分布式训练、去中心化 AI：Prime Intellect 让 AI 便宜到无需计量的总计划

Vincent Weisser and Johannes Hagemann, founders of Prime Intellect, welcome to The Cognitive Revolution.

Thanks for having us.

Thanks, yeah. I’m excited for this conversation.

I’m excited for this conversation. For a while, I’ve been wondering about distributed training: whether it’s going to work, what the trade-offs are going to be, and generally how to think about the concept of decentralized AI. When I recently mentioned on an episode that I was looking for the right people to talk to about this, your names came to my attention. I appreciate you being willing to do this, and I’m looking forward to getting into the nitty-gritty details of all the work you’ve done to realize this decentralized AI vision.

Before getting into that, one of my common refrains is that a positive vision for the future is currently the scarcest resource. I’m always amazed by how little of that we hear, including from people leading the frontier AI companies. We get very vague visions of what’s going to be good about the future. One thing that has struck me about the two of you, as I’ve explored your work in preparation for this, is that I think you do have a positive vision for the future.

I’d love to start by giving you an opportunity to articulate that positive vision for us a little bit.

I can kick it off by framing it in terms of what we plan to build with Prime Intellect, and then more broadly how I see the positive vision of the future with AI.

Our goal is to make intelligence and compute too cheap to meter. I think we’re entering that era right now. The key goal for us—and a key part of making it utopian—is to make it widely accessible and open to everyone.

If we zoom forward a few years or decades, the biggest risk would be having this super-powerful intelligence and compute but not making it accessible to everyone. Instead, it would be accessible only to the select few: the big states, the big tech giants, and certain individuals. I think the most positive future is one in which AI empowers every human and every creator to leverage their intelligence and be able to do more with less.

People would have agents at their disposal to create more advanced science, have better access to medicine and education, and gain access to all of these things. I think AI has the potential to usher us into an era of widely distributed abundance, if done right. But if done wrong, it also has the risk of creating a much harder-to-access intelligence age.

The most important piece is that large parts of society are able to participate in this upgrade.

We’re pretty much aligned on that. I think it needs to be rightly distributed. The most dangerous outcome is if there’s only 1 superintelligence. The safety factor is 0, obviously, and the distribution goes in the opposite direction from what we’re aiming for.

That’s how we think about the future of problem-solving through AGI and ASI.

It’s interesting that you mention AGI and ASI in that answer. So many AI worldviews or expectations really diverge on how powerful people expect AI systems to become, and on what timeline. How would you compare and contrast your own expectations with what we’ve heard recently from Sam Altman and Dario Amodei, who seem to be increasingly confident that we’re going to see superhuman AI in the next few years?

I think most of the views put forward are very plausible. Maybe they’ve been plausible for the last 15 years, since reading The Singularity Is Near and Superintelligence when they came out. I think we’re still on that path. It’s actually surprising how well those predictions have held up.

I think the most common views of the big lab leaders—from Demis Hassabis and Dario Amodei to Sam Altman and people like Leopold Aschenbrenner—are scenarios on which I would put a high probability. They base those views on the research progress we’ve seen over the last few years.

The biggest question is how general it will be in the broadest sense. A lot of people I talk to at the labs have high confidence that AI will reach superhuman coding, superhuman mathematics, and superhuman software organizations that run autonomously. But that doesn’t mean we get superhuman humanoids tomorrow.

I think a lot of these things will look like autonomous cars. It’s easy to make radical progress, but it’s hard to get to 100% accuracy and robustness. An autonomous car that isn’t 100% robust but is 95% robust is still pretty useless. I think we’ll see the same with superintelligence. The deployment of it will happen when it has extremely high robustness in its application.

It’s not enough to have a superhuman investment agent if it can screw up 1% of the time and lose you all your money. I think that’s how we’ll see progress shape up. It will move much faster in some areas, as we’ve seen with the new inference-time-compute scaling paradigm, and slightly slower than some of the biggest optimists might make it out to be in trickier, harder-to-simulate areas.

I think the framework from people like Sten Tamkivi about how computationally reusable different areas are is useful. It’s more difficult to solve all diseases than to write a bunch of code autonomously. We’ll run into the limits of what you can compute. You can’t simulate everything and you can’t compute everything, but you can approximate reality with pretty good simulations, including for things like biology.

I’m quite optimistic. I think most people are not optimistic enough, especially in the face of all the progress and all the infrastructure build-out. I think we’re on track for superintelligence in probably the next decade.

Can you put a little more color on the future that you envision? I hear you on the accessibility part. I definitely think people are underestimating what’s coming, but I’m more unsure about whether it’s going to be good by default or bad by default, and what the hinge points are.

It’s funny that you mentioned Kurzweil, too. I’ve increasingly been using the phrase “Kurzweil’s revenge.” In retrospect, it seems like there was really no other way this was going to go. The fact that so many different algorithms seem to work, combined with the fact that we have compute and data at scale, means there was really no trajectory in which we had these compute and data resources without somebody figuring out a workable algorithm.

In that fundamental sense, if you have the physical and informational inputs, somebody will figure out how to unlock the lock. We’ll get reasonably powerful, and potentially very powerful, AI systems. That seems to have been borne out very well.

Do you have a vision for daily life? In your successful scenario, am I working a job? Is there a stable equilibrium? I really struggle to envision a stable equilibrium. The decentralized-AI line of thought in general seems to try to get there, but I’d love to have a more concrete, detailed, richer vision than I do.

It really depends on the different dimensions of AI and also on the time frame and personal preferences. In many ways, people who are financially independent could retire, but they could also still work on a startup or do other things. I think that’s already playing out.

Every billionaire in society is probably working harder, not less hard, on philanthropic efforts, entrepreneurial efforts, and other efforts. I think we’ll see more people behaving as if they’re financially independent. In that sense, there are already glimpses of how abundance manifests. We are, by definition, living in abundance, but people have different priorities in life.

A recent book I read on this was Nick Bostrom’s Deep Utopia: Life and Meaning in a World Without Work. Some of his points make a lot of sense. There are so many other aspects to life that people care about, whether it’s having children, doing art, creating things, understanding the world, or reading. People will be able to spend more time on those things.

Even if one sees further jumps in human progress, people will be able to get more freedom and ultimately have an easier time reaching financial independence. That’s one example, but more broadly, people will be able to do whatever they want to do.

Most people I know could stop working, but they don’t. I don’t see them completely stopping work once we have AGI. I think what’s interesting is that even now, among people working on this technology, there are people who resign from their AGI jobs because they think it’s time to enjoy the last 2 or 3 years before superintelligence, while on the other side there are pessimists who think the end is near. I think that’s dangerous and unlikely. I think the nature of jobs will change, but there will probably be a return to very human labor. People might still go to a restaurant or a theater, even if they don’t drive an Uber because it drives itself.

There will be a different sort of labor that becomes even bigger: everything where you and I, even after AGI, would pay for a human because it’s more joyful to have a human around than a steel robot. If you want your kids to be taken care of, maybe you don’t want your Optimus robot to do it. You want a human you trust and can understand.

That’s how I think it will play out. In many ways, it won’t be absurdly different from today. There will just be far less knowledge work that isn’t exceptional. There will be much more curation, and I think knowledge workers will still work, but they’ll work in very amplified ways on different functions. They’ll have an army of people working for them for free, which are, to an extent, the AIs.

The one point I want to push on is the billionaire point. I feel like there may be a contradiction, although maybe you can resolve it. Why are billionaires working today? In part, I think it’s because they feel they can make a meaningful contribution. They want to make a positive difference.

Does your vision imply that, for people in that position, there isn’t an AI they can hand that responsibility off to that would do a better job? You might imagine Bill Gates, Dustin Moskovitz, or somebody else saying, “My goal is to eliminate these diseases. There’s a superintelligence that could probably make better grant-evaluation decisions. Why would I sit here and run this foundation?”

Are you imagining that they don’t have that option, or that they choose not to use it for some reason? That part is always a little weird to me, because if you have a superintelligence, what are even the billionaires doing?

There was a good interview between Sam Altman and Bill Gates that touched on exactly this. You could see Bill Gates grappling with it, saying that he gets so much meaning out of being good at fighting malaria, and that if superintelligence could do it for him, he would lose some meaning because he’s so good at it.

I think that’s the wrong worldview. If you really care about something, you should be happy if you can hand it over to your employee or your agent to do it better for you.

The best founders or philanthropists build teams that are better than they are in every dimension. Maybe they have broad skills, but ideally they hire people who are superintelligent compared with them on something very specific. I hope Elon Musk has someone who is much more intelligent than he is on propulsion, autonomous cars, AI, and systems across all of his efforts.

