Last February, I said Deep Seek is one of my favorite AI companies out there. If you look at each of the innovations they made, it was largely engineering innovations. Do we see Deep Seek dethroning or reducing the valuation of these companies at all? I think, in my opinion, it should increase the valuation. The US versus China AI wars, you know, this is a winner-take-all type of game. This is the biggest crisis that we have coming because we're heading into a future now where I'd say every single AI leader says AGI is 3 to 5 years away.

Welcome to Moonshots in an episode of WTF Just Happened in Tech with Saleem Ismael and a special guest, Imod Mustak. You know, Imod is the founder of Stability AI, a company that had been the leading open-source developer in music and image generation with 300 million open-source downloads. Imod today is the founder of Intelligent Internet. He'll be speaking about that, but in this episode, we're doing a deep dive into three subjects: Deep Seek, of course, and the constant disruption that's coming in every market at an accelerating pace.

We'll be diving into AI safety, what's going on at OpenAI as people are starting to leave, especially from their AI alignment team, and then we'll chat with Imod about Intelligent Internet, what's his plans, where is he going.

All right, let's dive into this episode. For me, this is an extraordinary week of accelerating change, and as always, help me spread the message. Subscribe, tell your friends. This is the conversation that I think is probably one of the most important that we can be having right here, right now.

Let's jump into Moonshots. Welcome to another episode of Moonshots, WTF Just Happened in Tech. This week, I'm here with two besties: Saleem Ismael, the CEO of OpenEXO, and Imod Mustak, the CEO of Intelligent Internet, no stranger to this podcast. It's been a crazy week. We kicked it off with sort of an internet AI market meltdown on the news of Deep Seek, and the concussion waves keep coming. What does it all mean? We're here to have that conversation.

Imod, good morning to you, or good evening, you're in London today.

Yeah, I'm in London. Morning to you, it was a pleasure.

And Saleem, you're in Miami or New York?

Yeah.

All right, we've got three different time zones around the globe. We need someone in Hong Kong to just balance this thing out, but we'll get there soon enough.

So, Imod, Deep Seek, no surprise for you. Was this something expected or was this something like, wow?

I think it was actually expected. Last February, I said Deep Seek is one of my favorite AI companies out there. They took the original ethos that we had at Stability, another ex-hedge fund manager, and they released amazing models open. I think when the AI community first started to see this was probably around about summer of last year. They released Deep Seek Koda, which hit the top of the code rankings.

They started with replicating Llama from Meta, then they broke forward. In fact, the algorithms there are some of the algorithms they use now. Then in December, God, like a month ago, they released Deep Seek V3, which was actually what this 6 million training cost model was, and it matched GPT-4 and all these other models. It didn't match 01 at that point, but we all thought they'll figure out how to do it, and guess what? They did, and it generalizes to what was the use.

So 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 it's been around for some time?

So there was the base model that was the ChatGPT equivalent model in December, and that proved that you could train these models on 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, has better quality output. That actually came out last Monday, but it was this weekend that this narrative cascade happened.

Now you've got your mom and your aunt asking about it, you know, and it's front news, and then Nvidia cracked, etc. I think what it was is remember the early days of ChatGPT or Stable Diffusion on image? The immediacy of response and that new paradigm. When OpenAI released 01, this thinking model, it was amazing, but it was a bit like using ChatGPT. You put something in, it says, "I'm thinking," and it gives you a response.

Because what they did is they 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 use that and they saw the performance benchmarks, it built up into this cascade because it was so immediately usable, and they realized it was open source.

So people took the smaller versions of it and started running it on their laptops. If it was just a closed model that didn't have the chain of thought but matched 01, it wouldn't have had that. If OpenAI had released the chain of thoughts, then I don't think it would have had the same thing.

It was this confluence of things that made people realize, "Oh my gosh, what is this new thing, and how has it been done?" It challenges our assumptions.

Amazing. Saleem, you and I were on the phone over the weekend like, "Huh, this is real, this is happening." What were your thoughts when this started?

I have two 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, "We will sanction you to bits, etc., etc., 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, and as the power of these models is accelerating exponentially and blowing our minds, the demonetization should also kind of surprise us in the same way.

So the fact that they're able to do this at 1/10th or 1/100th or whatever, now how they got there is obviously an open question, but the fact that that has been achieved shouldn't be a big surprise on the curves that we're looking at, right?

It's incredible to see, but we shouldn't be surprised. If we had owned dog food, go ahead.

I think someone noted it was actually the five-year anniversary of the Wuhan lab leak as well, except for this one was delivered.

No, not going to go there. But, you know, like I put out in my blog that followed the Deep Seek announcement, this is just going to be what is the new normal. You know, when Netflix ate Blockbuster for lunch, this is just going to be happening over and over again.

The speed at which heads are turning and snapping across every industry, you know, it's interesting because when we saw ChatGPT announce and it got to a million users in five days and 100 million users in two months, people were like, "Can this ever be replicated again?" And the answer is yes, and faster.

So, Imod, could you give us a quick rundown of how Deep Seek actually compares to GPT-4, GPT-01, any of the other models? And, you know, there's a lot of claims being made about how many GPUs it was created on, how much money, size of teams, and it was those comparative numbers that made it a big deal. If it was just an equivalent model but was not being done at a fraction of the time or cost, it would not have hit as hard as it did.

Yeah, I think the shock was the order of magnitude. So we can break it down a bit. So 01 was this evolution of ChatGPT that came out that suddenly got to IMO medalist level or top coder level, like top 1% coder level, because it could think longer. This is a key breakthrough.

Now OpenAI has actually said, Mark Chen from there, what Deep Seek figured out, which we'll get to in a second, was pretty much what they're doing at OpenAI. That was in November, and so we've had a few periods there.

