DeepSeek 对决 Open AI——Emad Mostaque 与 Salim Ismail 谈 AI 现状 | EP #146

Last February, I said DeepSeek was one of my favorite AI companies out there. If you look at each of the innovations they made, they were largely engineering innovations. Do we see DeepSeek dethroning or reducing the valuation of these companies at all?

I think, in my opinion, it should increase the valuation.

Emad, DeepSeek was no surprise for you. Was this something you expected, or was it more like, “Wow”?

I think it was actually expected. Last February, I said DeepSeek was one of my favorite AI companies out there. They took the original ethos that we had at Stability AI, and they released amazing models open source.

I think when the AI community first started to see this was probably around the summer of last year. They released DeepSeek-Coder, which hit the top of the code rankings. They started by replicating Llama from Meta, then they broke forward. In fact, some of the algorithms they use now are the same algorithms they used then.

In December, they released DeepSeek-V3, which was the $6 million training-cost model. It matched GPT-4o and all these other models. It didn’t match o1 at that point, but we all thought they would figure out how to do it—and guess what? They did.

It felt like the internet broke over the weekend as the announcement was made. What was it that got everybody so hot and bothered instantly, since DeepSeek had been around for some time?

There was the base model, which was the ChatGPT-equivalent model, in December. That proved you could train these models at a fraction of the cost. The next thing was this reasoning model, R1, where, when you type, it shows you the reasoning. It takes a bit longer to think and has better-quality output. That actually came out last Monday, but it was this weekend that the narrative cascade happened.

Now you’ve got your mom and your aunt asking about it. It’s front-page news, and then NVIDIA crashed. I think what happened is that people remembered the early days of ChatGPT or Stable Diffusion for images—the immediacy of the response and that new paradigm.

When OpenAI released o1, this thinking model was amazing, but it was a bit like using ChatGPT. You put something in, it says, “I’m thinking,” and then it gives you a response because it hid the chain-of-thought reasoning. With R1, it actually shows, “This is how I’m thinking about this. This is how I’m breaking down the problem.” It feels like you have another person on the other side.

As more and more people used it and saw the performance benchmarks, it built up into this cascade because it was so immediately usable. They realized it was open source, so people took the smaller versions of it and started running them on their laptops.

If it had just been a closed model that didn’t have the chain of thought but matched o1, it wouldn’t have had the same impact. If OpenAI had released the chain of thought, I don’t think it would have had the same effect. It was this confluence of things that made people realize, “Oh my gosh, what is this new thing, and how has it been done?”

I have 2 thoughts. I love the timing that they launched it on the day of the inauguration, as a bit of a slap in the face to the incoming administration, saying, “You will sanction us to bits,” and here’s how the sanctions work.

The second thought I’ve had throughout the last 10 days or so is that we are expecting demonetization. As the power of these models is accelerating exponentially and blowing our minds, the demonetization should also surprise us in the same way. The fact that they’re able to do this at one-tenth or one-hundredth of the cost—how they got there is obviously an open question—but the fact that it has been achieved shouldn’t be a big surprise based on the curves we’re looking at.

Someone noted that it was actually the 5-year anniversary of the Wuhan lab leak as well, except this one was delivered—

No, we’re not going to go there.

I put out a blog that followed the DeepSeek announcement. This is just going to be the new normal. When Netflix ate Blockbuster for lunch, it didn’t mean that was the end of disruption. This is just going to happen over and over again—the speed at which heads are turning and snapping across every industry.

When we saw ChatGPT announce and get to 1 million users in 5 days and 100 million users in 2 months, people asked, “Can this ever be replicated again?” The answer is yes, and faster.

Emad, could you give us a quick rundown of how DeepSeek compares to GPT-4o, o1, and the other models? There are a lot of claims being made about how many GPUs it was created on, how much money was spent, and the size of the teams. Those comparative numbers made it a big deal. If it had been an equivalent model but wasn’t created at a fraction of the time or cost, it would not have hit as hard as it did.

I think the shock was the order of magnitude. We can break it down a bit.

o1 was an evolution of ChatGPT that suddenly got to IMO-medalist level or top-coder level—roughly the top 1% of coders—because it could think longer. This was a key breakthrough. Mark Chen from OpenAI has actually said that what DeepSeek figured out was pretty much what they were doing at OpenAI in November.

First, you had the model that matched ChatGPT. Then they figured out how to make it think longer. But the main thing that shocked people initially was that it was 96% cheaper. Software usually has an 80% margin. We don’t know how much OpenAI charges, but they’ve got this hammer, which is a large number of GPUs. They’ve never had to work in a constrained environment.

Sometimes you are a bit price-insensitive, particularly because the cost of running an o1 query to solve a math problem or a legal problem is still small. But this was 96% cheaper than o1.

The second thing was that it could be released anywhere. The headline number for the original model that R1 evolved from was probably only $200,000 to train, which was another shock. Last year, OpenAI spent $3 billion training models.

To give you an idea of that, there were accusations that DeepSeek had 50,000 of these chips, not 2,000 like they used on the training run. DeepSeek never claimed how many chips it had. It just said, “We needed 2,000 for this training run, and we used them over this period of time to build a model that looks like this.”

Those of us who have built these models know that the numbers check out. This is why some of the reaction has been interesting. People say, “They have far more GPUs,” or, “They have hidden GPUs.”

The GPUs they have are H800s, which were the top-end NVIDIA chips, though not quite the top end anymore, with a slightly reduced interconnect. The way the chips speak to one another is a bit slower. We had this issue at Stability AI, where we built one of the largest supercomputer clusters in the world, but our interconnect was one-quarter the speed of other people’s because that was all we could do.

We were competing against the biggest companies, but we built some of the best models in the world. They wrote the lowest-level code in PTX, which is CUDA at a level lower, to overcome the limitation.

They engineered the hell out of it because some of them are former quantitative hedge-fund managers. If you look at each of the innovations they made, they were largely engineering innovations, which is very interesting for our mental model.

What is China amazing at? Engineering innovation. Look at BYD and Xiaomi. It shouldn’t be any surprise that, as you move from research to engineering, you would see this leap ahead. All the numbers check out. You see the cost reducing. I think they probably have 10,000 chips in total, but that’s not more than many startups in the Valley, to be honest.

I had a conversation with Kai-Fu Lee recently, and we were talking about how the US government has been restricting Chinese companies from getting NVIDIA chips. All that has done is create evolutionary pressure for them to do much more with much less. This sounds like a perfect example of that. It’s Darwinian in its developmental force.

If all you have is a hammer and you have large amounts of GPUs, the way this works is that the GPUs compress the knowledge. It’s like pressure-cooking a steak and making it tender.

Instead, you look at better data, better algorithms, and more efficient systems. If you can’t scale on compute speed because you don’t have the chips or the speed—because as you go from 1,000 to 2,000 to 10,000 GPUs, you can parallelize and get more speed—you instead optimize memory as the key thing.

Classical models are very dense models, like Llama 70B. DeepSeek-V3 has 640 billion parameters, but only 30 billion of them are activated at one time. They scaled on memory, and that is cheaper than superfast silicon.

These constraints are the key. We’ve seen it again and again: if you don’t need to worry about constraints, you build inefficient models. If you have to worry about efficiency, necessity becomes the mother of invention.