I think that’s more how it will play out. Execution is cheap; ideas are worth everything. In a world where you can just do anything, it almost inverts the current reality. I think it would lead to billions of startups, billions of movies, and much more: an explosion of content, companies, and autonomous creations that a human is either seeking or creating.

I’m kind of extrapolating from where we are today, and from how I see and try to adapt to what seems plausible in 1, 5, and 10 years. I think there are still a lot of things we will want to do and work on.

To your point, I think the ultimate thing is almost like Maslow’s hierarchy of needs. There are things like impact, status, meaning, and fulfillment that are timeless. Even if you have superintelligence, you might still get meaning out of specific things that you see or do.

I see it as humanity having a gigantic workforce of AI systems and agents that do a lot of the work, while humanity hopefully stays on top of it and manages and directs it toward the aims that humanity cares about.

In some ways, our biggest corporations or states are already superintelligences with a lot of agents in them. They’re directed through democracy, capitalism, or whatever, toward specific directions. I think the biggest risk is that nation-states and for-profit corporations have the superintelligence, instead of everyone having it to do what they want.

When you said “hopefully,” the interjection I was going to ask about is whether you entertain the concept of p(doom) at all.

I don’t think it’s an extremely useful concept. A lot of people are way too confident in weird cyber scenarios and in probabilities that are not actually probabilities. If you ask them to unpack their reasoning, they can’t really do it.

I’ve seen too many people say, “My p(doom) is 20%.” Then you ask why, and they say, “Because this other person also has it.” Usually, the argument falls apart.

I think p(doom) is a dangerous concept. I wouldn’t put strong confidence in anyone who puts out random numbers. Probability is between 0% and 100%, and the universe will go on for billions of years. Will something eventually kill it? Sure. But when, how, and what kind of event it would be are so hard to know that it’s almost not worth trying to pin down.

One of the problems that utilitarian philosophy brought to the space is having very high confidence in future scenarios and events that are impossible to know, and then basing policy and actions on them.

What’s more important—and the ironic thing is that, in this worldview, one needs to realize that on shorter time frames the probabilities are stacked in favor of good outcomes. Policy actions can sometimes have the exact opposite consequence. Someone with a low p(doom) might say, “That’s why we need to regulate it heavily in Europe.” Then AI leaves Europe. Or they say, “We need to regulate it heavily in the United States,” and suddenly things move outside the country.

There’s a cascade of very bad epistemics and reasoning that follows from people going too deep on p(doom).

I’ve certainly seen some of that. At the same time, I do worry that things could get out of control. When I’m asked for my own p(doom), I usually say 5% to 95%, or 10% to 90%.

What I mean is that nobody has said anything compelling enough to me that I shouldn’t worry about things getting totally out of control. At the same time, nobody has said anything compelling enough that it’s so likely to happen that I should check out and spend the rest of my life at the beach.

I do worry about the possibility of things getting out of control. Do you think that worry is misplaced, or do you share it to some degree?

I don’t think quantifying it is particularly useful, but I do think considering it is useful. I see it as similar to the Industrial Revolution. It created chaos, created a lot of harm, and in some ways created civil war and other consequences that weren’t good. The internet also created complexities and political consequences.

Ultimately, I think the benefits of AI will far outweigh the risks. Some of the paths to avoid the risks can become risks themselves. In a world where everyone is worried about existential risk and calls for a world government, the world government might be a bigger risk than the science-fiction scenarios people are worried about.

All of these are complex systems. It’s hard to know what the specific risks will be. I think the biggest risk is people losing their autonomy: putting life on autopilot, losing their freedom, and handing it over heavily to the state and to overregulation.

That’s my biggest worry: that even the most well-intentioned people create the worst outcome unknowingly and unintentionally, and that their efforts are misused by people who seek power.

A lot of the people who were fighting for alignment, safety, and policy have been corrupted by the big labs and big tech companies that have their own goals and intentions. I know stories about almost all of the AGI labs that are extremely worrying in terms of what has happened behind the scenes and how they seek power, including people whom the whole community reveres.

The thing I’m really worried about is full centralization of superintelligence by a few nation-states and big tech giants, completely disenfranchising large parts of humanity.

It seems like everything is in play in a way that it hasn’t been before in my life. I do hear you on the idea that a crazy imbalance of power—or 1 superintelligence in the hands of 1 organization, 1 company, or even 1 individual—would be a problematic situation.

In biology, one lesson I’ve learned over and over again is that anything sufficiently concentrated is dangerous. Drugs can become dangerous when you take something that was fine as a leaf growing on a tree, purify it down to 100%, and start inhaling it. That seems to come up over and over again.

I’m sympathetic to the idea that if we can maintain a buffered solution and a balance of power, where everybody is checking everybody else—much like things work today, where nobody has the ability to dominate the world unilaterally—that would be appealing. If we can maintain that as we bring AI online across all aspects of life, I think that sounds attractive.

Of course, I don’t really know what that could bring. Anything seems possible there. But I find a lot to like in the vision of everybody rising simultaneously, everybody’s interests keeping one another in check, and no single actor being able to make a mistake that throws everybody else into a huge problem.

Let’s talk about the company and your master plan for getting there. I love a good master plan, and you have one. One thing I wasn’t able to tell from my research is what kind of company it is. Is it just a normal corporation, or something else? Give us a little bit of the foundational backstory, and then take us through your master plan.

Our broader goal is to make compute and intelligence too cheap to meter. We want to do that in multiple stages.

In the first stage, we started by aggregating global compute and creating an efficient compute market, along with developer interfaces such as an API and command-line interface. The goal is to allow people, no matter what kind of compute they’re looking for—whether it’s 1 H100 or 1,000 H100s—to make a request and find the cheapest compute that fits their needs.

That’s how we started and how we launched 2 or 3 months into creating the company. It was the biggest thing we initially did, and it gained a lot of traction.

Building on top of that, and in conjunction with it, we’re building decentralized training and other approaches, such as distributed synthetic-data generation. These approaches can leverage the global compute fabric. If there’s extremely cheap compute in 1 place or another and it dynamically shifts, it’s extremely beneficial if you can train in a decentralized, fault-tolerant way.

You can save massive amounts of money and leverage global compute, especially when idle compute is available somewhere in the world. Ultimately, this reduces the cost of intelligence and compute because you can make more efficient use of the resource.

That’s stage 2. We’ve also made a lot of progress there.

Closely connected to that is the goal of training leading open models collaboratively. Anyone in the world could contribute their idle compute to the network and get rewarded for it, making the highest possible return on their compute because it gets fully utilized.

If the compute is idle, people can contribute it to frontier models. For example, we’ve been looking at contributing to the continuous improvement of the R1 model from DeepSeek.

That’s almost the third pillar, which connects more broadly to peer-to-peer compute and intelligence. The goal is for anyone in the world to be able to contribute compute, and for anyone in the world to be able to use that compute. It becomes an extremely efficient market for compute.

On the other side, the same can be said about intelligence itself. Any AI API is also a kind of market. There are different people hosting Llama models, DeepSeek models, diffusion models, and so on. They’re competing, and their cost is ultimately determined by compute.

You can lower the cost of those intelligence endpoints—of agents, LLMs, and other models—by having the most efficient compute market. An efficient peer-to-peer compute market enables efficient peer-to-peer intelligence sitting on top of it.

The goal is to create a system that is more like a protocol, similar to parts of the internet, that can be maintained almost like a public good. People can use it trustlessly and permissionlessly.

The closest structural analogy is something like Ethereum. The idea is to create technology, make it openly accessible, have a foundation-like structure to support it, and then over time allow anyone to own and use it at the cheapest possible cost.

To answer your question about how we’re structured, it’s a Delaware corporation, but the setup is more akin to Ethereum. We’re creating a foundation and giving grants for people to develop this in the open through an open-source approach.

The goal is to make it fully accessible to anyone. Anyone could improve the system, contribute back to it, or participate in it. The broader design principle is to create a public utility, in conjunction with systems like Ethereum, where anyone can create agents and commercialize them. An agent can make money for you.

Going back to the utopian vision, anyone in the world could contribute to the system, own a piece of it, and have it generate revenue for the end users who participate in it.

A lot of points there. Let’s start with the compute market, and work forward through the master plan.

How would you describe the compute market today? We’ve had a hyped cycle of people raising large equity rounds and pouring the proceeds directly into GPUs. In some cases, Nvidia has even taken equity positions in exchange for GPUs. It seems like that phase may be over.

Are those the kinds of companies that are now saying, “Maybe we overbought a little bit. We have more cluster capacity than we thought we could use,” and are now contributing to your marketplace? I have a million questions about the compute marketplace, but tell me what you think is interesting.

