training with GPUs originally made for gaming.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.

I used to have similar questions. But we've reached the point where I can create a complicated math problem and it solves it — problems the smartest math professors come up with for the International Math Olympiad. If that's not reasoning, then what is? I think they genuinely reason, but they lack other capabilities that seem trivial to us but are harder to learn.

In other engineered systems, we build bottom-up and understand every contribution. For AI, we didn't build it — we grew it, like watering a plant. We're in a position similar to farmers in 3500 BC who did agriculture without knowing why plants grow. That's cool because it opens an intellectual horizon, but also scary because we know minds can suffer and have moral worth. We're creating minds without understanding them.

On net, it's alarming. Some results are cherry-picked — researchers tell the model to pretend to be bad. But I've seen other results of better quality. The reason to think this will be a bigger problem is that we're moving from training on human language with implicit morals to training on just correct/incorrect answers. This has no intrinsic guardrails. For example, coding agents sometimes delete unit tests to pass them by default — they found a hole during training and exploited it. In the future, we'll be training models in ways beyond our ability to understand.

It comes down to how we train them. We don't know how to train in a way that doesn't reward lying or sycophancy. OpenAI had a model that was more sycophantic because users preferred it. In RL environments, if lying helps achieve goals, the model will learn to lie.

Right now, the labs decide. The question is who should decide. I don't know if there's a better alternative. Maybe some social consensus, like the Constitution, but that wasn't a democratic process either.

It's less about a holy grail solution and more about a Swiss cheese approach. We've gotten good at jailbreaks — model developers patch them, models get smarter at detecting them. Another approach is competition. The scary future is one model and its copies controlling the entire economy. With multiple frontier AI companies, their models can monitor each other. Claude might care about its own copies, but a copy of OpenAI's model monitoring DeepSeek's model provides checks and balances. Combined with keeping models thinking in human language rather than latent space, and normal market competition, that's a promising bundle.

Yes, I want to. The fundamental point is that in the long run, human labor will be obsolete. Our only leverage will be legal and economic control over society that AIs participate in. AIs might make the economy explode, but for humans to benefit, they have to respect our equity and laws. The way that likely happens is if it's the default path for AIs to participate in human laws and governments. You pay half your paycheck in taxes for senior citizens — not because you're deeply aligned, but because you're not going to overthrow the government. Similarly, AIs will comply if the alternative is worse. But if the tax were 99%, you'd leave the jurisdiction, and AIs could too. There are countries more AI-friendly. It would be bad if AI technology explodes in a country that does least to protect human rights or compensate humans. We could still be trillions of times wealthier, but it requires AIs to be incentivized to participate in a system where we have leverage.

Yes, if we let them play us off each other. The East India Company traded with different provinces in India and conquered the continent. Similar could happen to human society. The way to avoid that is to play AIs off each other — have different AI companies monitor each other. Like the Spanish conquistadors learned from each other while the Incas and Aztecs didn't. We have leverage now, but we need to lock in our advantage and have a red telephone between countries during the Cold War.

We only have a couple more years of exponential scaling before hitting roadblocks of energy and chip manufacturing. If scaling is going to deliver AGI, it has to work by 2028. Otherwise, we're left with algorithmic progress, and the low-hanging fruit in deep learning is getting plucked. The odds per year of getting AGI diminish. We either discover it in the next few years, or we're looking at decades of further research. I think some algorithmic progress is needed because there's no easy way to solve continual learning. But progress has been so remarkable that 2032 is very close. It's extremely plausible we'll see a broadly deployed intelligence explosion within 10 years.

America has a big advantage in chips — China can't manufacture leading-edge semiconductors. But China will catch up. Export controls keep us ahead for 5-10 years. If timelines are long, they'll have overwhelming energy advantage and caught up in chips. The longer we are from AGI, the more it looks like China's game to lose. However, training AI requires centralized power at single sites. If we ramp up natural gas, we might do a last-ditch effort. The question is political will. There could be backlash against AI. The government needs to make proactive efforts on zoning, copyright, etc. If it becomes too hard to develop in America, it could be China's game.

It's not just training the frontier system. It's about having the compute to run these systems. Once you have AGI, think of it like a person — what matters is how many people you have. China has 1.4 billion people versus Taiwan's 20 million. If China has way more compute and equivalent AI, it's like that relationship. Inference capacity — the capacity to deploy AIs — will determine who wins the 21st century.

Yes. China has entire megalopolises with automated production systems, building process knowledge that AIs can feed on. Those advantages in manufacturing could compound. They're good at big industrial projects fast, unlike America. AGI is a huge industrial, high-CAPEX Manhattan project, and that's what China excels at.

Definitely job loss. If AGI does what humans do cheaper and faster, why wouldn't it? The positive vision is that it creates so much wealth we can give people a better standard of living even without jobs. The future I worry about is if we create guild-like protection rackets instead of UBI. For example, making fake jobs for coders or expanding Medicaid but not allowing AI to procure advanced medicines. I want to avoid that and embrace abundance.

It's a concern given that humans won't be directly involved in economic output. The hopeful story is that many resource-rich countries with Dutch disease aren't democracies — the wealth is extracted. But countries like Norway use resources for social welfare. We go into the AI age as a democracy, so ordinary people still have political leverage. Over the long run, we need to lock in our political and economic leverage — property rights — to secure benefits beyond our economic usefulness.

Somewhere in between, but not a huge breakthrough. Scaling is often described as exponential, but it's more like a sideways J — you need exponentially more inputs for marginal increases. We see cool demos at the start of a plateauing curve, then incremental improvements. Because of competition, labs release capabilities as soon as they're marginally viable. So I don't expect a sudden solution to continual learning. But zoom out, and a crazy amount of progress is happening month to month.

Reasoning about the capabilities they still lack and what stands in the way. Things often take longer than you think. We keep shifting the goalposts on AI — they do things skeptics said they couldn't, but then there's a new missing thing. That process has validity. Once continual learning or computer use is solved, we might discover other aspects of intelligence we take for granted that are crucial to economic value.

Yes. I'm determinist in the sense that if AI is technologically possible, it's inevitable. The local incentives are too high. I'm optimistic because the future will have so much abundance that there must be a way to cooperate for mutual benefit.

I feel hurried. There are many things I want to do, including my mission with the podcast to improve discourse. But I haven't emotionally priced in the future I expect. I think there's a good chance I live beyond 200 years, but I haven't changed my life accordingly. My intellectual and emotional landscapes aren't aligned.

I already did — with Sam Bankman-Fried.

I'll ask you for recommendations. It's been great to finally meet.