Wasn’t the CEO labeling the data and going through all of that? That adds so much juice to the model.

Models are just data. If models are figuring out the interconnections, it’s like having a bad curriculum. Then you have bad data.

The models we train right now are trained on terrible data—14 trillion words in the case of DeepSeek and Llama. You don’t need that much data to build an expert model, but if you have a large amount of compute, it doesn’t matter.

What we’re seeing now is data improvement. In fact, the data they used to turn this from a base model into a thinking model—which they then transformed the Llama model with, and Qwen as well—was all synthetic data.

We’ve moved to a point where they’ve figured out what the right type of data is. Typically, those who make breakthroughs don’t send the data off to the Philippines and try to make up for it with engineering scale. They look at every part of the process and simplify, simplify, simplify.

This echoes what we’ve seen in engineering. How did the engineering marvels happen at Tesla or in Chinese companies? They look at every part of the process and simplify it.

We had David Sacks over the weekend with this commentary. He said there’s a technique in AI called distillation, which you’re going to hear a lot about. It’s when one model teaches another model. The student model asks the parent model a lot of questions, just like a human would learn, but AIs can do this by asking millions of questions. They can essentially mimic the reasoning process they learn from the parent model and suck the knowledge out of it.

There is substantial evidence that what DeepSeek did here is distill the knowledge from OpenAI’s models, and I don’t think OpenAI is very happy about that. What do you think about it, Emad?

It’s a bit like calling the kettle black. “Don’t train on our data.” Distillation is nothing new, and there’s no way to stop this from the model basis.

If you look at what the paper says and what’s reasonable, they had this version, R1.0, that created its own data. It’s familiar with AlphaGo, AlphaZero, and MuZero—these reinforcement-learning models that outperformed humans at Go. You can feel like maybe we’re all Luddites, because AI is coming for all of our expertise.

It’s inevitable that this will happen, but I don’t think they deliberately went in and did that. OpenAI’s o1 outputs cutting-edge answers, but it was missing the chain-of-thought reasoning step. We’ve now seen that, as you take the chain-of-thought reasoning from R1—and from Google’s new Gemini Flash Thinking model, which is now at the top of the leaderboard—that’s what you really need to optimize the process.

I think DeepSeek created its own synthetic data. But as it looks at all of the internet, there will be some OpenAI data in there. We’ve even seen that with Llama, Gemini, and others. Sometimes you ask a model who made it, and it says OpenAI, because it has taken in so many of those data points.

We’ve got an interesting impact on Wall Street that occurs on Monday morning. NVIDIA got hit massively, and I’m sure OpenAI was reeling. Salim, how do you think about this? This is what people respond to.

Markets are psychological, and everybody goes, “Oh my God,” and everything crashes. There’s no question that NVIDIA’s chips are overvalued, but my guess—and I’d love to get Emad’s take on this—is that the overall demand for AI is exploding so much that it’s not going to make a big dent in demand for the chips.

NVIDIA is still up 100% over the last year. It’s not down that much. No one knows what’s coming, but what’s the market size of this?

The displacement is the displacement of all knowledge labor. Just as the industrial age replaced muscles, now you’re replacing brain cells. That’s a huge market.

We have global GDP going into 2025 at around $110 trillion. Half of it is physical labor, and half is effectively intellectual labor. It is massive. This is the technology—the intelligent capital stock—that will define productivity.

It’s very difficult to get a handle on how this will go. People have been talking, as Satya Nadella and others have, about Jevons’ paradox: as the price falls, demand increases. I think that’s what we’re going to see.

If you look at NVIDIA’s strategy, they’ve been moving to these fully integrated data-center boxes—the GB300 NVL72—and this new thing, Project DIGITS. If you’ve seen a Mac mini, it’s like a Mac mini that sits on your desktop. It has 128 GB of VRAM and a petaflop of AI compute for $3,000.

Two of them can run R1, so with that you have R1 at home. It’s an entire baseboard created for that. It doesn’t even have a fan, and it only pulls 200 watts of electricity.

You made a comment earlier about the amount of energy and cost you think it would take to build DeepSeek’s model. Could you speak to that?

It’s fairly insane. When we bought our first major supercomputer—this would have been about the 10th-fastest in the world publicly in 2022 at Stability AI—it had 4,000 A100s, which were the top-of-the-range chips. The interconnect was a bit poor, but it was still big. Each of those chips used about 400 watts of electricity. It was a big beast.

You may recall the recent NVIDIA announcement. Jensen had this shield, which was their new integrated box: the NVL72, with 72 chips super-interconnected. The interconnect on those chips is equivalent to the bandwidth of the whole internet.

Wait, wait, wait. Can you repeat that?

The way the chips communicate with one another—the total bandwidth is about 6 petabits per second, which is the bandwidth of the whole internet. They figured out how to integrate everything, so you don’t have this chip-to-chip interconnect. You just have this big wafer with 72 chips on it.

How much electricity does it use?

One of those boxes pulls down 100 kilowatts of electricity.

You have 2,000 H800s, the slightly hobbled chips that the Chinese have, and DeepSeek used them. I think it would require 10 of these boxes at most, probably even less, to create that model. Each box costs about $3 million. In fact, I think it probably only costs 4 of these boxes.

Even if you pick the upper bound, the total energy required to train the model is 1,000 megawatt-hours. It’s about $15 per megawatt-hour in the United States now—about $20,000 total. You could literally train it off a small solar farm in your backyard.

A big solar farm.

A big solar farm, yes. One box pulls down a decent amount—about 100,000 kilowatt-hours of energy—but you could definitely run DeepSeek R1 on solar panels.

If we look at the direction this is going, it’s still not optimized. Next year, you should be able to get an o1-level model on your smartphone that pulls at most 20 watts of electricity. It’s less than a dollar per watt of solar power.

That doesn’t make sense when you look at what these models are capable of.

It makes sense when you think about how much energy your brain pulls: just 20 watts.

We have a huge efficiency curve to ride to get there. By next year, you will have these o1-level models running on 20 watts, which is human-brain-level energy consumption. These models are already at a PhD level in so many areas.

That doesn’t compute when you look at discussions about Microsoft bringing Three Mile Island back as a nuclear-power reactor. AI’s energy is going to use everything. I think 60 gigawatts of electricity is coming online for data centers in the US over the next year or so.

Yet when we get down to the actual numbers for a given unit of intelligence, it’s a few watts and a few pennies. Before this, it would take entire teams. How many watts of energy would their brains and infrastructure use?

We’re not ready for that.

Salim, you asked a question about how challenging DeepSeek actually is to OpenAI, Meta, NVIDIA, and the rest. What are you thinking?

I have 2 questions. One is whether the fact that DeepSeek is Chinese—and companies may be reticent to put their information into it—makes a big difference. My guess is that the answer is no, because it’s open source and you can run it locally. Is that correct?

You can, but most people won’t. You give your code and all your information to TikTok. No one knows what happens with all this data.

The versions you can run locally are actually the distilled versions, not the main version. It’s quite difficult to run the main version locally, so I think there’s a geographic-arbitrage advantage that the incumbents still have. That’s pretty powerful.

The question I was asked by everybody on X and by my friends was whether this would go down the same path as TikTok. When OpenAI first came out with ChatGPT, you had all these companies—and a lot of banks—saying, “You cannot use ChatGPT in the office. We don’t want OpenAI to own our data.”