If you zoom out, GPUs were a tiny, tiny market before ChatGPT. They’re growing at an exponential rate, and I don’t see a slowdown. This week was anomalous in the sense that there were more than $100 billion in compute commitments from individual AI companies. There’s much more demand for compute, including from a long tail of startups seeking it.

To your point about how the market works structurally, more than a year ago we were in a period when supply for H100s was extremely constrained. The big players that paid extra got more of them and received priority.

That shows you the problem. The big tech giants got priority because they paid more and had gigantic orders for long-term clusters. What has shifted is that there is now more supply. Production has started to catch up, and new generations, such as Nvidia’s B200, are rolling out.

Many of the small startups that raised capital and deployed it bought long contracts. They bought 1- or 2-year contracts and paid hundreds of millions of dollars. If you wanted to train a model and needed 1,000 GPUs, you had to buy a 2-year contract to get them. It was impossible to rent them on demand.

Fast-forward to now, and on platforms like ours you can rent 1,000 H100s on demand. That wasn’t possible even 3 to 6 months ago.

It shows that a lot of the supply was initially allocated through long-term contracts. From a capital perspective, those are more like structured financial products. CoreWeave buys tens of billions of dollars in compute, gives it to OpenAI, and takes out loans against the GPUs. That model is evolving, and more of that compute is moving to on-demand supply.

That’s what powered our growth and positioning. A year or 2 ago, it was impossible to find an H100. Every AI setup and developer we talked to had a very hard time finding H100s. We were able to find them by going directly to data centers and talking to billionaires who had bought tens of thousands of H100s and wanted to sell them on demand.

Fast-forward again, and there are even more data centers and more cloud providers. The market is insanely fragmented. To an outsider, it may not be obvious that there are hundreds of clouds and thousands of data centers, and there will be more of them. There will also be more chips than Nvidia chips.

Nvidia keeps the market fragmented by design. It doesn’t want an overly powerful buyer. If OpenAI or Microsoft were Nvidia’s only buyer, Nvidia wouldn’t be in a good position. On the other hand, its biggest clients also compete with it. Google, Amazon, and the other big buyers are producing their own chips and competing directly with Nvidia.

Nvidia needs to be careful, so it gives many allocations to smaller clouds such as Lambda Labs and CoreWeave rather than to the very biggest tech clouds. Those smaller providers are more aligned with Nvidia in many ways.

That’s why supply is distributed across hundreds of data centers in America and around the world. As an end user, you want to discover, “I need 200 GPUs. Where can I find them right now?” That’s not easy or clear because demand and supply shift in real time across all these platforms.

There were also well-known cases of startups that raised a lot of money and started pretraining models. That continues to be the case. There are more AGI labs today than a year ago. There are also many applications that need more and more compute, especially new use cases such as video, heavier reasoning, coding agents, and agents in general.

I don’t think there will be a slowdown in compute. On the other hand, compute will probably become one of the biggest chunks of GDP as we ramp up toward AGI or ASI, powering all of the different intelligence use cases.

Would it be fair to say that you’re playing a role like an aggregator of people who are also selling directly? I don’t know whether to pick Kayak as the analogy: I could buy directly from the airlines, or I could go to Kayak and get a menu of all the different airlines.

That’s as opposed to eBay, where somebody says, “I have 3 H100s here. Can I plug them into your marketplace?” It sounds more like the former.

Ultimately, it’s not so different from the biggest marketplace in the world, Amazon. Anyone in the world can sell compute on our platform. We’re in touch with everyone selling compute, from hyperscalers and individual data centers to cloud providers, billionaires who have compute, startups that have too much compute, and startups that want to buy compute and put it back onto the platform when they don’t need it.

The best analogy is that a lot of platforms are isolated islands. We’re trying to create the ocean that connects all the islands. You have some compute in one cloud, some in another, and some in one big tech company and more in another, but it’s very hard to orchestrate all of it.

A lot of the people we talk to start with 2 people building their own APIs to query the availability of all the GPUs. They have different APIs, and that’s all the work we simplify for the end user. You can orchestrate this global compute fabric.

A lot of these providers are almost like private-equity products. They borrow $1 billion to buy $1 billion of compute and lend money against the compute. It’s a very high-capital-expenditure business.

What’s different is that we don’t buy hundreds of millions or billions of dollars in compute. We’re not a hotel; we’re more like Airbnb. We’re a marketplace sitting on top of other marketplaces.

Some of the marketplaces we talk to that are trying to do something similar don’t cover the full spectrum. They might not work with hyperscalers or larger clouds; they might work only with individual data centers. We’re trying to aggregate all of them over time onto 1 market.

What’s the smallest unit of compute that would make practical economic sense for somebody to contribute to the marketplace today?

We have every type of chip that’s in demand. The majority of demand right now is for H100s, B200s, A100s to some extent, and 3090s. Over time, there’s reason to believe that, for synthetic-data generation, you could even use a MacBook or something similar. That’s not our primary focus right now, but compute is everywhere. It’s not just in data centers or H100s.

Over time, the goal is to leverage all the compute out there. Some of it is much more useful for different AI workloads, but as we shift to the synthetic-data paradigm, some new types of compute become useful.

Nvidia is now shipping its own home GPUs, and those could become interesting over the next few years, especially if a lot of people get them. The answer is all compute, but compute that’s in the highest demand will be more successful on the market. It’s very market-driven.

For users on the platform, all kinds of compute resources would be useful for development purposes. For the distributed-training component, it’s much more focused on high-capacity GPUs: an H100 or A100 node, a full node, or even full clusters.

There are still use cases for 4090s, smaller clusters, and data generation. For training, you need quite a bit of memory, so it’s more focused on A100s and H100s.

Part of why I’m interested in the economics and practicality of contributing to the market is that it seems like, in the long term, we may be headed for a fight about who can compute, under what controls, under what circumstances, and with what oversight.

The trillion-dollar data center is one idea. It’s highly aggregated and highly centralized. From a geostrategic standpoint, it’s probably relatively easy for a government to control: you can fence the perimeter and know who comes in and out.

On the other end, there isn’t really anything like this for AI yet, but there is Bitcoin and Ethereum. These are systems that are truly not something people can shut down. It would be very difficult, even for the United States government, to shut down those networks if it really wanted to.

Where are we on that spectrum today with these global, decentralized compute fabrics, and where do you think we’re going? If you need to be at significant scale and basically be a data center, we’re not at the extreme of a trillion-dollar data center, but we’re closer to that. It could be controlled or governed.

If you could get down to the level of, “I have a Mac mini on my desk that’s idle, and I’ll rent it to you for $0.05 an hour—not even necessarily for the money, but because I want to contribute to a global network nobody can shut down”—we would be in a very different regime.

On the first point, I don’t think the $2 billion cluster is going to be 1 single cluster. Even the Stargate build-out will be distributed. All the big tech companies are looking into how to train across distributed clusters.

They have a more favorable environment than we do. They have cables between their data centers, with at least 40-gigabit or 100-gigabit interconnects, which we obviously don’t have in a globally distributed setting. But even there, I would argue that it’s not easy to control. You don’t have 1 single location that you need to prevent somebody from accessing.

It will be distributed across the globe. If you want to build something like a single cluster, energy requirements mean you can’t build a 10-gigawatt cluster right now. We’ll be limited to multiple 500-megawatt clusters over the next couple of years.

What we’ve already seen with our distributed-training run is that we trained across Europe and Asia. Both things can be true at the same time. Compute is already distributed across Europe, America, Asia, China, India, and other parts of Asia.

A lot of compute is shifting to places where it’s freer, where you can generate revenue and aren’t encumbered by stringent regulation. Places like Malaysia and Singapore are seeing a boom in compute.

There’s also a gigantic long tail of compute. A lot of data centers have hundreds or thousands of GPUs, but most don’t have 100,000 or 1 million. You can count those on 1 hand. They’re more like the xAI, OpenAI, and Microsoft clusters.

There’s also a lot of lower-powered compute in people’s homes. A ton of people have 4090s and 3090s across the world. So I think it’s not only extremely naive to try to regulate compute; it’s also extremely dangerous.

David Deutsch said in a recent interview that regulating compute is like regulating code or mathematics. It’s extremely dangerous, but also extremely naive, because it has the exact opposite effect. It pushes compute toward nations that are even harder to regulate and observe. It shifts compute to the boundaries and to actors that don’t obey regulations.

You can see this with Bitcoin and Ethereum. A lot of Bitcoin mining was in China, and then China banned it. A lot moved to the United States, but energy was cheaper in other regions, so it also moved to parts of Africa and elsewhere.

I think the same thing will play out with AI. Europe is perhaps the best example. We’re both from Europe, and we both left Europe. One reason is that Europe is extremely bureaucratic and co-opted by regulators who ultimately don’t have the best interests of European citizens, the future, or the economy in mind.