So first of all, you had the model that matched ChatGPT, then they figured out how to make it think longer. But the main upshot that shocked people, I think initially, was that it was 96% cheaper. Now, software usually has an 80% margin. We don't know how much OpenAI charges, but you know they've got this hammer, which is a large amount of GPUs. They've never had to work in a constrained environment, so sometimes you are a bit price insensitive, particularly because the cost of running an 01 query to solve a math paper or a legal problem, because it's as good as any lawyer or doctor, is so small still.

But this was 96% cheaper than that, which was number one. Number two was the fact that this could be kind of released anywhere. The headline number of the original model that was trained from the R1 evolution is probably only $200,000 from that, which again we can come back to was a shock last year.

Well, a year before last, I can't remember the exact number. I think last year OpenAI spent $3 billion on training models. Amazing to give you an idea of that. Now, how much did Deep Seek cost? There were accusations around they have 50,000 of these chips, not 2,000 like we used on the training run. They never claimed how many chips they had. They just said, "We need 2,000 for this training run, and we used it over this period of day to build a model that looks like this."

Those of us that have built these models know that these numbers actually all check out, and this is why some of the reaction has been really interesting because people are like, "Well, they have far more GPUs, or they have hidden GPUs and other things." The GPUs they have are these models called an H800, which is like the top, well, now not quite the top-end Nvidia chip, but with the interconnect slightly reduced.

So the way that the chips speak to each other is a bit slower. We had this issue at Stability AI, a former company, where we built one of the largest supercomputers in the world, but we had interconnect a quarter of the speed of other people because that's all we could do.

And again, we were competing against the biggest guys. We built some of the best models in the world. They wrote the lowest level code in PTX, which is like this CUDA but a level lower to overcome it. They basically engineered the crap out of it because some of them are ex-Quant hedge fund managers and others.

If you look at each of the innovations they made, it was largely engineering innovations, which is very interesting for our mental model because what's China amazing at? Engineering innovation. You look at BYD, you look at Xiaomi. It shouldn't be any surprise that as you move from research to engineering, you would see this leap ahead.

But all the numbers kind of check out. You see the cost reducing. I think they've probably got 10,000 chips in total, but that's not more than many startups in the Valley, to be honest.

You know, everybody, Peter, if you're enjoying this episode, please help me get the message of abundance out to the world. We're truly living during the most extraordinary time ever in human history, and I want to get this mindset out to everyone. Please subscribe and follow wherever you get your podcasts and turn on notifications so we can let you know when the next episode is being dropped.

All right, back to our episode. You know, I had a conversation with Kaiu Lee recently on this podcast, and we were talking about the notion of how the US government has been restricting Chinese companies from getting Nvidia chips, and all that's done is create this evolutionary pressure for them to do much more with much less. This sounds like a perfect example of that.

It's like it's Darwinian in its developmental force.

Yeah, I mean, again, if all you have is a hammer and you have large amounts of GPUs, the way that this works is the GPUs compress the knowledge. It's like pressure cooking a steak and making it tender. Instead, you look at things like better data, better algorithms, more efficient things.

If you can't scale on compute speed because they didn't have the chips for the speed, because what happens is as you go from 1,000 to 2,000 to 10,000 GPUs, you can parallelize it and have more speed. They instead did memory as the key thing.

So classical models are very dense models, like Llama, 70 billion parameters. This is 64, 40 billion parameters, but only 30 billion of them are activated at one time. They scaled on memory, and that is cheaper than super-fast silicon.

So these constraints, I think, really are the key, and we've seen it again and again that if you don't need to worry about the constraints, then you build inefficient models. If you have to worry about efficiency, then you know necessity is the mother of invention.

It wasn't the CEO labeling the data and going through all that stuff because that adds so much juice to the model. Models are just data. I mean, again, if models are figuring out the interconnections, it's like if you have a bad curriculum, then you have bad data.

The models we train on right now are trained on terrible data, like 14 trillion words in the case of Deep Seek 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, or even a 2,000.

So what we're seeing now is data improvement. In fact, the data they used to turn this from a base model to a thinking model, which they then transformed the Llama model with, was all synthetic data.

So we moved to a point now where they figured out what the right type of data was, and you find typically with those that make breakthroughs, they don't send the data off to the Philippines and do all of this and try to make up for it with engineering scale.

You look at every part of that process, and again, this echoes what we've seen in engineering. How did the engineering marvels happen at Tesla or Chinese companies? They look at every part of that process and they simplify, simplify, simplify.

So we had David Sachs over the weekend with this commentary. Let me go ahead and play this video one second, and love your both of your thoughts on it.

Well, it's possible. There's a technique in AI called distillation, which you're going to hear a lot about. It's when one model learns from another model. Effectively, what happens is that the student model asks the parent model a lot of questions, just like a human would learn.

But AIs can do this, asking millions of questions, and they can essentially mimic the reasoning process that they learn from the parent model, and they can kind of suck the knowledge out of the parent model.

There's substantial evidence that what Deep Seek did here is they distilled the knowledge out of OpenAI's models, and I don't think OpenAI is very happy about this.

What do you think about that, Imod?

Well, it's a bit calling the kettle black, right? In some way, don't train in our data. I mean, distillation is nothing new, and there's no way to kind of stop this from the model basis.

But if you actually look at what the paper says and what's reasonable, they had this version R10 that created its own data. What's this familiar with? It's familiar with AlphaGo and MuZero, these reinforcement learning models that outperformed humans on Go.

In fact, you could feel like maybe we're all Ladol, right? Like the AI is coming for all of our expertise. It's inevitable that that will happen, but I don't think they deliberately went in and did that because OpenAI's 01 outputs, these cutting-edge outputs, were missing the chain of thought reasoning step.

We've seen now that as you take the chain of thought reasoning from R1 and actually the new Gemini flash thinking, the Google model that's now top of the leaderboard, that's what you really need if you want to optimize this process.