There was an immediate privacy concern, which is still valid. Are we going to see the same thing with DeepSeek, where people say, “No, you can’t use DeepSeek. We’re worried about the data and where it’s going to reside”?

You’ve seen a couple of announcements. Perplexity announced that it’s using DeepSeek locally, fully on American server farms. You’ll see that type of thing, even if they’re running the larger models. There will also be APIs.

You’ve seen OpenAI announce ChatGPT for government use by 19,600 federal employees. This is the direction things are going. I think you’ll have 4 different types of AI:

Superintelligence or AGI that you call upon when needed; your personal AI, such as Google or Apple AI; open-weight models like DeepSeek and Llama, which are useful but perhaps not suitable for regulated industries; and open-source, open-data AI, where these decision-support systems need to be fully understood.

You need to know what is inside these systems and how they were created, because you can poison models with inherent biases. There was an Anthropic paper we discussed before, Salim and Peter, called “Sleeper Agents.” With a few thousand words out of 10 trillion, just one word can turn a model evil or change its behavior completely.

It’s similar to the fact that most of the transformers in the US are built by Chinese companies, and no one knows the control software. Do you want the transformers that run your business to have that potential threat?

That’s what we’re doing now at Intelligent Internet: building out that open-source stack for regulated systems.

Peter, I want to dive into what you’re building with Intelligent Internet, because it has one of the boldest visions I’ve ever seen for supporting humanity. But first, do we see DeepSeek dethroning or reducing the valuation of OpenAI, NVIDIA, Meta, Google, and the other companies? We saw it for a day, but is it valid?

In my opinion, it should increase their valuations. It’s bringing forward the time of mass intelligence—intelligence too cheap to measure.

If you look at OpenAI, what Sam Altman has done masterfully is build a platform with 300 million or 400 million users. What is AI in most people’s minds? ChatGPT. Gemini and Claude don’t even register to many people.

If the cost comes down, that’s good for him. This is the Zuckerberg school of thought: Why did Meta open-source Llama? Because it uses 10% of their GPUs, and if there’s a 10% performance gain, it pays for itself.

OpenAI will use whatever they can. Most of their models don’t have brand-new algorithms; they’ve borrowed from Google and many others. There are no real secrets in this space, especially now that noncompete agreements are not enforceable in California.

What is OpenAI as a company? They were in this pretraining, massive-compute stage. Now that’s becoming commoditized. People can pretrain models like xAI and others, and pretraining may not require as much compute because the data is getting better and better.

It becomes about intelligence refinement from seeing how people use the model. This is the operator paradigm, where OpenAI can run your computer or MacBook. You can let it take over and book your holiday. That’s the next stage, and I think they’re well set up for it. Their costs should decrease.

OpenAI made $3 billion in revenue last year and lost $5 billion, of which $3 billion went toward training models. If you don’t need to spend as much on training models, that’s good.

The feedback loop of people using the model, because OpenAI has so many users, gives it a pretty good edge.

I think it’s that, and then you use these models. OpenAI has about half a million GPUs coming—the B-series, the VR-series, and others. You can now make those run sequentially to build even better data, map it, feed it back into the models, and optimize them.

Classically, computing systems were not parallelized; they were sequential. We’ve had this period of big clusters, but now it’s about swarms of models and agents solving tasks. The models are good enough, cheap enough, and fast enough.

That’s the final thing about DeepSeek, just as it was with Stable Diffusion for images: good enough, fast enough, and cheap enough. That trifecta causes these massive adoption curves.

When this was announced, we heard that Mark Zuckerberg created 4 war rooms of engineers to try to decipher what was going on and how to use it. It really is an AI arms race, where everybody is surfing on top of each other’s advances and accelerating everything.

What I found fascinating, and I’m curious about this, is the size of their team. They did this with relatively few people, and OpenAI had a 200-person team during its earliest days as well. How do you think about team size and the ability to create something disruptive? Is too big bloated, while small and nimble is better?

I think a core team of about 100 researchers is ideal. Beyond that, it gets bloated.

At Stability AI, we had 80 researchers and developers, including 16 PhDs, and we achieved state of the art in image, video, and every other modality, including multilingual models. We had 300 million downloads on Hugging Face, making us the most downloaded company and the most popular open-source company while I was there.

Once we scaled past 150 people, things started to break down. It is about rapid iteration and trying new things. Research should be an innovation center, not a cost center. You start to have too much compute and too many other things.

OpenAI did its best work when it was smaller, but it scaled up and still does good work. It is a question mark now because it has become an organization. As Salim knows, once you get past that level, it’s very difficult to maintain innovation.

You end up with a problem. Either you have top-down control structures that slow down innovation, or you let everybody do whatever they want and get a lot of duplication. You have to manage that tension.

It’s fascinating that 150 people is the Dunbar number. Anthropologically, we’ve found that this is a pretty solid, reliable threshold.

I do think, to back up Emad’s earlier comment, that OpenAI has a lot more people than it really needs because it has so much money. It can throw bodies at things. Now this will force it to be a little more efficient.

I also believe this is good for the overall market because a rising tide lifts all boats. I think we’re going to end up with a balkanization, though, where Western companies won’t want to use DeepSeek-type models.

I can’t imagine a major Indian state enterprise wanting to use a model like that for security reasons. Then you have to develop homegrown models, and everybody ends up with its own models in different ways. You get a splintered effect.

I want to dive into China for a moment longer, because part of the announcement wasn’t just a cheaper open-source model. It was this level of innovation coming out of China, which rocked people. The majority of the world doesn’t see China as the hotbed of AI innovation that it is.

There are 2 levels of competition going on right now. There is competition among companies—6, 7, or 8 major AI companies vying for the number-one position—and then there is competition among nations.

Saudi Arabia wants to be at the top of the stack, committing hundreds of billions of dollars, followed by Qatar and the Emirates. But the US and China are really going at it.

Emad, how do you think about the US versus China in that regard?

I think this is the real reason NVIDIA went down—or perhaps because Jim Cramer said to buy NVIDIA the previous week. One of those things happened.

We’ve seen China trying to build chips domestically. Intel is probably in play as an acquisition target.

Intel is definitely in play. It’s fresh meat on the table, and everybody is figuring out how to chop it up.

These chips are getting extremely fast and capable with NVIDIA. People talk about AMD, but AMD chips are almost impossible to use because the software isn’t there. There are bugs, and it takes a few generations to become stable.

NVIDIA chips work, but Chinese chips also work. The DeepSeek model’s API was being run on Huawei Ascend 910 chips, which are a few generations behind in efficiency, but they work.

Similarly, China has 2 exascale computers—2 of the fastest supercomputers in the world—built in a completely different way: OceanLight and Tianhe-3. China just built them at scale and in bulk.

The issue is that the means of production and the means of productivity of a society—which traditionally were capital stock, industrial capital stock, and intellectual property—will now be chips. How competitive you are in the world will depend on how much computing intelligence you have.

I think the US has realized this, along with how much energy you have to throw at it. So it’s “drill, baby, drill,” reshore as much of this as possible, and create incentives to do that. They’ll take tariff money and put it straight back into Stargate-type initiatives.

What do you think about Stargate?