Europe overregulates. It didn’t even have a real AI community. It was barely alive, and now it’s essentially nonexistent. It’s the best example of a continent that fumbled its participation in the new intelligence age because of a few corrupted regulators.

I know some of those regulators. They had the best intentions. They didn’t plan to destroy the future of the entire continent, but they did. That’s a cautionary lesson.

Fortunately, I’m very excited for America. America isn’t repeating that mistake. It was on track to repeat it under the last administration, but I think we’re on a bright path toward an optimistic future. Otherwise, the United States would have said goodbye to its future and its stake in AGI.

That would have been extremely dangerous and would have pushed development even harder toward Asia, especially with progress from DeepSeek.

It’s important to think about second- and third-order consequences. A lot of policy people I know, including some of the most powerful ones, regret policies they argued for. Some of the people who indirectly funded the biggest initiatives aren’t fans of them anymore. They publicly state that, or will do so in the future.

You’d be surprised by how many people who enabled major policies now regret having done so. That will become more public as the Overton window shifts.

There’s a lot to update on. Funding superalignment is great, and there should be more of it. But funding policy quite quickly goes wrong, gets co-opted, and has unintended consequences. It can go horribly wrong.

On the other side, most policy proposals are confused about governance. People like Jack Clark have commented on our work and said that compute governance needs to be updated. Compute governance in the face of distributed training makes little sense, especially in the way it’s done right now.

Policy proposals have been very slow and lacking, but also very naive. They aren’t in the spirit of freedom, democracy, free markets, or individual autonomy. That’s very problematic.

What policies do you think people are regretting?

I think the EU AI Act will go down in history as one of the worst ideas, not only for Europe but also because it has had essentially zero impact. It didn’t improve anything for anyone. It made things worse for small AI startups.

That’s part of a general problem Europe has: it’s very proud to regulate and overregulate industries that are tiny or nonexistent. The EU AI Act is a concrete example. The California AI Safety Act is another. Even the people arguing for it said it should have been done at the national level rather than only in California.

Those are the 2 most well-known policy initiatives. I know some key policymakers at the big AGI labs who have left and regret supporting those efforts. I also know some of the biggest funders of AI policy who regret making those donations.

There will be more of that, and it will become more public. People will realize the outcomes of the policies they supported.

The same goes for funding superalignment. It’s good to fund that work, but funding policy can quickly go wrong. It can be co-opted and have unintended consequences. I think that happened to some extent with a lot of the governance and policy work.

The majority of the proposals were confused about governance. Jack Clark also commented on our work and said that people need to update. Compute governance in the face of distributed training makes zero sense, especially as it’s done now.

I think David Deutsch gave a good interview on this. If you’re worried about an unsolvable problem, the only path is to solve it. AI can help solve those problems. One of our first model releases showed that AI can make the world safer with respect to biosecurity.

The philosophy is well encapsulated by what Vinod Khosla wrote about accelerating defensive technologies and differential development. Democracy and decentralization are probably the path to the brightest future in this direction.

That doesn’t mean governments should overregulate. It means making sure that good actors are supported while bad actors are stifled. Policy and regulation often have the opposite effect. The best and most well-intentioned actors follow the rules, while bad actors don’t care.

I agree with the counterproductive nature of many regulations. I’m personally worried about that right now with policies directed toward China. I can’t shake the feeling, no matter how many arguments I hear to the contrary, that the actions we’re taking will make the situation worse.

You also mentioned superalignment. What do you think is reasonable for society to expect from the frontier labs?

I went back and forth on SB 1047. Initially, I thought it seemed reasonable. Then there was a version in which a panel or commission appeared to be overpowered. I didn’t love the idea of 5 people appointed by the governor making all these decisions. When that was the version being considered, I was opposed.

Then that provision was cut from the final version. I would summarize the final version as saying that if you’re doing frontier AI work, you have to have a safety plan, you have to publish that safety plan, and you have to be open to some amount of scrutiny.

That seemed pretty reasonable and light-touch. Even if you don’t frame it as law, policy, or regulation, what does it mean for a frontier AI developer to act responsibly in today’s world? They’re taking potentially disruptive steps, so they should proceed with caution and try hard to act responsibly.

What does that mean in your mind? What should we expect from them or demand from them?

A key aspect is differential development of defenses. You and I can’t defend against AGI if we don’t have one ourselves, especially against cyber, biological, or other threats.

Going back to the frontier labs, I think much of what they’re currently doing is reasonable. They red-team their systems and have goals of publishing something safe that can’t easily be misused.

Ultimately, you can create a technology, but the burden is also to some extent on the user. A lot of things can be misused, whether it’s the internet, a car, or anything else. Most people don’t misuse them. They use them lawfully and carefully.

We should assume the same with AGI. The vast majority of humanity will use it for good purposes. There will be a tiny minority that misuses it. That’s where the focus should be: stopping the misuse of technology.

The details depend on the flavor of AI research. I think all the labs have taken reasonable steps on safety, alignment, and related issues.

The benefit of open source is that you have more oversight and transparency. People can jailbreak the models, and others can quickly create solutions for those jailbreaks. That makes systems more robust.

If I’m going to hand over my health care, my children’s education, and key parts of social functions to a system, I want to see the code, verify the code, and run tests against it. A black box served by an AGI lab doesn’t instill much confidence in me.

I think there are reasonable demands for people to gain more insight into frontier models than just an API endpoint with limited access to capabilities. I know you red-team some of the models. You can discover a lot by using closed black boxes, but you can’t look inside them or understand what’s really happening.

I think we would be much safer if all frontier AGI models were completely open source. The building blocks of the internet are extremely open, as are blockchains. They’re safer because they’re open and operate in an adversarial environment where people try to break them.

That would be the safest path for this technology. Given that closed models are currently black boxes, I think they should face more stringent regulation than open-source models. Closed labs should open their code to people who can examine it deeply and perform tests beyond API access.

The jury is still out on whether closed models can achieve the same rigorous testing and transparency that open-source models get by nature.

I’m of mixed minds on this. My background is generally libertarian. I’m skeptical of regulation, broadly pro-capitalism, and supportive of free markets across a wide range of outcomes.

I do wonder how well the open-source label really applies to models. Models are not that much less of a black box just because you have the weights. They can be somewhat less of a black box. I’m a small angel investor in Goodfire, for example, and they’re doing interpretability work on Llama 3 70B, or 3.1 or 3.3, whatever it is exactly.

There’s some degree to which the black box is starting to get cracked open, but we’re still nowhere close to saying that a model is fully under control or won’t do bad things.

I’m always looking for a way to square these objectives—to get the best of both worlds. I think more access than an API, especially for safety and academic researchers at the big labs, should be a significant priority. The access I’ve had has been purely through APIs, with very little information. It has gotten somewhat better, but I’d like to see it improve further.

At the same time, I worry that we don’t know what the trajectory of these technologies will be. The R1 moment is a good example. We don’t know exactly what the trajectory of these technologies is going to be, so there’s something risky about open-sourcing frontier systems that you can’t take back.

If you open-source Llama 70B and somebody performs gain-of-function research on it, as happened with R1, you can get a dramatic capability increase. In the R1 paper, they showed that they could take midsize open-source models to much higher levels of reasoning capability.

If you had asked the people who created those models what the best possible result would be if somebody fine-tuned them on certain math benchmarks, they might have said it wouldn’t get that good. Then suddenly there’s a huge unlock, and society has to treat it as the new norm.

I wish there were a way to enable access, research, and understanding while still giving us a way to take things back if we make mistakes. So far, we’ve mostly been protected by the fact that the AI systems available to us aren’t yet that powerful.

It seems realistic that in the not-too-distant future, something more powerful could be open-sourced and seem fine at the time. Then another post-training unlock could take it to a whole new level, and we’d realize that it had become a real problem. There wouldn’t be much we could do about it.

You have a good example of how to get the best of both worlds. One of the best open-source AI projects I’ve seen that fits this description is the recent work you did in partnership with SecureBio.

Tell us about the biology work you’ve done that illustrates how we might get the best of both worlds, at least in some domains.

For context, we supported the training of a metagenomic foundation model for early pandemic detection in wastewater. It can also be extended more broadly.

There are 2 or 3 interesting things about it. First, it took around $20,000 to $30,000 of compute to create a state-of-the-art frontier model for detecting pandemics earlier. We’ll continue training it in a distributed fashion to improve it, because there’s much more data we can train on.

The other interesting fact is that there are many defense-favoring open-source models that could be built. A lot of objectively high-impact, positive models haven’t been built yet. Scientific foundation models are a good example. If you look at something like AlphaFold, there’s very little misuse risk and enormous upside for humanity.