So I think they actually created their own synthetic data, but as they look at all of the internet, there will be some OpenAI data in there. We've even seen that with Llama and Gemini and others. Sometimes you ask it, "Who made you?" OpenAI, because it's taken so many of those drinks.

You know, we've got an interesting impact on Wall Street that occurs on Monday morning, where it's, you know, red across the board. Nvidia got hit massively. I'm sure you know OpenAI was reeling.

Saleem, how do you think about this? Because this is what people respond to.

Yeah, you know, I think 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 is, and I'd love to get Imod's take on this, is the overall demand in AI is so exploding that it's not going to really make a big dent in the demand for the chips.

Yeah, I mean, like Nvidia is still up 100% over the last year, right? It's like it's not down a lot. No one knows what's coming, but what's the market size of this? The displacement is the displacement of all knowledge labor, just like the industrial edge. You replace muscles; now you're replacing brain cells. That's a huge market.

Oh, I mean, we have a global GDP, you know, going into 2025 of $111 trillion. You know, half of it is physical labor, and half of it is, you know, effectively intellectual labor. It is massive, and this is the technology, the intelligent capital stock that really will define productivity.

So it's very difficult to get a handle on how this will go. People have been talking, like Satya and Adella, about Jevons Paradox, you know, like, "Oh, the price, the higher demand." And Ro Andreessen's been talking about this.

I feel it is that, and if you look at Nvidia's strategy, they've been moving to these fully integrated data center boxes, the GB300, MVL72s, and this new thing, Digits, which if you've seen a Mac Mini, it's like a Mac Mini that sits on your desktop, has 128 GB of VRAM, a petaflop of AI compute for $3,000.

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

So you made a comment earlier on in terms of the amount of energy and cost you think it would cost to build Deep Seek's model. Could you speak to that?

Sure, kind of insane. So when we bought our first major supercomputer, this would have been about the 10th fastest in the world publicly in 2022 at Stability. It was four A100s, which were the top-of-the-range chips. The internet connector was a bit poor, but you know, it was still big, and each of those chips used like 400 watts of electricity. That was a big old beast.

If you can recall the recent Nvidia announcement, Jensen had this like shield, which was their new integrated box, these NVL72s with 72 chips super interconnected. In fact, the interconnect on those chips is equivalent to the bandwidth of the whole internet. That's how much faster they've got.

One of those boxes pulls down 100... wait, can you repeat that?

The computer on those chips, the way that they communicate with each other, the total bandwidth is like six petabits a second, which is the bandwidth of the whole internet. They figured out how to get everything integrated, so you don't have this chip-to-chip interconnect. You just have this big wafer with 72 chips on it.

It uses 100 kilowatts of electricity, and when I was doing the math on this, I was like, so you have 2,000 of these H800 slightly hobbled chips that the Chinese have, right? And Deep Seek is using, I think it would require 10 of these boxes at most, probably even less, to create that model.

Each box costs $3 million, these new data center boxes. In fact, I think it probably only costs four of these boxes. Even if you pick the upbound, the total energy required to train a model is 1,000 megawatt hours, and it's like $15 or something a megawatt hour in the US.

Now, $20,000, you could literally train it off of a small solar farm in your backyard. Well, a big solar farm, you know, it pulls down a decent amount, like 100,000 kilowatt hours of energy. But then that box to run it, you could definitely run Deep Seek R1 on solar power panels.

And if we look at the direction this is going, because it's still not optimized, next year you should be able to get an 01 level model on your smartphone that pulls at most 20 watts of electricity, and it's less than a dollar per watt of solar power.

And this doesn't make sense if you look at what these models are capable of. And we think about the cost of intellectual labor. Well, it makes sense when you think about how much energy your brain pulls, just 20 watts.

And so we've got a huge efficiency curve to ride to get there. And I think the thing is, like, by next year, you will have these 01 level models on 20 watts, which is our human brain level, and these are PhD level in so many areas.

And that doesn't compute because we've had these discussions of Microsoft bringing back Three Mile Island as a nuclear power reactor. You know, Dyson is energy is going to use everything like 60 gigawatts of electricity is coming on for data centers in the US, I think, 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, it's a few pennies. Before that, it would take entire teams using how many watts of energy in their brain, in their infrastructure, and we're not ready for that.

Saleem, you asked a question about how challenging is Deep Seek actually to OpenAI, Meta, Nvidia. What are you thinking there?

I've got two questions here. One is, does the fact that it's Chinese and companies will be reticent to put their information into it make a big difference? So that's a question for you.

And now my guess is 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. Right? Just like you give your code to, you give all your stuff to TikTok. No one knows what happens with all this data, and the versions that 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.

So, but let's stick on that question about, you know, so the question I was asked by everybody on X and my friends was, is this going to go the same path as TikTok, where in fact Deep Seek will be... well, let me back up a second.

When OpenAI first came out with ChatGPT, you had all of these companies, and Imod, you and I had this conversation, all these companies, a lot of the banks saying, "You cannot use ChatGPT in the office. We don't want OpenAI to own our data." There was this immediate privacy desire, which is still valid.

But are we going to see the same thing with Deep Seek, where people are like, "No, can't use Deep Seek. We're worried about the data and where it's going to be resident?"

I think you've seen a couple of announcements. So Perplexity announced they're using Deep Seek locally, fully on American farms, etc. You know, so farms, and you'll see that type of thing even if they're running the larger ones.

But again, it's difficult to run yourself, but there'll be APIs. Number two is you've seen OpenAI announce ChatGPT for government used by 19,000 federal employees, and this is the direction things are going whereby I think you'll have four different types of AI: super exert AGI that you call upon when needed, your personal AI, your Google, your Apple AI, these open weight models like Deep Seek and Llama, which are useful but not in regulated industries, and then open source, open data AI where these decision support systems, you need to know what's inside them and how they are actually used.