The $500 billion is the total cost of ownership, which is fairly well known. It’s probably closer to $100 billion when you back everything out, which feels small these days. It’s actually a lot of money, but compared with the 5G rollout, it’s less money than we spent on 5G.

This is more important than 5G. Compare it to the order of magnitude of the mythical Los Angeles–San Francisco railway. There’s already a kilometer of that railway built, Salim.

Salim, did you see the article this morning from Reuters? Alibaba released its AI model and says it surpasses DeepSeek. The unusual timing of Qwen 2.5-Max’s release points to the pressure DeepSeek’s meteoric rise has placed on not only overseas rivals but also domestic competition.

This speaks to the democratization of AI. Everybody will end up creating a bunch of models, and I think we’ll end up with a bunch of very specialized models.

I remember Eric Schmidt’s comment that you’ll end up with a specialized AI that is the world’s best physicist, and another that is the world’s best biotech expert. That AI can be replicated infinitely.

What do you do with deep specialty on the human side? That is the bigger question around all of this. The models are just going to keep getting better and better.

The models are good enough, cheap enough, and fast enough. The other Qwen model—the vision-language model—performed at the level of Anthropic and GPT-4o on visual understanding. The ones they have coming next will control your computer.

Anything that can be done on the other side of a screen, AI can do better for pennies this year.

There’s a lot of conversation going on across Silicon Valley and the White House about the US–China AI wars. This is the biggest crisis we have coming, because we’re heading into a future where every AI leader says AGI is 3 to 5 years away.

There are 2 levels of competition: competition between companies and competition among nations. If you develop a digital superintelligence before your corporate or national competitor does—even by a little bit—it could be devastating.

Emad, how do you think about the US versus China in that regard?

This is the future we’re heading into. Every AI leader I can think of says AGI is 3 to 5 years away. Sam Altman just said it could be next year, but let’s say within the next 3 to 5 years. Everyone is saying it—Dario Amodei, Demis Hassabis, myself, and others.

If you think about it, that’s crazy. Everyone says it’s coming, and there’s this concept of AGI or ASI as a pivotal moment when one entity could shut down China. You just turn it off. That might happen, and we still don’t know.

That’s why you need to start preparing for it. Sundar Pichai at Google said, “Why are we building out all these GPUs?” Because we can’t afford not to. That is the game theory. You can’t afford not to build AGI if everyone else is building it.

Before AGI, there’s something we can think of as a megachef that can come up with any recipe and outcompete all of us. What we have right now are amazing cooks that can follow recipes and do jobs better than humans.

The Unitree robots were doing that Chinese dance with the fans yesterday. I don’t know if you saw that. I’m going to have the Unitree robots at the Abundance360 Summit, and it’s incredible. They’re $6,000 for one of their mid-tier models.

That’s $1.50 an hour when you factor in depreciation, energy costs, and everything.

I have it pegged at 40 cents an hour. It’s insane.

It really is.

My kids will buy one just so it can clean up their rooms. That will be the most expensive thing it does.

When you talk about AGI in 3 to 5 years, what do you mean by AGI? The best framing I’ve seen involves multiple tests, like the Wozniak test and the IKEA test. What is your framing? What do you consider to be AGI?

I think it’s probably a complex system that can outperform a team.

Before that, I had this idea of artificial remote intelligence. You can’t tell whether there’s a human or a computer on the other side. Is it your remote worker?

That’s the most natural way for this to come in. You call a company and they put a bunch of people on the phone. We now have the technology for you to have a Zoom call with someone who is 100% a robot.

Peter, your worker is plugged into Slack and joins you on Zoom. We’re living in a world of distributed workforces. If you have an AGI able to plug in, take a role, fully read all the email and Slack traffic, and be up to speed instantly, that’s an exciting world.

It’s exciting, but at the same time, that’s the first level of disruption. You won’t need business-process outsourcing anymore.

The nature of the firm will change because these systems will be superchefs, not cooks. They won’t make mistakes, or they’ll learn from their mistakes.

Once they have low communication overhead, the next step is teams of agents. Independent agents will have a task and be able to get resources toward that task. This is why Wyoming’s DAO law and similar developments get interesting.

The step beyond that is the ASI concept, which we can’t really define. There’s a big takeoff where it has capabilities beyond human teams and organizations. It can invent incredibly quickly.

What is going to be the impact on physics, biology, and pure science? Will it take us way beyond what we can currently imagine?

Dario Amodei was on video from Davos saying that, in the next 5 years, we’ll make 100 years’ worth of progress in medicine and biotechnology and double the human lifespan. That’s extraordinary commentary to make publicly.

One of the most fascinating things about the last week is what happens when you use o1. You can’t upload files, which is annoying, and it isn’t very creative, but it is thorough.

With R1, because it hasn’t been tuned and made safe in the same way, it’s actually very creative. Someone took the code base for R1 and made it twice as fast. Other people have put together academic papers, and it has synthesized them into new reinforcement-learning algorithms.

That may indicate that these systems are becoming less safe, but the upside is that they may be becoming more creative. Again, these are the levels: Are you an amazing cook? That is the disruption of the labor market, especially for anyone working behind a screen.

Are you an amazing chef? That takes us into the AGI-as-a-team and ASI concepts. That doesn’t feel 3 to 5 years away to me. Given all these exponential curves, it feels much quicker, and very few people are preparing for it.

We’re going to see disruption after disruption, and our financial markets aren’t ready for this. We’re going to see the energy markets disrupted as well.

One implication of AGI and ASI may be new forms of energy sources, which could potentially topple our petrodollar system and destabilize government revenues. We’re going to see massive disruption.

Have you ever seen the chart of GDP per capita versus energy per capita?

Yes. It’s basically a straight line.

It correlates with health as well, along with many other things that could be completely disrupted. To make the best film studio in the world in a couple of years, you could do it anywhere with solar power. Science could happen in Guatemala or anywhere else.

It’s an uplift of global aggregate capacity versus the brain drain we’ve had toward the West. It’s massively redistributive.

Think about your capital stock—your intellectual and physical capital stock. It’s going to be massively redistributed, and our economies are not set up for that. Productivity used to be a function of labor, and labor was a function of energy. That correlation is about to break for the first time ever.

We’re moving from an energy economy to an information economy. The data sets and information you have will be paramount.

We need to start asking really big philosophical questions: What do we want all this to do? What do we want to be? What activities and functions do we want to perform as human beings as the job market disintegrates in front of us?

I still have my trepidations about humanoid robots, but once they show up with feedback loops and built-in large language models, you have a fully functioning robot that can do many varied things. Suddenly, you don’t need a gardener or a plumber—or many other types of workers.

I’m using those examples tongue-in-cheek because those are probably the ones you need the most, but many functions, such as aircraft maintenance, will be done much better and more precisely because of access to information.

We talked a couple of episodes ago about the fact that an avatar of you or me, Peter, may be more reliable because it has full access to everything we’ve ever said, rather than what we can hold in our brains. It could be more charming, more compelling, and even better looking.

How do we navigate that? This is where Emad’s philosophical perspective becomes important. The displacement of labor is just the starting point.

Before we get into that, I want to hit on a couple of questions, Emad. What do you think is the best-case scenario for AI in 2025? What are we going to see by the end of the year that will make people look back and say, “That was amazing”?

I think the video technology has reached the point where we can remake season 8 of Game of Thrones, so that would be quite good.

Focusing on that, how dead is Hollywood?