Even from a utilitarian perspective, the outcomes are heavily skewed toward the positive. The data for this model is mainly wastewater data, used to detect what’s in the water and identify pandemic outbreaks early.

COVID-19 was detected through wastewater surveillance. This kind of system can help monitor the spread and global distribution of risks. If we put these systems in place, we’re in a much safer world, irrespective of AGI.

COVID-19 probably didn’t happen because of intelligent systems. It happened through mistakes or misuse involving nation-state actors, including the United States and China and their involvement in gain-of-function research. Society lacked the antibodies and defense mechanisms to prevent it from becoming a major problem.

That’s how we also get to a utopian, safe AGI future. We look concretely at the risks, such as biosafety, and solve them. Even if there’s another COVID-19 or another pandemic—which there surely will be—that isn’t necessarily a problem because we have vaccines, early detection, and other defenses to prevent it from spiraling into a crisis.

The model was a collaboration with University College London, the Nucleic Acid Observatory, and the Safe DNA team. They did the research work, and we helped with compute and training to make sure a good model came out of it.

It’s a good example of how there are many more projects where, with millions of dollars of compute, you could transform scientific and human progress. This is central to our broader goal: using intelligence to solve everything else.

The most important thing to solve is science: solving diseases, natural risks, climate, energy, and other problems that will create an enormous amount of human flourishing and progress.

A key part of that is working toward autonomous AI research and autonomous scientists that can safely advance different areas of science. It’s ultimately determined by how much compute you can give them to reason, solve problems, and simulate them.

That’s a world we’re already starting to move into: a world in which AI massively accelerates science and scientific progress.

I’ve been tracking that general phenomenon in many different domains. It is ramping up everything.

What I found especially interesting about this model, MetaGen-1, was how thoughtfully it was designed. We’ve heard from many people, myself included, saying, “Do we really want expert virologists in everybody’s pocket?” That may not be a great technology to make freely available in a way that we can’t take back.

This model is almost the exact inverse of that. The key architectural decision is that it has only a 512-token context length. That’s obviously much shorter than any genome, so it can’t be used to generate full genomes.

There’s been a lot of discussion, although perhaps less research, around creating AI systems that are myopic. We don’t necessarily want these systems thinking through enormous, long-term plans or doing huge things that we don’t understand. We want them to do 1 thing very well: the thing we put in front of them.

This model does that. With 512 tokens, it can only do so much, but that’s enough to detect anomalies. You’ve created something that seems overwhelmingly likely to be safe to put in everybody’s pocket. The only thing people could really do with it would be to monitor local wastewater for anomalies.

This is a unilateral provision of a global public good, which is one of my favorite concepts. For the people who need it, it’s incredible. In terms of the balance between attack and defense, it seems extremely useful, but it’s difficult to imagine how anybody could abuse it. That safety is baked into the artifact itself.

When you imagine a general-purpose scientist, I’m not sure that property always holds. What other projects do you have in mind that have a similar set of properties?

That’s really the philosophy: figuring out the biggest contribution that has a disproportionate impact relative to the amount of compute it requires.

There are things you could do with hundreds of millions of dollars that would have less impact. The question is how to identify the low-hanging fruit for human progress and differential progress.

I’ve spoken with people in biosafety about this work, and they’re excited to open up a distributed pathogen-detection and monitoring effort. At some point, you could have millions of places analyzing samples—not on an individual-home basis, but more at the level of cities, airports, and similar locations.

More broadly, we want to build the key pieces and conduct the research needed to scale progress radically. We want to create systems with a human in the loop that can do science or AI research autonomously in a safe way.

That’s also central to solving superintelligence alignment. The only reasonable path to having an autonomous scientific system solve pandemics is to have the right guardrails and mechanisms in place. The benefits of open source and broad participation can help drive these systems in the best direction.

We have multiple things planned in this direction, including other scientific foundation models. A virtual-cell foundation model is 1 idea. Another is a simple autonomous scientist with mechanisms that allow humans to interact with the AI scientist.

Those are some of the most promising paths toward radical positive human progress. You need to do this iteratively: deploy in small steps, improve it, battle-test it, scale it, and make it accessible to everyone.

Let’s switch gears to distributed training. We’ve covered a lot of ground, but now might be a good time to dive into the technical details.

To remind myself and everybody else, why does distributed training matter?

There are multiple reasons. In the limit, it could create truly decentralized AI infrastructure that nobody can control. It’s also a challenge to build clusters beyond a certain scale. There’s the spare-compute notion and major implications for compute governance.

Your blog post on distributed training is 9 months old, but I still found it very worthwhile to read today. That’s a rare accomplishment in AI: having something remain worth reading 9 months later.

Let’s start with what makes distributed training hard. Why was it ever a question whether it would work?

It’s mainly about the bandwidth requirements of distributed training. Even distributed training inside a single cluster is difficult to do.

You need to make sure that, for every parallelization strategy you use to distribute workloads across all the GPUs in a data center, you do it efficiently. The different parallelization strategies have different memory and bandwidth requirements.

One of the techniques used in our approach is normal data-parallel training. You do a forward pass and a backward pass of the whole model, and then synchronize the gradients at the end using an efficient all-reduce operation across all the nodes in the network.

Other techniques have more memory requirements. One is tensor model parallelism, where you split the weights of the model across different GPUs. You then have to communicate between GPUs for every single layer in the Transformer model.

Another technique is pipeline parallelism. You have different stages, with each stage containing different blocks of Transformer layers. You only need to synchronize the final activation or hidden state to the next pipeline stage. That has lower bandwidth requirements than tensor model parallelism, but still significantly more memory and bandwidth requirements than data parallelism.

Those are the 3 main techniques. There are some other approaches that people have developed over the last couple of years, but those are the primary ones.

A lot of people working on distributed training across non-collocated clusters have focused on the data-parallel regime. In a normal setting, that’s still too communication-intensive to run with an interconnect slower than 100 gigabits per second for large models.

You need techniques that require less communication between nodes. One technique we’ve been working on heavily builds on the DiLoCo approach from DeepMind—distributed low-communication training.

You train on different islands of devices in a data-parallel fashion, but instead of synchronizing every training step, you synchronize every few hundred steps. You synchronize the pseudo-gradients and use an optimization loop to merge the results from all the different nodes.

It works almost as efficiently as centralized training, with some limitations on how many nodes you can scale to and where it works well. It doesn’t work as well at the beginning of training, but in the later stages it’s almost as efficient as normal data-parallel training.

Let me take that back from the top to make sure I understand it. You can correct me and elaborate on anything I simplify.

In terms of the fundamental challenge, data parallelism is the first type of parallelism. If I’m training a giant model and want to train on, say, 15 trillion tokens, I’m obviously not going to process 1 token at a time.

The reason for having large clusters is to parallelize the work. I can’t have 1 model do a forward pass, take the gradient update, and keep using that 1 copy of the model. I’d never get there.

Data parallelism means making a lot of copies of the model. You run the forward and backward passes on independent copies, each processing its own data. The aggregation step is where a lot of information has to move around.

The gradient is basically the same size as the model itself. If I have 100 billion parameters sitting on a server and take a backward pass, I have, in the naive approach, 1 adjustment to each of those 100 billion parameters.

If I have 10,000 GPUs, each of which has processed its own data and calculated a gradient, I now have 10,000 sets of 100 billion adjustments. I need to aggregate those into a cumulative gradient for the training step, apply the update to all the copies of the model, and then take the next step.

Tensor and pipeline parallelism are important because the model may be too large to fit on a single GPU. You need to split it across multiple GPUs, which adds complexity and overhead. But even if you had giant GPUs that could hold all the parameters, you would still have the data-parallelism problem.

You would still have to run multiple copies and figure out how to aggregate the changes that each model instance wants to make so the overall model can improve.

Is there anything else someone who won’t do this themselves but wants a good intuition about the problem needs to know?

That’s a great explanation. It probably makes sense to stay in the data-parallel regime because the other forms of parallelism are even harder to distribute.

As you said, the models also have major memory requirements. You can’t train a 100-billion-parameter model on a 4090 because the parameters alone are too large.

For training, you need the model parameters, the gradients—which are the same size as the parameters—and a large optimizer state used to update the model parameters. Those things take a lot of memory.

That’s why, in our 10-billion-parameter run, the largest model we could fit on a single node, given the memory requirements, was around 10 billion parameters. We then used data parallelism across the nodes.

The momentum term helps stabilize training. You could take the gradients without it, but then you have convergence problems because everything is thrashing around. The momentum term keeps individual updates from going too far in unusual directions.

You have 1 momentum value to track for every parameter. If the model has 100 billion parameters, you need to keep track of 100 billion additional values for the momentum, along with the other optimizer state.