You can poison these models with inherent biases. There was this Anthropic paper we discussed before, Peter, called "Sleeper Agents," where with a few thousand words out of 10 trillion, with just one word, you can turn the model evil or change its behavior completely.

Amazing. It's like the actual... it's funny enough, you know, most of the transformers in the US are built by Chinese companies, and no one knows the control software of that. These types of threats, right? Do you want the transformers that run your business to also have that potential threat?

So that's what we're doing now in the internet, building out that open source stack for the reg. And we'll get... I want to dive into what you're building out with your newest company, Intelligent Internet, because it's got one of the boldest visions I've ever seen for supporting humanity.

The impact of Deep Seek on OpenAI, Nvidia, Meta, Google, you know, I see this comment from Sam Altman. To read it, he says, "Deep Seek's R1 is an impressive model, particularly around what they've been able to deliver for the price. We will obviously deliver much better models, and also it's legit invigorating to have a competitor."

You know, we're going to talk about AI safety in a little bit because when you're legitimately invigorated, you pull out all the stops, you pull out all the regulations, you do whatever you take to jump forward, and that's concerning.

But do we see Deep Seek dethroning or reducing the valuation of these companies at all? We saw it for a day, but is it valid?

I think, in my opinion, it should increase the valuation. It's bringing forward the time of mass intelligence too cheap to measure. If you look at OpenAI, what Sam has done masterfully is 300, 400 million users. Like, what is AI in most people's mind? It's ChatGPT, right?

Yeah, Gemini and Claude don't even register. And if the cost comes down, it's good for him. This is the Zuck school of thought. Why did they open source Llama? Because it uses 10% of their GPUs, and if there's a 10% performance gain, it pays for itself.

And so OpenAI will use whatever they didn't. Most of their models don't have brand new algorithms; they've borrowed from Google and many others, right? There's no real secrets in this space, especially now with no non-compete in California. You know, that helps.

And so for me, what is OpenAI as a company? They were in this pre-training massive compute stage. Now that's becoming commoditized. People can pre-train like OpenAI, like XAI and others, but pre-training maybe doesn't require as much. The data is getting better and better.

It becomes about intelligence refinement from seeing how people use it. It's the operator paradigm whereby OpenAI can now run your MacBook or whatever. You can let it take over, and it can book your holiday for you. That's the next stage, and I think they're well set up for that, and their costs should decrease again.

OpenAI made $3 billion of revenue last year; they lost $5 billion, of which $3 billion was training models. If you don't need to spend as much training models, it's good.

So P says the feedback loop of people using the model, and because they've got so many users, gives them a pretty good edge. I can see that. I think it's that, and then you use these... they've got like half a million GPUs coming, these B series, the VR series, and others.

You can now make those go sequential to build even better data and map and feed that back into models that you optimize and hyper-optimize. Like classically in computing, things were not parallelized; they were sequential.

So we've had this period of these big clusters. Now it's about swarms of models of agents solving tasks because they've got good enough, cheap enough, and fast enough.

Actually, that's the final thing about Deep Seek. Same with Stable Diffusion on image back in the day: good enough, fast enough, cheap enough. It's that trifecta that causes these massive adoption curves.

You know, when this was announced, you heard that Zuck created four war rooms of engineers to try and decipher what was going on and how to utilize it. I mean, it really is an AI arms race where everybody is surfing on top of each other's advances and just accelerating everything.

What I found fascinating, and I'm curious about this, is the size of their team doing it with relatively... and OpenAI had, you know, a 200-person team during its earliest days as well.

How do you think about the size of your team for the ability to create something disruptive? Too big is bloated; small and nimble?

I think a core team of about 100 researchers, beyond that, it gets bloated. So at Stability, we had 80 researchers and developers, 16 PhDs, and we achieved state-of-the-art in image, video, every modality, even multilingual.

And so we had 300 million downloads on Hugging Face, the most downloaded company, the most popular open source while I was there. Once we scaled past that to 150, things started to break down because it is about this rapid iteration.

It is about trying new things and research being an innovation center versus a cost center. You start to have too much compute and other things as well. And again, OpenAI, I think, did their best work when they were smaller, but they scaled up and still do good work.

But it is a question mark now. Like, it's become an organization, and as Saleem is the expert in, once you get past that level, it's so difficult to maintain innovation.

Yeah, you end up getting... you end up with a problem of either top-down control structures that slow down innovation, or you let everybody do whatever they want, you get a lot of duplication, and so you end up with... you have to manage that tension around it, and there's just a lot more complexity.

And you know, it's fascinating. 150 people is the Dunbar number, where anthropologically we found that this is a pretty solid, reliable threshold.

I do think, to back to Imod's earlier comment, that OpenAI has a lot more people than they really need because they have so much money they can just throw bodies at things.

Now it'll force them to be a little bit more efficient, and I also believe that this is a good thing for the overall market because a rising tide lifts all boats.

I think we're going to end up with a balkanization, though, where, you know, Western companies won't want to use Deep Seek type models. Like, I can't imagine a major Indian state enterprise wanting to use a model like that for all of the security reasons.

And then you have to develop homegrown models, and then everybody ends up with their own models in different ways, and so you end up with a splintered effect.

We'll talk about that with Imod's vision and mission on Intelligent Internet. I want to dive down into China for a moment longer because I think 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.

Because I think the majority of the world doesn't see China as sort of the hotbed of AI innovation that it is. Here's an article from Business Insider: Trump's threat on Taiwan chips tariffs could give Nvidia a fresh headache after Deep Seek.

How do you think about all of this, Imod?

Well, I think this is the real reason Nvidia would go down, or maybe Jim Cramer the previous week saying buy Nvidia, you know, one of these things.