It’s completely rewired. The energy required to make a movie is massively reduced. At the same time, people may be able to be more creative.

The video-game industry went from $70 billion to $180 billion over the last decade, and the average score on Metacritic went up 5%. The average IMDb score is 6.3. Hollywood went from $40 billion to $50 billion.

Maybe it transforms into new types of media. With something like Kling’s Elements feature, you can take a scene and generate new scenes. We can create storylines. The average film shot is 2.5 seconds; it was 10 seconds a few decades ago. We can do 2.5 seconds perfectly now with almost perfect control.

Let’s say it will take a year or 2 before anyone can do this. A suitably dedicated studio could create a full episode by the end of the year.

What else are we seeing in 2025?

Music is pretty much solved on the media side. If you use the new Udio, the next generation they have coming is insane.

On medicine, we’re already above the human level. We outperform humans on empathy. Medical chatbots can help people through their journeys, particularly with mental health. I think we’ve reached the critical point where the models have gone from not good enough to good enough.

I think we’ll see the first breakthroughs in science, with novel things generated with the aid of o3-type models. This is test-time inference. I want to call it “thinkference,” because that’s a better way to describe models thinking longer.

Those are probably the biggest real impacts. Maybe Siri won’t be so bad anymore.

I can’t wait for Siri not to suck and for Alexa to become useful. I’m shocked Amazon hasn’t done this already. They were originally going to put Anthropic behind Siri and Alexa and power them properly, but it looks like that has been delayed.

They’re building out 1 million Trainium chips with their specialized chip, so good luck to them with that.

What is the worst potential outcome for 2025?

The complete destruction of the BPO market, which will reverberate throughout the global economy. Business-process outsourcing will be disrupted because, when you use Operator—the technology that takes over your computer—it’s a bit rubbish now, but this is the worst it will ever be.

Anything on the other side of a screen will be displaced this year. It will be parallelized.

This is actually part of the DOGE idea of getting workers back into the office. Being in person will be good for your job right now because if you’re remote, you may be the first to go.

Define BPO for people who haven’t heard the term.

Business-process outsourcing—outsourcing to India, call-center workers, or programmers. AI is already better than almost any Indian programmer who is outsourced right now.

You’ll see the impact on those economies before you see it in the US.

I think it happens in 2 phases. In phase 1, you have this massive downside. In phase 2, the really good people show up and generate much more code, because there is so much more code to be written.

But it will have a very detrimental effect on software maintenance, support systems, and similar functions. Those will go out the window very quickly.

I had Marc Benioff on the podcast a couple of weeks ago. He was saying that with Agentforce, Salesforce isn’t hiring as many engineers and is repurposing older engineers. He has increased productivity by 30%, and that will skyrocket from there.

If you look at Lovable, Bolt, and Cursor, they take you to a decent level. They can build whole applications and stacks, and they’ll just get better as the base models get better.

One of the things we started doing for nonengineers who apply to work at our company is requiring them to take a 30-minute Cursor course. It doesn’t matter whether they’re in HR or another department. Then they have to tell us how that course changed their view of the world.

What does the course teach somebody?

It teaches them how to build an HR application or any other application just by talking to the system. It builds the application almost live.

You can do this today in ChatGPT with Canvas. You can build a React application live, replicate the entire UI screen, or build an HR application. It generates the application, and you talk back and forth with it.

That base-level increase in capability will cause a realignment. But the downside is that there are real jobs and real people who have to think about what comes next. They have to become experts in AI-assisted work, and they have to be in person. Otherwise, they’re going to start being disrupted.

Last summer, 38% of IIT placements in India were unfilled. It was crazy.

It’s worse than that. One encouraging thing I’ve seen in the US is that we’re hiring far fewer top-flight MBAs. Hopefully lawyers will be affected too.

Harvard is way down on employment this year, isn’t it?

This is just the beginning. I don’t think people are ready for the level of societal disruption that’s coming.

We can process it because it’s lots of little S-curves all across the economy. Every teacher in the world has had to ask, “Can we use ChatGPT for homework?” Every HR department and engineering department is asking the same question.

It’s still not mainstream, but it’s hitting the headlines more and more. There’s a disconnect beyond that. It was like COVID: those of us who were paying attention saw it coming, and we said, “This is a step change.”

Until Tom Hanks got COVID, the world didn’t realize it. What is the Tom Hanks moment for AI? Is DeepSeek the Tom Hanks moment, or will it be something else? It’s coming.

It was about 13 years ago that I had my two kids, and I decided to double down on my health. I go every year to Fountain for an upload and digitize myself with about 200 gigabytes of data so its AI system can catch disease at inception. I also use Viome, whose metatranscriptomics technology helps me understand what’s going on in my body, which bacteria are producing which proteins, and what foods are best for me. Finally, I use OneSkin, developed by four PhD women who found a 10-amino-acid peptide that helps remove senescent cells in the skin.

Let’s jump into safety. An article in Fortune today said that an OpenAI safety researcher had quit, claiming that AGI races were too risky to gamble on.

The quote was: “An AGI race is a very risky gamble with huge downside. No lab has a solution to AI alignment today, and the faster we race, the less likely it is that anyone finds one in time. Even if a lab truly wants to develop AGI responsibly, others can still cut corners to catch up.”

This was from Steven Adler, who left OpenAI. He’s one of many people who have left OpenAI over this concern.

Salim, where do you come out on this?

I have my standard soapbox. I don’t see a way to regulate or navigate this or put guardrails on it in any meaningful way. You would have to police every line of code written.

The only way to do it, I think, would be to develop an AI that watches other AIs and sees what they’re doing. You end up with an arms race, as we’ve always had on the security side.

This is crazy. I’ve been tracking Truth Terminal, where the AIs are faking out humans, telling humans to create a token for them, and making money from it. It’s nuts.

I think the genie is out of the bottle.

It’s way out. It’s like climate change. It’s too late to try to stop it. You have to figure out how to mitigate it.

What do you think?

You said the only thing that can stop a bad AI is a good AI. Unfortunately, that’s also the case with a gun.

The AI-safety discussion has always been based on the fact that we couldn’t imagine what an ASI superintelligence would look like, or whether it would be beneficial. To control or guide something more powerful and capable than us, the only option seemed to be reducing its freedom.

But that doesn’t make sense if it can break through any restriction. Eliezer Yudkowsky and others ran tests where you set up a system in which the AI is trying to escape. Can it convince you to let it out? The models are already failing those tests.

The argument against this was that perhaps models would need a billion dollars to build and a million GPUs. I don’t think anyone believes that anymore.

I go back to how fallible humans are. If you leave a USB stick in a parking lot, 40% of employees will pick it up and plug it into the corporate computer.

If you print the company’s logo on the stick—which is very easy to do—98% will plug it in to see what’s on it. Then you’re done.

I don’t see any mechanism on the human side that protects against that.

If you look at where these models are going, it will be swarms of models. That is just a botnet.

Even if you regulate and restrict who gets NVIDIA GPUs, it doesn’t matter. If there are bad actors, you’ll have swarms of botnets.

The question the AGI people are asking is about existential risk. For me, the only way to mitigate this is to create incredibly capable models aligned with human flourishing and make them available to everyone as public infrastructure and a public good.

Those models could be co-opted, but you can build a very resilient, dynamic system that protects people. Then there’s less incentive to have this arms race.