Is DiLoCo as simple as realizing that you can just run more forward passes, keep track of things locally, and avoid aggregating every step, while the process still works? Or is there more to it?

It’s a very empirical result. The idea is not entirely new. There’s a lot of federated-learning literature in this direction, and the general algorithm is often called local SGD.

You can do local steps and then have an outer step. The DeepMind paper showed that this works well, up to a 400-million-parameter model, which isn’t large by today’s language-model standards.

We scaled it to 1 billion parameters, where it was still working, and then to a 10-billion-parameter model with Intellect-1. Unfortunately, at that scale we didn’t have a baseline run for comparison because the cost was too high for a small startup to run those experiments.

The general intuition is that you can do local steps and synchronize less frequently. It works somewhat less well at the beginning of training. At the beginning, local convergence is a little slower, and then it catches up to the baseline later.

I initially had the opposite intuition. When you’re learning quickly, it seems like maybe you would need to share the updates sooner. Why is the opposite true?

At the beginning, there isn’t much signal because all the gradients point in roughly the same direction. Later in training, there’s more signal coming from the different data-parallel ranks, and the technique works better.

It’s an empirical result. We’ve seen that local convergence is a little slower at the beginning, but it eventually catches up to the baseline.

There are still things to solve. Since Intellect-1, we’ve been working on 3 major components.

First, the communication requirements are still too large to train a 100-billion-parameter model with this technique. We need even less communication. We’ve been working on quantizing the gradients and using other representations.

Second, there are memory requirements. If you use larger nodes with multiple H100s, those requirements can be addressed, but if you want everybody to be able to join with a minimum of an A100 or H100 node, there are still limitations.

Third, there are diminishing returns in how many workers you can scale to. When you synchronize after 500 steps, the signal has been diluted if you have too many workers. We empirically found that Intellect-1 scaled well to around 16 workers, but a truly distributed Intellect-2 run would need to scale to hundreds or thousands of workers.

The intuition there is that each worker is somewhat overlearning in whatever random direction it happens to be exploring. If there are too many workers, the process becomes noisy, and those different over-learnings cancel one another out.

It doesn’t stop working entirely. We’ve scaled to more workers in our testing, but it becomes less compute-efficient than normal data-parallel training.

The goal is to accept some trade-offs because, in a globally distributed setting, the FLOPs may be cheaper. We can run on spot instances, and people can contribute idle compute. The FLOPs are cheaper, while the bandwidth is more expensive.

But we still want to be roughly as efficient as centralized training.

You mentioned quantizing the gradients. Is that basically rounding many gradients down to 0, so anything below a certain size doesn’t have to be sent across the network? What other tools are in your toolbox?

We’re working on several things to improve the algorithm, and we’ll submit a paper on it soon.

For Intellect-1, the values we sent after the local steps were not the actual gradients. They were the difference between the weights at the beginning and the weights at the end after all the local updates.

Those pseudo-gradients are relatively easy to quantize. For Intellect-1, we didn’t send them in 32-bit precision. We sent them in 8-bit precision, which gave us another 4-fold reduction in communication requirements.

For that run, it was enough to synchronize every 100 steps, giving us an overall communication reduction of roughly 400 times. That allowed us to train efficiently across the globe with the interconnects we had for a 10-billion-parameter model.

If we do more local steps and find additional quantization techniques, we could probably scale to much larger models.

The other thing that stood out to me was the DiLoCo scheme itself. I had been studying mixture-of-experts models, and I was interested to learn that the relationship between experts is generally very opaque.

You shouldn’t model the experts as domain experts in recognizable areas of human pursuit. Just like everything else, we don’t know why a token gets routed to 1 expert or a set of experts.

DiLoCo starts to segment data by longer sequences. You don’t have to move things around at the token level. That seems interesting both from an efficiency standpoint and potentially from an interpretability standpoint.

What if we sent all our science data to 1 subcluster and all our literature data to another? Could that allow the systems to do more local work before aggregating? Could it create an architecture with a separation of concerns, more like a traditional software project and less like a giant spaghetti black box?

That’s what a lot of people get wrong when they think about mixture-of-experts models. They assume the experts have semantic routing and are specialized in particular domains. That doesn’t appear to be the case.

Mixture-of-experts models are usually a more efficient way to compute a lower loss. That’s the normal way mixture-of-experts models work, including recent DeepSeek models. They route at the token level, sending every token to a different expert in the MLP.

The parallelization approach we’re using means that we can’t distribute mixture-of-experts models if we route every token to a different expert. We have to route at the sequence level. For every sequence, we route to a different expert.

Theoretically, that would allow routing to domain experts. Unfortunately, I’m not very bullish on routing by domain. It’s a mixture-of-experts model that doesn’t seem to work as well as the normal token-level approach.

I’d be interested to see more research replicating the DiLoCo approach with sequence-level routing and testing whether domain-level routing works as well. My intuition is that it won’t.

Is that because it seems too clever, and clever things don’t work, or is there a more specific reason?

A semantic separation between experts would be a huge win. In the DeepSeek V3 and R1 architecture, if I remember correctly, there are 671 billion parameters, of which around 37 billion are active at any given time. Only around 6% of the total parameters are used in a single forward pass.

If you could get sufficient semantic segmentation between experts, you could have a virologist package, for example. That could be removed from a model, and you could distribute a version that didn’t contain the virology knowledge.

It could be a way to square the circle. You could say, “Everybody gets what they need, but you don’t get the virologist,” and people probably shouldn’t complain too much about that.

I think it can work in the sense that you can build it. I’m not sure it works as well as simply allowing the model to learn.

Adding an inductive bias that says, “We want to route these tokens to these domains,” is probably less efficient than allowing the mixture-of-experts model to learn where to route them. Empirically, mixture-of-experts models don’t learn to route by domain. They route in a completely different way that’s not interpretable to us.

That’s unfortunate for interpretability research, but it probably would have learned domain routing if that were a better architecture.

There are 2 interesting points here. On the 1 hand, you can build architectures that are better suited to distributed settings. On the other, we’re entering a new scaling paradigm: inference-time compute.

That paradigm scales much better. You get more reasoning for less compute, and it’s almost perfectly suited for distributed training by coincidence. The R1 scaling paradigm is a good fit for a decentralized setting.

It has low communication requirements because much of the scaling comes through synthetic data and inference-time compute. From an interpretability perspective, the reasoning chains are useful because you can read them, even though you can’t always do that with models like o1 or o3.

As an end user, you can see how the system is reasoning. You can align with it, edit out parts of the reasoning, or add things such as safety checks.

To be realistic, it’s still a strange way of reasoning. A model may suddenly start reasoning in Chinese or drop random symbols that you can’t understand. Scaling laws don’t produce reasoning in the same way human reasoning works.

These systems may take different paths through a multidimensional space of mathematics and concepts. Those patterns and structures probably exist, but they may not map anthropomorphically onto human reasoning or language.

That’s the paradigm we’re in, and the paradigm we plan to scale: inference-time compute in a decentralized fashion.

On the practical question of why R1-style reinforcement-learning training is more favorable, you’re doing many more forward passes relative to the number of backward passes. In pretraining, you do a backward pass for every document or sequence processed through the model.

With reinforcement learning, you might sample many generations, take the best or worst ones, and only do the backward pass for the generations that are especially good or bad. That dramatically reduces the bandwidth overhead.

Is there a ratio in the R1 paper? I don’t think it said how many generations they used or what their best-of-n strategy was.

Unfortunately, DeepSeek is not always transparent about those details. The R1 paper leaves out some information about how they trained it, how the infrastructure worked, how long they trained, and what the communication requirements were.

In general, the DeepSeek-R1 process has 2 components. First, supervised fine-tuning uses reasoning chains generated by another model, along with answers from verifiers. That synthetic-data generation is very viable in a distributed setting.

You can generate reasoning chains across a split network on H100s or H200s. You can also use smaller distilled models for synthetic-data generation and create more supervised fine-tuning data.

The second stage is reinforcement learning. DeepSeek didn’t provide all the details, but in general, reinforcement-learning approaches for language models use many rollout phases.

You might generate 256 rollouts for a set of questions, then perform an update step after all those rollouts. You do the backward pass afterward, accepting or rejecting some of the generations.

We’ve seen minutes of forward passes for every backward pass, or potentially hours if there are many rollout phases. It’s still early, but the paradigm clearly seems much more favorable to distributed training.

There won’t be a single ratio of forward passes to backward passes, because it also interacts with curriculum learning. You need to choose problems with an appropriate success rate.

If you’re working on problems you get right 1 time in 1,000, you need 1,000 attempts to get 1 correct answer and receive a reinforcement signal. The sampling strategy matters.