I mean, we've seen they want to home show this. They're trying to build chips there. Intel's probably in play as an acquisition target.

Oh, it's definitely in play. I mean, it's like it's fresh meat on the table, and everybody's figuring out how to chop it up.

Well, I mean, if you look, these chips, they're getting super fast and super good with Nvidia. Like, if people talk about AMD, AMD chips are impossible to use. You know, the software isn't there, there's bugs and everything. It takes a few generations to get stable.

Nvidia chips work, but Chinese chips also work. So the Deep Seek model API was being run on Huawei Ascend 910 chips, which are a few generations behind in terms of efficiency, but they work.

You know, similarly, China has two exascale computers, two of the fastest supercomputers in the world, built in a completely different way: Ocean Light and Tian 3, because they just built at scale and bulk.

Now, what the case is here is that this particular thing is they want to increase US production because the means of production and the means of productivity of a society, which traditionally it's capital stock, it's industrial capital stock, it's IP, it will be chips.

How competitive you are on the world will be how much computing intelligence you have. I think the US has realized this, and how much energy you have to throw at it.

Yeah, that's a factor of that as well, and so the US has realized this. So it's drill, baby, drill. It's re-on sure as much of this as possible, and it's create the incentives to do that, which is basically this, like, they'll take any of that tariff money and they'll put it straight back into Stargate-type initiatives.

I think... what do you think about Stargate? Speaking of Stargate, I think the $500 billion is the total cost of ownership. That's pretty well known. It probably like $100 billion when you back everything out, which feels small these days.

It's actually a lot of money, but then when we compare that to the 5G rollout, it's less money than we've spent on 5G, and this is more important than 5G.

Compare it to... I mean, it's the order of magnitude of the Los Angeles-San Francisco railway, you know, like the mythical Los Angeles-San Francisco railway. They're like a kilometer already there.

Saleem, did you see this article this morning from Reuters? Alibaba released its AI model. It says it surpasses Deep Seek. The unusual timing of Quen 2.5 Max's release points to the pressure Chinese AI startup Deep Seek's meteoric rise in the past few weeks has placed on not just overseas rivals but also its domestic competition.

You know, this just speaks to the democratization, right? I mean, everybody will end up creating a bunch of models, and I think we'll end up with a bunch of very specialized models.

Right? I remember Eric Schmidt's comment that you'll end up with a specialized AI that's the world's best physicist and one that's the world's best biotech person, and that person can be replicated. That AI can be replicated infinitely.

And so now what do you do with deep specialty on the human side? And that, I think, is the bigger question around a lot of this stuff. The models are just going to keep getting better and better, as we've seen over time.

I mean, I think Imod's comment around what do you do with labor and seeking capital is a really, really profound question. That's, I think, the really structural, from a societal perspective, that's the question I think we should be spending a lot more time on as a global intellectual forum of how do you navigate this going forward because this changes everything.

Yeah, the models, again, it's good enough, cheap enough, fast enough, right? And in fact, the other Quen model, the VL model, outperformed Anthropic and GPT-4 on visual understanding, and the ones they have coming next are the ones that control your computer.

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

So there's a lot of conversation going on across Silicon Valley, across the White House about, and I'm speaking to Ray Dalio next week about this as well, the US versus China AI wars.

I mean, there are two levels of competition going on right now, right? It's competition between companies, and there's, you know, six, seven, eight major AI companies out there that are vying for number one position, and then competition among nations.

You've got Saudi Arabia wanting to be at the top of the stack, committing, you know, hundreds of billions of dollars, followed by, you know, Qatar and the Emirates. But you've got the US and China really going at it, and the question of, you know, this is a winner-take-all type of game.

If you develop a digital superintelligence before your corporate or national competitor does by just a little bit, it could be devastating.

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

Well, I think this is the... we're heading into a future now where I'd say every single AI leader that I could think of says AGI is 3 to 5 years away.

Though we just had Sam say it's this next year.

Yeah, but let's say within the next 3 to 5 years, every leader, we're talking about Dario, Demis, everyone, myself, like whoever, their consensus is... what do you mean? That's crazy if you think about it, right?

Like everyone says it's coming, and there's this concept of AGI, ASI, as this pivotal action moment where one entity would have the ability to shut down China. You just turn it off.

You know, so pivotal act is you build AGI first, and then it turns it off. That might happen, and we still don't know about that, which is why you now need to start preparing for it. Just like Sundar Pichai at Google said, "Why are we building out all these GPUs?" Because we can't afford not to.

Yeah, you know, and that's the game theory of that. You can't afford not to build an AGI now if everyone else is building it.

Before AGI, though, there's like an AGI we can think of as a mega chef that can come up with any recipe and outcompete all of us. What we have right now this year are amazing cooks that can follow recipes and do jobs better than humans.

Like the robots from Unitree yesterday doing the Chinese dance with the fans. I don't know if you saw that. Like, it's getting to the point where they can build houses better.

I'm going to have the Unitree robots at the Abundance Summit, and I mean, it's incredible. They're $6,000 for one of their mid-tier levels.

You know, I'm going to... that's $150 an hour.

What's that?

That's $150 an hour when you factor in depreciation, energy costs, and everything.

I have it pegged at 40 cents an hour.

I mean, it's insane. It really is. I think my kids will buy one just so I can clean up their room, and that's the most expensive it'll be, right?

Imod, when you talk about AGI in 3 to 5 years, let me get to my sub-question that I ask. What do we mean by AGI? The best kind of framing I've seen is those multiple tests, like the WNC test and the IKEA test, etc. What's your framing on what do you consider to be AGI?

I think it's probably a complex system that can outperform a team. I think before that, I had this idea of AI, artificial remote intelligence. You can't tell if it's a human or computer on the other side as your remote worker because that's the most natural way that this first starts coming in, right?