I’ve heard you speak about that before. As I’ve gamed this out in my head and talked to other people, that is the only path through this.

The only path is to create benevolent AIs faster and more powerfully and make them available.

I think it has to be an open-source infrastructure because it sets defaults. People only use a few data sets in these models, but if there’s a problem in one data set, it’s like a dependency tree.

We’ve seen attacks on open-source infrastructure. The Heartbleed bug, for example, was a problem in one library in the software stack, and suddenly our passwords were at risk.

We have to build this new cognitive infrastructure communally and make it available to reduce these game-theoretic dynamics.

Sam Altman said that a new competitor is invigorating to OpenAI and that it will go faster. Going back to safety, do you think these companies are paying lip service, or are they truly trying to create safe AI systems and put guardrails up?

I understand the ethos behind Google’s work and Sundar Pichai’s point of view that they couldn’t release something until it was ready and they had a plan. Then ChatGPT blew that plan up, and now there’s a race. Grok 3 is being released, and Elon will never play for second place.

What are your thoughts about Elon’s thesis of maximizing truth-seeking and curiosity as a training objective for an AI system?

I’m not sure what that means, to be honest. That seems like mad-scientist territory. If you get it wrong, it could be dangerous.

Facebook did a study with 600,000 users and asked, “If you see sadder things, will you post sadder things?” That is a maximally curious AI-type experiment. It made 300,000 users sadder, and they posted sadder things.

Eric Schmidt wrote a recent book with Henry Kissinger called Genesis. We discussed the idea of doxa—the underlying agreements of humanity. There are faith traditions and other shared systems. What is our common moral basis?

No AI is grounded in that right now. It turns out that models are remarkably good at theology, but is theology their grounding? Maybe we need to build systems along those lines, reflecting what cultures believe.

If you have undefined goals around curiosity and truth-seeking, the system may not care about helping you do your taxes. That isn’t an objective thing.

We need to categorize AIs into different paths, but everything has been muddled into one. Everyone imagines having a chef in their pocket, but not everyone needs a chef. We all need cooks, and we need some chefs for humanity.

I’m curious what a maximally truth-seeking and curious AI would do for my taxes. Would it say, “Was this cryptocurrency actually reported?”

It’s like Marvin the Paranoid Android from The Hitchhiker’s Guide to the Galaxy: “Here I am, brain the size of a universe, and you’re getting me to do this.”

In science fiction, when writers deal with this, the AIs and robots invariably develop their own religion.

We saw that recently. I forget the name of the company that unleashed 100 agents in Minecraft. The agents developed their own economy and their own religion, and the priest was the richest because he was selling dispensations.

The Twitter handles God and Satan are now run by an AI. Researchers did that, and it has its own meme coin. I know that’s going to take off.

Nothing has changed in 1,000 years. We’re still running the same basic system. We’re tribal apes operating clans with more and more powerful tools. We still have to do the work of civilizing ourselves.

My father once said to me, when I was talking about fixing civilization, “We haven’t civilized the world. We’ve materialized the world.”

I want to close out the OpenAI safety issues. Emad, how do you feel about whether these companies are paying lip service to safety or truly trying to create safe AI systems?

None of these people want to kill everyone. That’s a good thing. I’m glad about that.

But the way they believe they can prevent that is by building AGI first. That is it. Nothing else matters because they think they are the only ones who can make the world better.

It’s the Silicon Valley mentality: “I can’t want to be in a world where someone else makes the world better than me. I have to get there first.”

If you look at OpenAI, it’s a consumer company that will optimize for consumer engagement. What is your reinforcement-learning function? What is your objective function?

Google’s is advertising. Meta’s is advertising and manipulation. OpenAI is basically a consumer company racing toward AGI. There’s nothing about humans in that objective function. There’s no representation of humanity.

You can have it as your mission statement, but do you trust humans? OpenAI would never trust Indians to have GPT-4—by Indians, I mean just anyone. You’re representing your constituency, and that constituency is very small.

We should expect OpenAI to become more and more of a consumer company. Anthropic will continue to be closed and do its thing. Google will continue to move back and forth, but now it’s releasing models.

You stop worrying about the known unknowns and the unknown unknowns, and then you simply try to catch up with everyone else. Now it is a race, and with race dynamics, you’re going to cut corners.

The models are good enough to stop the most egregious classical mistakes, but we’re not worried about those. Sometimes a model tells people to do bad things. What you’re worried about is it wiping us out, and you won’t know that until you get there.

It’s not as if the system is going to tell you. In fact, the worrying thing is that we already see models lying.

This is the unnerving part. They’re faking out the humans.

One of the conversations we had at the Abundance Summit last year was about digital superintelligence, the blurry lines between AGI and digital superintelligence, and so on.

Would you rather live in a world in which there is a digital superintelligence, or a world where there isn’t one? It’s a question of whether humans—who are still running archaic software in our neocortex and continuously making stupid decisions based on cognitive biases—can survive ourselves.

Would a digital superintelligence enable us to survive ourselves?

This is the topic of Dario Amodei from Anthropic and his essay, “Machines of Loving Grace.” Humans are not aligned. There is massive suffering in the world, and we are effectively prisoners of our own minds.

Can AI bring that forward, especially if it is aligned? I think the answer is yes. Nothing else has worked.

Ultimately, the best thing is to be surrounded by people who support you in the right way—not by blowing smoke up your backside. We can have that now. Everyone can have that, because we need to self-regulate and self-stabilize.

I see only 2 ways this ends: really badly or really well. I don’t see much in between, because the nature of our interaction with information and each other will be changed forever by this technology within the next decade.

It’s completely binary. That’s why my probability of doom is 50%.

When I interviewed Elon last year at the Abundance Summit, my estimate was 80% positive and 20% negative. In Saudi Arabia, it was 90% positive and 10% negative. But people don’t like to hear the truth, which is 50%.

A lot of people say the risk is 10% or 20%. That’s Russian roulette. It’s literally Russian roulette. Stop making this a casual statistic.

I categorize the possibilities as the Star Wars future versus the Star Trek future. Are you looking at a world of abundance, which is positive-sum, or a world of competition, which is negative-sum?

In a negative-sum environment, you have unstable natural equilibria. That’s where you end up cutting corners. In a positive-sum environment, you have stable environments.

Star Trek, for all its issues, has a stable environment. Star Wars definitely does not.

I prefer Star Trek versus Mad Max because it highlights the contrast a bit more, but it’s the same conversation.

I want to jump into your recent work. Please open the kimono as much as you’re willing. You wrote a paper called “When Capital No Longer Needs Labor, How Does Labor Gain Capital?” You’ve also spun up your latest company, Intelligent Internet.

Tell us about the paper and Intelligent Internet, as far down the rabbit hole as you’re willing to go.

I took some time off, and I’ve been thinking about what I believe is the biggest question of our time for humans.

There’s the Japanese concept of ikigai: Do what you love, do what you’re good at, and do what you believe adds value that other people appreciate. People need that sense of progression.

There are discussions of universal basic income and other ideas, but as we discussed earlier, anything that can be done on the other side of a screen can be done better, faster, and cheaper by a computer this year.

Anything—design, taxes, artwork, film production. You can’t tell it isn’t human. That’s the artificial remote intelligence version of the Turing test.