You need problems that are hard enough to produce meaningful learning but not so difficult that you have to try 1,000 times before getting 1 right. Both the R1 work and the Kimi paper discuss algorithms that calibrate the difficulty of the problems.

What else from the blog post or your work since then is worth highlighting?

One thing mentioned in the blog post is swarm parallelism. It’s a paper by Maxime, whom we’ve worked with a lot and who is currently at Together AI. He’s done tremendous work in distributed training over the last 6 years.

Swarm parallelism means that you don’t only use data parallelism across the world; you can also use pipeline parallelism across the world. For small sequence lengths, that isn’t necessarily very communication-intensive because you only need to synchronize the final activation state of each pipeline stage.

For a Transformer model, and for the different hybrid architectures being developed, that tensor has the size of the sequence length multiplied by the batch size and the hidden dimension. For small sequence lengths, the tensor isn’t necessarily very large.

In theory, you can send that over the internet. But there are other things to be cautious about. In a data-parallel regime, you don’t have to think much about latency. In a single data center, it isn’t an issue.

Across the world, latency is a major issue. In data parallelism, it doesn’t matter much because you do local update steps infrequently. If you have 100 milliseconds of latency, it doesn’t add much overhead.

With pipeline parallelism, you have to communicate frequently with relatively small tensors. That makes latency much more important.

Pipeline parallelism is still a useful technique for scaling to larger model sizes, which is one of the limitations of how large a model you can currently fit on 1 node.

Could pipeline parallelism eventually allow truly end-user compute to be contributed to training projects? My laptop can’t hold DeepSeek V3 or R1, but it might be able to hold 1 expert or 1 layer. How far away are we from that?

It’s possible for smaller models. You still can’t train a very large model across a completely distributed network because you would need too many pipeline stages. There are also latency bottlenecks in a home-GPU setting.

The reason I’m somewhat bearish on doing this in a completely distributed way is that the research paradigm has been moving toward larger context lengths. We need long sequences for R1-style reasoning.

When you increase the sequence length, it doesn’t affect data-parallel training. You still only send the gradients, which don’t grow with sequence length. In pipeline parallelism, the activations grow with the sequence length.

If you try to train a model with 1-million-token sequences, as you might for an o1-style model—the activations will be too large to send efficiently.

There are better techniques. We’ve been focusing more on loading the optimizer state onto a single node. There are approaches such as ZeRO-Offload and more modern methods that make this efficient.

You can train a 100-billion-parameter model on a single node of H100s or A100s. You still won’t be able to train a 650-billion-parameter model such as DeepSeek R1 on 1 node, but you can at least enter the 100-billion-parameter range.

Can you describe offloading in more detail?

It means moving some of the optimizer state from the high-bandwidth memory of the GPU to other storage on the node.

You want to offload the optimizer state that takes up the most memory. In mixed-precision training with AdamW, you have the parameters and gradients stored mostly in 16-bit precision. But the optimizer state needs copies of the parameters and gradients in full precision, along with the momentum and variance, also in full precision.

That takes much more memory than the parameters and gradients alone. If you can offload that efficiently while maintaining good training efficiency, you can fit some of the largest models on a single node.

I’ve been thinking about a blog post by Steve Yegge about buying GPUs in the wild. He had worked at Google and then left to start something else before returning to Google. The post was illuminating because he said he had no idea how spoiled he was at Google. The infrastructure just worked.

Then he entered the real world and tried to buy compute from different clusters. At the time, your company didn’t exist, and the market was even more chaotic. It sounded like a jungle.

At what point do you expect the GPU ecosystem to be comparable to CPUs? The hyperscalers figured out how to abstract away differences in underlying hardware and create consistent containers. They don’t need to know what physical devices a workload is running on. They know the virtual environment, and that’s enough.

They also tolerate hardware failures. If 1 computer fails, they route around it. That’s why we have high uptime.

We aren’t there with GPUs. I’ve heard stories about 1 GPU failing and stopping an entire training run. That seems like a dramatic departure from the way large-scale web products were already being supported 10 years ago.

Is it simply that the software needs time to mature, or is there something more fundamental?

Fault tolerance is a huge part of it. It’s a problem on both the software and hardware sides of the GPU ecosystem.

If you train across 100,000 GPUs in 1 cluster, a node will fail every few hours. Maybe that will happen less often in the future, but current training frameworks aren’t necessarily fault-tolerant. The training run crashes when 1 node out of thousands fails, and you have to resume from the last checkpoint.

The big labs have figured out a lot of this. The others are racing to figure it out because otherwise they won’t be able to train models across the largest clusters.

For us, it’s even more complicated. We don’t only need fault tolerance; we also need to onboard different GPU nodes during training.

That’s what we’ve implemented in Prime, our training framework, which we’ve open-sourced. It’s fault-tolerant. If a data-parallel rank drops out, training doesn’t stop. It continues.

That’s why the big labs are interested in these solutions. The challenge is making fault tolerance work across a broad, heterogeneous network.

Over the long term, the software will improve, and hardware will become more like a commodity. That isn’t the case yet.

You mentioned the ReCa paper, which showed a 100-fold difference in reliability between different GPU providers. There will probably always be better and worse providers, but the ecosystem will become more commoditized as fault tolerance improves and hardware gets more reliable.

When you trained the 10-billion-parameter model in a distributed way, how much of your time went toward managing that mess—making sure that heterogeneous hardware worked and that nodes came online and offline?

Did you have service-level agreements with providers that they wouldn’t pull resources away from you? Or were you using preemptible nodes and taking whatever you could get?

How much of the return comes from managing that infrastructure versus the more theoretical information-management improvements such as DiLoCo?

For the initial Intellect-1 run, most of the work was hard-core engineering. We had to implement fault tolerance in the framework and cover all the other edge cases.

A node can drop out, then rejoin, or join in the middle of a training run. Other frameworks don’t really support that right now.

Our focus was primarily on engineering. To be honest, a lot of things went wrong initially. We had to fix things on the fly, including edge cases we hadn’t anticipated in a globally distributed setting.

The training stopped for a few hours on several occasions, and then we continued with the available nodes. By the end of the Intellect-1 run, we had a relatively stable solution. Many of the components were fixed, although I’m sure that when we start Intellect-2, we’ll still encounter issues and have to restart training runs because of edge cases.

That sounds like a decent metaphor for the AI phenomenon as a whole: there are a lot of edge cases we haven’t anticipated, there may be crashes, and hopefully we’ll recover and reach a stable equilibrium.

Let me try to summarize where we are on distributed training. There’s a mess of hardware management, which you’ve made substantial progress toward taming but haven’t fully solved.

There are also the fundamentals of information management and bandwidth. It seems like the low-hanging fruit has been picked: you can perform less communication-intensive training and still get acceptable results. Now we’re getting into more tangible trade-offs. You can do certain things, but convergence will be slower or efficiency will be worse.

In terms of scale, you’ve trained a 10-billion-parameter model, and you can see a path to 100 billion. A path to 1 trillion parameters isn’t yet clear. Is that fair?

I think there’s still a path to AGI, and possibly beyond, within this paradigm. That path doesn’t require simply maximizing parameter count.

The R1 model shows that scaling a different paradigm might get us to AGI without requiring trillions of active parameters. There are other pieces of the puzzle that are important to solve.

I think we’re largely on track to solve decentralized training. There will always be trade-offs, but those come with huge benefits. You can train across the whole world without relying on enormous amounts of energy in 1 place or an absurdly large cluster.

The big labs are all extremely interested in solving this. OpenAI, Google, and all the others are working on distributed training. It’s not only a project for the decentralized-AI ecosystem.

The big labs have a different setup because they sometimes have extremely fast interconnects between clusters. But even when you look at OpenAI, its clusters aren’t all next to one another. The build-out is spread throughout the United States.

Distributed training is necessary to win support from Congress for the massive build-out because every state and district needs a piece of it.

It also unlocks global communication. Fiber under the oceans, connections between continents, and systems like Starlink could theoretically allow compute to be located almost anywhere.

We’ve had people reach out because the cost of energy in the ocean is much lower. If you have wave energy in the ocean, suddenly a different infrastructure paradigm opens up. Compute doesn’t need to be located in established hubs; it can be placed in extremely rural areas.

That future is already beginning to play out. The question is how quickly the infrastructure build-out continues. It’s already in the trillions of dollars. I think it will reach tens of trillions.

Microsoft alone is spending $80 billion a year on data-center build-out. OpenAI is now committing more than $100 billion, and there are many other efforts that aren’t reported.

Governments have enormous infrastructure build-outs that aren’t well reported, including the United States government. There is more compute being built than people realize, and much of it isn’t publicly known.