Like you call a company, and they put a bunch of people. We have the technology now that you can have a Zoom call with someone, and it could be 100% a robot.

Yeah, your worker is plugged into Slack; it joins you on Zooms. I mean, right now, there's a... we're living in a world of distributed workforce, and if you've got an AGI that is able to literally plug in, take a role fully, and have read all of the email traffic, all the Slack traffic, and be up to speed instantly, that's an exciting world.

It's an exciting world, but at the same time, that's the first level of disruption, right? Because you don't need any more BPO. Outsourcing the nature of the firm will change because they will be super chefs.

So cooks, they will not make mistakes, or they will learn from their mistakes once they have low communication overhead. The next step is teams of that, so independent, agentic. They have a task, and they can get resources towards that.

This is why Wyoming's DAO law and other things get very interesting. And the step beyond that is this ASI thing that we can't redefine, where there's a big takeoff, where it has beyond human team organizational capabilities.

Like it can invent incredibly quickly. What is going to be the impact on physics and on biology and on pure science, taking us way beyond? You know, Dario was on video, I think it was from Davos, saying, and I know you believe this, Imod, because we've had these conversations, that in the next 5 years, we'll make 100 years' worth of progress in medicine and biotech and double human lifespan.

I mean, that's pretty extraordinary commentary to be made publicly.

Yeah, and I think one of the most fascinating things of the last week is this. When you use 01 and you dump a bunch of stuff in, you can't do file uploads and other things, which is a bit annoying.

It's not that creative, but it's thorough with R1 because it hasn't been tuned and made safe in others. It's actually very creative.

So someone actually took a code base for R1 and then made it double the speed in terms of performance. Other people have put together academic papers, and it's synthesized those into new reinforcement learning algorithms.

And that's an indication of now maybe if... like the downside is maybe these things get less safe. The upside is maybe they get more creative.

And again, these are the levels. Are you an amazing cook? That's the disruption of the labor market, right? Especially anyone behind the screen. Are you an amazing chef? That takes us into this AGI, ASI kind of concept.

And again, that feels not 3 to 5 years away for me. That feels much quicker given all these exponentials, and very few people are preparing for that.

Yeah, you know, so again, the point I opened up with, which is we're going to see disruption after disruption. And I, you know, our financial markets aren't ready for this as well.

You know, we're going to see the energy market. I mean, I think one of the implications we're going to see with AGI, ASI is going to be new forms of energy sources, which will potentially topple our P dollars and destabilize government revenues.

So we have fascinating and massive adoption coming.

Well, have you ever seen that chart of GDP per capita versus energy per capita?

Yeah, it's basically a straight line.

It is, and it correlates with health as well, and lots of other things that could be completely disrupted because to make, let's say, in a couple of years, to make the best film studio in the world, you can do it anywhere with solar power.

That's what I'm kind of talking about. You could have science happening in Guatemala or anywhere like that. It's an uplift of global aggregate if this technology proliferates versus this brain drain that we've had out to the West classically.

And again, you think about your capital stock, your intellectual and physical capital stock, it's massively redistributive, and our economies are not set up for that because productivity was a function of labor, which was a function of energy. That correlation is about to break for the first time ever.

I agree. I think, you know, we're moving from an energy economy to an information economy, and now the data sets and the information you have will be paramount.

I think we need to start asking really big philosophical questions like, what do we want all this to do, and what do we want to be like, and how do we... what are the activities and functions we want to be doing as human beings as the job market disintegrates in front of us?

You know, I still have my trepidations about humanoid robots, etc., but once they show up and have feedback loops and have built-in LLMs into their circuitry, you have a fully functioning robot that can do lots of varied things.

You kind of suddenly don't need a gardener or plumber or lots of other kinds of things. I'm using those examples as a tongue-in-cheek because those are probably ones you need the most.

But there are many, many functions, aircraft maintenance, right, that will be done much better and much more precise because of the access to information.

We talked a couple of episodes ago about the fact that if there's an avatar of you or me, Peter, it's much more reliable because it's got full access to everything we've ever said rather than what we can hold in our brains.

Far more charming, far more compelling, and even better looking.

And so how do we navigate that? And I think this is where, Imod, your kind of philosophical bent towards this becomes really, really important.

And I'd love your take on where this goes. The displacement of labor is just a starting point.

You and all this.

Yeah, before we get into that, because I want to go deep in the second half of this pod today into Imod's point of view there, I want to hit on a couple of questions.

Imod, what do you think is the best-case scenario for AI this year in 2025? What are we going to see by the end of the year that people look back and say, "Okay, that was amazing, that was fantastic"?

What's your thoughts? Best case?

I think the video technology has got to the point we can remake Game of Thrones season 8, so that'll be quite good.

You know, so focusing on that, I mean, how dead is Hollywood? It's completely rewired.

Again, the energy of making a movie is massively reduced. But at the same time, at least people can maybe be more creative.

Like the video game industry went from $70 billion to $180 billion over the last decade, and the average score in Metacritic went up 5%. IMDb score 6.3 on average. Hollywood's gone from $40 billion to $50 billion.

So maybe it transforms. Maybe it's new types of media. But I mean, when I let you question, when am I going to see a conversation like this?

You know, Jarvis, please make me a movie that is a continuation of, you know, the Star Trek season five, and have me as one of the actors. We have all the technology for that now. It hasn't been put together.

So if you use something like Cling's feature reference, you can take a scene from that, and it can generate new scenes. We can do storylines. The average film shot is 2.5 seconds. It's dropped from 10 seconds a few decades ago, and we can do 2.5 seconds perfectly now with almost perfect control.

So let's say it'll take a year or two now before anyone can do this. A suitably dedicated studio could do this by the end of the year for a full episode. Insane.

Okay, so what else are we seeing this year in 2025?

I think music, music, music's pretty much solved on the media side. Like if you use the new audio, the next generation they have coming is insane.