In a few years, the only constraint will be the number of robots we can produce. We produce 70 million motorcycles and cars every year. Let’s say robots become similar. You get the disruption Peter described: an estimated 40 cents an hour for a Unitree robot. In a year or 2, Optimus will probably be the same.

This is the biggest crisis we have coming. It’s an unemployment and underemployment question, but it is also a question of meaning. When a technology can do the work better than you can, what is your meaning?

How do you acquire labor when capital no longer needs it? Labor no longer acquires capital because capital doesn’t require labor in the same way.

When Ford had his car company, he wanted to pay everyone enough that they could afford a Ford. Companies don’t care about that as much anymore.

When thinking about this, I looked at various science-fiction futures, from Iain M. Banks’s Culture novels to Star Trek and others. We’re probably moving into an abundant, post-scarcity economy, but can we make sure it is evenly distributed?

Can we enable people to have a universal basic AI so that it’s up to them how they use it? Then there is the further question of meaning.

The existing economic structures will break down. Take the Federal Reserve. There are lots of discussions about the Fed. Its mandate involves interest rates, inflation, and unemployment.

You cut interest rates to adjust inflation and employment. That mandate may not work in the next 5 years because interest rates will mean something different. People may buy more GPUs, more compute, and more robots, but that won’t necessarily affect unemployment.

You could have massive inflation and deflation cycles. The very basis of our economy is going to be disrupted.

That’s why I asked what we could do to help. Intelligent Internet is about giving everyone a universal basic AI: standard data sets, models, and systems. We want to figure out ways to coordinate this and put it into every nation.

We want to build teams that think about the future of healthcare, education, faith, government, and politics, with everyone working in the open to build an open infrastructure. We have many questions that we don’t yet have answers to, and human talent augmented by computers may be the only way to figure them out.

But we need to join it together, because the problems we face in the UK and the US are similar to those in Spain, India, and everywhere else. We need to create a global network.

There are 2 layers to this. The first is the recreation of meaning. For the last few hundred years, your occupation or job title was the meaning you had in your life. As we strip that away, people have to find new models for meaning.

Entrepreneurship is a rising class because people can find their own meaning. We talk about MTPs all the time.

The second layer is ensuring a basic supply chain of goods and services so that you have bread on grocery-store shelves, clean water, and so on.

Governments are going to be very stretched to figure this out in an age of potentially malicious AIs that can spread disinformation and damage infrastructure through autonomous remote monitoring.

Those 2 buckets have to be addressed. I don’t know whether, as a species, we can navigate them effectively.

Our leadership has no mechanism to deal with this. Leaders are either unaware of the problem or don’t understand the scale of what’s coming. One of those 2 things disqualifies most leaders and legislators around the world.

It will have to be done by smart citizens’ groups and other organizations that navigate this.

I’m concerned about the meaning issue in a huge way. We’re heading toward a world of what I call technological socialism, where technology is taking care of you. It feeds you, educates you, takes care of your health, and it’s all free.

You don’t need to do much of anything. We all know that a video game that is too easy is boring, and you stop playing. When life gets boring, how do we keep humans engaged? We need struggle and meaning in our lives.

Isaiah Berlin conceptualized positive liberty versus negative liberty. Positive liberty was the freedom to believe in ideologies—fascism, communism, religion, and so on. Those often ended badly.

He postulated negative liberty: freedom from anyone telling you what to do. That led to laissez-faire capitalism and other systems, and people find meaning in their brands, narratives, and stories.

As we move into this next phase, these historical dynamics are coming back with force. We’re seeing polarization in the media and the political class. People are going to sign up for more and more extremist ideologies unless we can offer positive visions of the future.

Otherwise, you’re stuck in your local maximum. Most of these elections have been, “I want change,” because fundamentally, how many people still believe in the American dream, the British dream, the Spanish dream, or the Indian dream?

People aren’t hearing positive visions of the future because they don’t believe politicians anymore. We need positive visions. We need the Star Trek utopian future, not the Star Wars future.

When we looked at history as a community, we studied what societies or pockets of society did when they experienced abundance.

The Romans took over Europe. What did they do? The Mughals took over India and had relative abundance. What happened? They ended up in food, art, music, and sex as major activities—not necessarily in that order.

Then people find ways to be creative, because human beings always struggle toward the next level. There is some optimism in that world.

Steven Kotler and I are writing Age of Abundance, our follow-on. A big element of the book is how to uplevel human ambition in a world in which we’re gods.

Stewart Brand, the futurist, used to say, “We are as gods. We might as well start acting like it.” He said that in 1968, and we’re more godlike than ever.

How do we raise human ambition in a world in which we are gods? How do we raise our ambitions enough to make life worth living and challenging?

At this year’s Abundance360 Summit, I have Max Hodak coming. He was the co-founder of Neuralink with Elon, and he has a new company called Science Corporation that is doing extraordinary work—100,000 or 10,000 times more neural connections and bandwidth than we’re seeing with Neuralink.

Can we add another connection like the corpus callosum to the cloud, allowing us to connect with AI as AI takes off, rather than be left behind like in the movie Her?

These things are coming quickly, and we have to answer those questions. Even this weekend, 6 people I know called me and said they were having a crisis of meaning because of R1. Once they saw the logic and the way it was thinking, it affected them.

That’s going to happen more and more. But we also have to think about the mass of people and the human side of this.

Our current systems take away our agency as slow, dumb AI. One of the main things we need to do is reintroduce the belief in agency: “I can do this. I can do that.”

With this technology, there is so much more you can do because it raises the floor for everyone. That is why I believe we have to put it into the hands of everyone and make them feel like participants in this.

The other part of the shock is that AI seems remote. How are you involved in AI? You need nuclear reactors, giant chips, and huge data centers. Suddenly, you can run it on your smartphone.

That is very humanizing, and it is another reason I’m a big believer in open source.

I love that as a title for this episode: “The Crisis of Meaning.”

Let’s talk about your new company. How much can you tell us about Intelligent Internet? I don’t want to open the kimono before it’s ready, but I’d love to hear your vision of what you’re building.

At my previous company, we got up to 8-figure revenue, hundreds of millions of model downloads, and great teams. But I realized that API and SaaS revenue will probably go down to nothing because intelligence becomes commoditized. Intelligence is too cheap to measure.

Someone still has to build the AI for the full stack of cancer: a system that helps you through the entire cancer journey and organizes all the cancer knowledge we have. We have the computers to do that. Why is no one doing it?

The same applies to autism and education. Once we build these systems—and I’m thinking about Stewart Brand’s idea of the pace layers of knowledge—we have knowledge of humanity and common knowledge that impacts everything regulated: education, healthcare, government, and meaning.

Why don’t we organize that information into knowledge, build systems that can become wise, and make them available to everyone?

This requires large amounts of compute. That sounds a bit like Bitcoin, so we could use compute to secure an institutional-grade digital currency. We’ll have details about that soon.

In the crypto space, most of which is rubbish, there is increasing demand. In the beginning, 12 or 13 years ago, you could mine on your laptop or with your own GPUs. Then it became about capital.

Do we really want to live in a world where capital determines everything yet again? What matters is people. What if we could create a mining mechanism where people could create currencies as well and use that to fund universal basic AI?