DeepSeek is a good example. A Scale AI executive recently claimed that DeepSeek has more than 50,000 H100s, even though the number is underreported because of export restrictions.

It’s difficult to get an accurate number for what the Defense Department is building. My understanding is that the number of GPUs is already in the hundreds of thousands and may be approaching 1 million. There are enormous clusters that people don’t know about.

The same is happening in China and many other places. There will be specific locations where compute explodes.

ByteDance was one of the biggest providers of compute capacity to different platforms because it bought so many GPUs—more than 600,000 at one point—and sometimes didn’t need them. It made those GPUs available to the market.

These dynamics will continue as corporations and nation-states ramp up their data-center build-outs.

It would be quite something if seasteading became a thing because people wanted to put compute in the open ocean for wave energy.

We may be headed toward a cyberpunk future. The energy is already much cheaper in some places, by an order of magnitude. That will probably be a major factor.

Compute will be everywhere it’s allowed and where there are geographic advantages, such as access to cheap energy and ease of maintenance.

With Starlink, you could even have decent connectivity. You could have floating compute islands in the middle of the Pacific. That’s quite something to contemplate.

What does this mean for Nvidia? This isn’t a stock show, but Nvidia has blown up to a $3.6 trillion market cap. AMD is down by roughly a third over the last year and has a market cap of around $200 billion, giving Nvidia an 18-to-1 ratio.

What do you think about Nvidia’s prospects for continued dominance? A naive reading of what we’ve discussed would be that abstraction is coming. If abstraction comes, that seems almost synonymous with Nvidia losing the moat and profit margins it enjoys today.

Several things can be true at the same time. This market was basically nonexistent 5 years ago, and now it’s worth trillions. It could reach hundreds of trillions.

It’s difficult to comprehend how large this market will become. Everyone in the stack could become one of the most valuable corporations in history: Nvidia, ASML, TSMC, AMD, and others.

On the other hand, it’s possible that Nvidia’s current monopoly will weaken. Nvidia currently captures more than 100% of the profits in the AI industry because everyone else is losing money while Nvidia accumulates cash.

Over time, Nvidia’s margins may compress, but its revenue could increase by 10 or 100 times. It could still be a much larger business in optimistic scenarios.

AMD is the clear number 2, but it has struggled on software, ecosystem development, and other fronts. Nvidia has built a substantial ecosystem.

Nvidia also has a large free-cash-flow advantage, which it can reinvest into the next generation of chips and use to extend its lead.

The strongest competition may come from Google’s TPUs. Google has been somewhat secretive about them, and other big tech companies such as Amazon and Apple are also developing their own chips.

I wouldn’t underrate the chip efforts of the big tech companies. They can force adoption. Amazon has its own chips, and Apple has its own chips.

AMD may go from nearly 0% market share to somewhat higher market share, but it probably won’t overtake Nvidia anytime soon. There are also specialized providers such as Groq and Etched, which are close to 0% market share today but will likely gain some share.

If you fast-forward, the entire class of companies may be worth more. Margins will be compressed, but revenues will be much larger.

We also plan to contribute to the software stack and help commoditize it. In a perfectly efficient economy, margins eventually compress toward the cost of capital plus a risk premium.

That’s also what I hope happens with superintelligence: over time, the margins and opportunities get compressed by cheap alternatives.

Are you familiar with Imad Mostaque’s Intelligent Internet project? I did an episode with him recently. Of everything I’ve seen, you seem to have the closest vibe match with him.

We’re quite close with him. I can’t fully wrap my head around what he’s planning yet, but it’s evolving and hasn’t fully launched.

There are some differences and similarities. We’re focusing on peer-to-peer compute and intelligence as foundational pieces. He seems more interested in somewhat different components.

We’re focused heavily on decentralized training and making that work. His interests may be more orthogonal to that.

We’re also trying to ship small increments as quickly as possible, open them up to the open-source community, and build different pieces of the puzzle. He’ll probably do the same. He has a track record of building relevant open-source models and creating a community through Stability AI.

I think that’s one of his superpowers, and he’ll continue to build on it. We’re collaborative and in touch, as we are with many other efforts in the space. We try to figure out how different projects can support one another.

People can build agents and models on top of our network or Imad’s network. His network may need compute and draw it from ours. That’s the future I envision.

It’s similar to Ethereum. One reason Ethereum became powerful and won in many dimensions is that it has co-ownership of the network and protocol. It’s fully open source, and thousands of people collaborate with strong incentives to make it successful, rather than having 5 or 10 team members and a few shareholders.

That’s a paradigm in which the open and decentralized community competes against closed-source labs. That union is growing, including through foundational open-source projects such as Python and the Llama community.

All of these systems can build on one another. Imad can leverage our progress on decentralized training, and everyone else can do the same. That’s the strong tailwind for the space.

Open source has now officially caught up with closed source. Since R1, the gap has closed. Many people didn’t think that was possible even 1 or 2 years ago.

The perception has shifted. The 1- or 2-year lead that OpenAI had has effectively melted toward 0. It’s now measured in months, not years.

I’m not sure I’d go quite that far. I’ve had very good experiences with DeepSeek, but it’s only been a few days, and a lot of people are enthusiastic about it.

If you gave me only 1 choice, I would still take o1 over R1, and o3 is coming. There’s something there. I would also bet that Google DeepMind and Anthropic have systems at least at the R1 level, even if we haven’t seen them.

We’ve seen Gemini Flash Thinking, and Dario’s recent comments about what he’s seen internally at Anthropic suggest that they have something moving in that direction.

Going back to the long-term vision and the structure of the company, you have compute partners that contribute to the marketplace and can also contribute to decentralized training runs.

Are they contributing in kind? What’s the governance model? I’m fuzzy on what’s in it for them, what the incentive design is, and how you create sustainability at that level.

There are different phases. Until now, everyone has contributed in kind. They’ve been friends and sponsors, including Hugging Face, Oran Patel from SemiAnalysis, and others in the open-source community.

The broader goal is to create a public utility that anyone who contributes to can have a piece of and help direct. We’re finalizing the details and working out how it should be structured.

It will have a nonprofit foundation structure that governs the provision of this public utility. In many ways, that’s counter to a traditional startup corporation.

It’s a structure designed to provision a utility efficiently, openly, and in a decentralized way. It’s permissionless. Vitalik can’t stop Ethereum, even if he wanted to. That’s by design. He isn’t the only person contributing to it. Hundreds of teams contribute to the infrastructure, and everyone who uses it or submits a pull request participates.

That’s what we’re setting up. Our company will be 1 of many contributors to the public infrastructure and utility, but it’s very different from a traditional corporation.

Do you envision a token or currency in which I contribute compute and receive some sort of claim on future governance that I could potentially resell?

The broad goal is to create an alternative system like Ethereum that is permissionless and tokenized. We’re not ready to comment too much on the concrete timeline, but that’s the broader direction.

It raises a larger question about how economies and currencies will change in the intelligence age. The default outcome may be a continuation of fiat currency losing value or being heavily inflated through measures such as universal basic income, potentially funded by the windfall from superintelligence.

People may want to move into less inflationary, harder assets. A currency backed by compute is harder than fiat currency, which is backed by nothing. Even a cryptocurrency such as Bitcoin isn’t necessarily a hard asset if you look at its utility.

In the intelligence age, you may want to own a piece of a superintelligent system that can generate value. Through your ownership, you would have access to the compute and intelligence.

We’ll share more about how we plan to build that. It will need to have stages that guarantee that anyone can get involved and own a piece of it, while remaining genuinely permissionless.

It should be in the spirit of the early cypherpunk internet, not the late-stage platform monopolies of big tech. We lost the internet to big tech platform monopolies, and I want to make sure we don’t lose superintelligence to them as well.

That’s the default outcome if we don’t choose a different path.

You brought us to a perfect ending point. Maybe one final bonus question: What are you looking for? That could mean hiring for your team, compute contributors, or anything else. What kind of bat signal do you want to put out into the world?

We’re actively hiring across many roles, including AI research, general development, marketing, design, and other areas. Anyone interested should feel free to reach out.

We’re also actively collaborating with dozens or hundreds of open-source AI researchers, leading scientific institutions, and AI labs at universities.

We want to support high-impact initiatives with compute and hands-on support to build and scale their models. That’s another area where people can work with us.

In general, we’re happy to connect with anyone interested in going deep on this, collaborating, or contributing. There are many surfaces for collaboration.

People who have compute can contribute it to models or agents they’re excited about. That’s what we’re building toward: enabling people to participate in this system.

This has been excellent. Vincent Weisser and Johannes Hagemann, founders of Prime Intellect, thank you both for being part of The Cognitive Revolution.