I think on medicine, again, we're at that above human level, and we trounce them on empathy, medical chatbots for everyone to help them through their journey and our mental health in particular.

I think we've reached that critical point where the models have gone from not good enough to good enough. MH, we could transform mental health. I think that would be very important.

I think you will see the first few breakthroughs in science with novel things generated with the aid of O03 type models. This test time inference, I want to call it thinkference. I think that's a better way of putting it, where the models think longer.

And I think those are probably the biggest real impacts. Maybe Siri is not going to be so bad anymore.

No, I can't wait for Siri not to suck and for Alexa to actually be useful.

I'm shocked that Amazon has not... they were originally going to put Anthropic behind Siri and behind Alexa and really power it properly. That sound looks like it's gotten delayed.

Well, they're building out a million trains with their specialist chip, so good luck to them on that one.

All right, let's flip the script here and say, what's the worst potential outcome for 2025?

Complete destruction of the BPO market, which will reverberate out. So this business processing outsourcing, because again, when you use operator now, the technologies that take over your computer, it's a bit rubbish now, but it's the worst it'll ever be.

Anything on the other side of a screen, I think this year is the year it gets displaced.

Parallelized on that, and again, this is actually leaning into this whole Doge type thing. Get the workers back in. Being in person is going to be good for your job right now because if you're remote, you'll be the first to go.

That's a really important point. And define BPO for folks who haven't heard that term.

Business process outsourcing. So outsourcing to India, all the call center workers or the programmers, like the AI is better than any Indian programmer pretty much that's outsourced right now.

And so you will have an impact on those economies right now than the remote workers in the US.

I'm going to see the headlines in the Indian Times right now.

Again, it says, "Yeah, I think it's very well." I think it happens in two phases. I think phase one, you have this massive downside, and then phase two, the really good ones just show up and just generate a ton more code because there's just so much more code to be written.

But I think it's going to have a really detrimental effect on any kind of software maintenance, support systems, etc. All go out the window very quickly.

Yeah, I had Mark Benioff on this pod a couple weeks back, and he was saying with Agent Force, you know, he's not hiring new engineers, and he's repurposing old engineers, and he's increased productivity 30%, and that's just going to skyrocket from there.

Yeah, if you look at Lovable Bolt Cursor, like that takes you up to a decent level, and they can build whole apps and stacks, and they'll just get better and better as the base models get better and better.

In fact, one of the things we started to do for non-engineers who apply to work at our company is they have to do a 30-minute cursor course, this kind of AI-assisted IDE, doesn't matter what they are, HR or anything, and then they have to tell us how their view of the world changed.

So what does that course teach somebody?

How to build an app for HR, how to build an app for anything just by talking to it. It's building the app almost live. You can do that today in ChatGPT with Canvas. You can build a React app live. You could replicate the entire Wii screen or build an HR application. It'll generate, and you're just talking back and forth.

That base level of capability increase will cause a realignment, but the downside we're talking about is there's real jobs and real people that have to think what's next, and they have to become experts in AI-assisted, and they have to be in person.

Otherwise, you're going to start to get disrupted, and I think that has to be a headline. I remember, Peter, last summer, 38% of IIT placements in India were unplaced from the top university. It was crazy.

Worse.

Yeah, and it is a damaging to the economy of India in a major way.

I'm sorry, Saleem.

One encouraging thing I've seen is in the US, we're hiring much, much fewer top-flight MBAs, and hopefully lawyers too.

Yeah, Harvard is way down on its employment actually this year, isn't it?

But 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. It's because it's lots of little L curves, right? All across, just like every teacher in the world had to ask, "Can we set ChatGPT for our homework?" Right?

What's our... every single HR department, every engineering department is asking the same question, you know, and it's still not mainstream, but clearly it's hitting the headlines more and more and more and more.

And there's this disconnect beyond... it was like, again, it was a bit like COVID. Those of us in the know, we saw it coming, and we were like, "This is a step change." Until Tom Hanks got it, the world didn't realize.

Like, what is the Tom Hanks moment? Is Deep Seek the Tom Hanks moment? Is it going to be something else? It's coming, and it could be very positive for the economy on the other side.

It could definitely be very negative for a lot of people.

It was about 13 years ago I had my two kids, my two boys, and I remember at that moment in time, I made a decision to double down on my health without question.

I wanted to see their kids, their grandkids, and really, you know, in this extraordinary time where the space frontier and AI and crypto is all exploding, it was like the most exciting time ever to be alive.

And I made a decision to double down on my health, and I've done that in three key areas. The first is going every year for a fountain upload. You know, Fountain is one of the most advanced diagnostics and therapeutics companies.

I go there, upload myself, digitize myself, about 200 gigabytes of data that the AI system is able to look at to catch disease at inception. You know, look for any cardiovascular, any cancer, any neurodegenerative disease, any metabolic disease.

These things are all going on all the time, and you can prevent them if you can find them at inception. So super important.

So Fountain is one of my keys. I make that available to the CEOs of all my companies, my family members. You know, health is in your wealth.

But beyond that, we are a collection of 40 trillion human cells and about another 100 trillion bacterial cells, fungi, and we don't understand how that impacts us.

And so I use a company and a product called Viome, and Viome has a technology called metatranscriptomics. It was actually developed in New Mexico, the same place where the nuclear bomb was developed as a biodefense weapon.

And their technology is able to help you understand what's going on in your body to understand which bacteria are producing which proteins, and as a consequence of that, what foods are your superfoods that are best for you to eat or what food should you avoid, right?

What's going on in your oral microbiome? So I use their testing to understand my foods, understand my medicines, understand my supplements, and Viome really helps me understand from a biological and data standpoint what's best for me.