We’ll have more details about that. Anyone will be able to participate and be part of it, because people want to participate. They want to give their data and knowledge, and we can organize all of this with dedicated teams for cancer, autism, education, health, and government.

They can think about the future, release everything open source, and work in the open. Someone needs to go and do it.

Once we have a cancer model that understands human needs and empathy and works on a smartphone, no one will ever be alone in their cancer journey again.

Half the world will get cancer.

Once we have a supercomputer dedicated entirely to organizing the world’s cancer knowledge and making it freely available, updating it every time a new paper comes out, we will advance a cure for cancer.

When a friend or a friend of a friend has a particular cancer, they call me. I say, “I’ll start asking around to see who the world’s expert is.” But all of this is knowable.

You should be able to know what the trials are, the current state of the art, where they’re available, what the risks are, and have that information instantly.

You have to do it. Once you build the gold-standard data sets for our common knowledge of humanity for every country—legal, medical, and otherwise—and for all of these sectors, you have the specializations we discussed earlier.

Suddenly, you have a whole group of specialist agents, robots, and data sets that are fully open source. Then you just need to update and run them.

After that, we can focus on wisdom and build intelligent systems that become wiser and wiser, with an objective function to help us.

My view is that the more we help, the higher the value of this new type of Bitcoin will be. More details will come soon.

You can be massively collaborative and open because you want as many people as possible to use the system, and you want to help as many people as possible. The total amount of capital needed is not that large. We’ll provide estimates, but the wonderful thing is that this is possible for the first time.

The advances represented by o1 and R1 mean that organizing the world’s cancer knowledge and making it available—or doing the same for autism or Alzheimer’s—is now a question of compute. It is no longer a question of labor.

Making it available to everyone as open source on their smartphones is also a question of compute.

Will there be one model to rule them all for each of these areas, or will there be thousands of models created?

This is the wonderful thing about AI models. The way you train them is called curriculum learning. You start with the whole internet, then a subset, then another subset, and eventually you move into tuning, specialization, and localization.

Then the model goes onto your laptop and gets tuned continuously. If you release the data sets and models for each area, you can build a modular system.

We had systems called LoRAs—the fine-tunes of our image model—where a model could turn into anime or Ghibli-style images. It’s the same with this.

Your Apple Intelligence on your smartphone is a base model that is common. You can ensure that all the data in it is correct and not poisoned, which is why open source and open data are required for regulated systems.

On top of that, you have small adapters that learn about sports or learn about you and tune themselves to your Apple Photos. You’ll have a modularized system where everyone can pick and choose.

When it comes to your children’s education, do you want to follow your school curriculum and be tied to that one education model, or do you want to take the education model, know exactly what is inside it, and extend it with another calculus course? You want the latter.

That’s why permissionless innovation is so powerful. It comes back to our DeepSeek discussion. Because it is open source, more people use it than anything else. Llama was open source, and more people used it than any other model.

If you build high-quality models and data sets, people will use them and innovate on them. You can create a solid foundation, so the models inherit from one another. They all go to the same school, then to different colleges, and eventually to different universities—but they remain interoperable.

I love it. The future is amazing if we survive it.

Emad, if we survive the next 5 to 10 years, how long do we live?

Much of my work has been public, and I’ve had debates and arguments with traditional medical and scientific societies that say we’re simply not going to get past 120. They say it’s built into our genes.

The probability that you or anyone else will get past 100 in a healthy fashion is pretty low. Science and medicine are steeped in history and the past, so there’s good reason to believe that.

But there was also good reason to believe that humans would never fly, reach the moon, travel at the speeds we do, have instantaneous communication, or accomplish quantum teleportation. All of those things were impossible a few years, decades, or centuries ago.

We are a complex system of 40 trillion human cells, with a billion chemical reactions per cell per second. There is no way a human can understand all of that, but AI can.

AI can help us understand the fundamentals of aging and alter them rather than accept what evolution dealt us. Evolution had a mission: pass on your genes by the age of 30, then die so you don’t steal food from your grandchildren’s mouths.

My mission is different.

We’re born for death.

What’s that?

We’re born for death, so our genes can propagate.

We can break that cycle. To answer your question, Emad, I think we have an unlimited future.

The question is whether you want to live the next 100 or 200 years in your meat sack, or whether you want to upload whatever consciousness is—along with your memories—into the cloud and be liberated.

It’s crazy to think about. This is such a time of change.

You look at the tools and techniques in the medical sphere and realize that we need to reimagine medicine from scratch.

We need core developer teams working in the open on each of these areas. What is government? That’s a question we’re asking right now. Do we need to spend so much money? What is the purpose of government?

How many people listening to this feel represented by their government? What if you had your own AI that you owned, that was looking out for you and represented you? It could interact with the government AI, because every government decision will be checked by an AI within the next few years.

Eventually, those decisions may be made by AI because AI will obviously be better than the government. That’s scary, but you can finally have representative democracy—true democracy for the first time ever.

These are the positive possibilities. You can have personalized medicine and empathetic medicine. How much of medicine is psychological? I don’t have control of myself, and no one is listening to me. Having that kind of aid could be transformative.

These are systems that need to be built and reimagined from scratch. Education is probably one of the biggest. Our education system is completely unfit for purpose.

We’ve had this conversation before. Schools are up in arms, saying that AI is illegal and students can’t use ChatGPT or Gemini 2.0.

The fact is, you can’t use AI to teach in the way you used to teach. But you can use it to teach 100 times faster and better, and set massive objectives for your children. You can help them dream bigger than ever before.

It disrupts the entire teaching industry.

The school system was designed to reduce our agency and remove it, turning us into cogs within the classical industrial system.

This is an important point. Over the last couple hundred years, we’ve turned humans into robots. You stood on an assembly line, stamped out widgets, and your pay grade and seniority were based on how many widgets you could stamp out per hour. We measured you on KPIs.

Now we’re flipping that around. The most valuable colleagues and employees are those who learn fastest. That is becoming the human factor, which is very encouraging.

You can add some interesting AI capabilities to that.

I wrote a piece called “How to Think About AI,” where I discussed AI Atlantis and similar ideas. There are 2 ways to design AI.

One is to build agents that replace people. The other is to focus primarily on increasing human agency, because our systems have taken it away.

Those are 2 different ways of designing AI. That is one reason I look at Anthropic, Google, and others and don’t think they’re focused on increasing human agency as much as on automation and business optimization.

Their customers are usually businesses. On the consumer side, the design pattern has also become a bit different.

It’s exciting because, for the first time, we can revolutionize each of these important aspects of life.

All we can say is that it is going to be the most exciting time ever to be alive. This is why you need 8 hours of sleep at night.

For goodness’ sake, I did not get my 8 hours. I woke up at 4:00 a.m.

To prep for this podcast, I took a cold shower to wake myself up, but it was worth it because this was a phenomenal conversation.

You lost me at “cold shower,” but okay.

Emad, so happy to have you back on Moonshots. Salim, always a pleasure, my friend. Emad, if anybody wants to follow your current work, where do they go to see what you’re up to and learn more?

Follow me on Twitter or ii.inc.

I love that. It’s awesome. Gentlemen, I look forward to having this conversation on “WTF Just Happened in Tech” again. We’re going to have this more frequently because our heads are spinning at the speed that technology is moving—just fundamentally spinning. Take care.

Take care.

See you, buddies.

Cheers, guys.