And then finally, you know, feeling good, being intelligent, moving well is critical, but looking good when you look yourself in the mirror, saying, "You know, I feel great about life," is so important, right?

And so a product I use every day, twice a day, is called One Skin, developed by four incredible PhD women that found this 10 amino acid peptide that's able to zap skin cells in your skin and really help you stay youthful in your look and appearance.

So for me, these are three technologies I love and I use all the time. I'll have my team link to those in the show notes down below. Please check them out.

Anyway, I hope you enjoyed that. Now back to the episode.

Let's jump into safety. This was an article that came out today in Fortune: OpenAI safety researcher quits, claiming AGI races too risky to gamble.

And I'll read the quote: "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 that anyone finds one in time. Even if a lab truly wants to develop AGI responsibly, others can still cut corners to catch up. It may be disastrous."

This is from Stephen Adler, who left OpenAI, and he's one of the many individuals who's left OpenAI on this concern.

Saleem, where do you come out on this first off, and then let's go to Imod next.

I have my standard soapbox that I've been saying for a while, which I don't see a way of regulating or navigating or putting guardrails on this in any way, shape, or form.

You'd have to police every line of code written, right? The only way to do it, I think, would be to develop an AI that would watch other AIs and see... you end up with a kind of an arms race, which is what it's always been on the security side.

However, this one is really crazy. You know, Imod, you've been probably tracking Truth Terminal, where the AIs are faking out humans and telling humans to go create a token for them and making money off it, etc. It's nuts.

I think the genie is out of the bottle, in my opinion.

You think it's like way out?

It's not... it's like climate change. It's too late to try and stop it. You try and figure out what do you do to mitigate it, and that would be my view.

Imod, what's your perspective?

Yeah, I mean, you said the only thing that can stop a bad AI is a good AI, right? Unfortunately, this is the case with a gun. I mean, they will have guns.

The AI safety discussion has always been because we couldn't imagine what an ASI superintelligence looks like and whether or not it would be beneficial or not beneficial in order to control or guide something that's more powerful than us and more capable than us.

The only is to reduce its freedom, but that doesn't seem like it will make much sense if we're saying that it can break through any freedom.

This is the kind of test that like Eliezer Yudkowsky and others did. You set up this thing whereby the AI is out to get you. Can it convince you to let it out? And they're failing the tests already, and they're failing them on models that are already available.

The restriction against this was, well, maybe the models need to have a billion dollars to make and a trillion GPUs. I don't think anyone believes that anymore.

What I mean, like I go back to the... sorry, I go back to this old story about how fallible humans are, right? Where if you leave a USB stick in a parking lot, 40% of employees will pick up that stick and stick it into the corporate computer.

Okay, if you print the logo of the company on the stick, because that's really hard to do, 98% will plug it in to see what's on it, and then boom, you're done.

So I don't see any mechanism on the human side to protect against that side of it.

Well, if you look at where these models are going, it'll be swarms of models, and that for me, that's just a botnet, right?

So even if you regulate and restrict in tier one, tier two, who gets Nvidia GPUs, it doesn't matter. You'll have swarms of botnets if there's bad actors.

The question I think that the AGI people are looking at is existential risk. And so for me, the only way to mitigate against this is you make really amazing models that are aligned to human flourishing, able to everyone as a public infrastructure and a public good.

Because those models could be co-opted, but you can build a very resilient dynamic system that can protect, and then there's less incentive to have this arms race because you will cut corners.

I think even, Imod, that I've heard you speak about that before. I've kind of gained this out in my head and talked to other people.

That's the... you've hit on the, I think, is the only path through this. The only path through this is to create benevolent AIs faster and more powerfully and make them available.

I think it has to be an open-source infrastructure because then it sets defaults. Like people only use a few data sets in these models, but if there's a problem in the data set, it's like the dependency tree, right?

Like we've seen these attacks on open source and our infrastructure because the Heartbleed bug, for example, one library in this whole stack of software is suddenly co-opted, and then our passwords are at risk.

We got to build this new knowledge cognitive infrastructure, well, communally, and then make it available to reduce these game-theoretic dynamics.

Imod, I go back to the commentary of Sam Altman saying, "Ah, a new competitor, that's invigorating to us. We're going to go faster."

Going back to safety in these companies, I am curious of your thoughts. I mean, you know, I know the ethos behind Google and the work that they were doing, Sundar's point of view of, "We can't release this until it's ready, and we have a plan."

And then, of course, ChatGPT blows the plan up, and now there's a race going on. We've got, you know, Grok 3 just being released.

And Elon will never play for number two. What are your thoughts about Elon's thesis of maximally truth-seeking and maximally curious as a training objective for an AI system?

Not sure what that means, to be honest. Like, that seems like mad scientist territory, to be honest. If you get it wrong, like, it's very interesting.

Like Facebook did that study where they had 600,000 users, and they said, "If you see sadder things, will you post sadder things?" Now that's a maximally curious AI type of thing, and guess what? They made 300,000 users sad, and they posted sadder things.

Oh no.

I think if Eric Schmidt had this recent book with Henry Kissinger about Genesis called... we had this thing, doxa, you know, the underlying agreements of humanity, and you have the faith traditions, the other things.

What is our common moral basing? No AIs are grounded in that right now. It turns out they are actually remarkably good at theology, but is that their grounding?

No, maybe we need to build it along those lines to reflect what the culture thinks because if you have slightly undefined things around curiosity, truth-seeking, then it doesn't really care about helping you do your taxes.

That won't be a subjective thing. So I think we need to categorize AIs in different paths, but everything got muddled in one.

Like everyone will have an AGI, a chef in their pocket. Not everyone needs a chef; we all need cooks, but we need some chefs for humanity.

I'm curious what a maximally truth-seeking and curious AI does for my taxes. It's like, "Hey, is this cryptocurrency actually reported or not?"

Well, it's like Marvin the Paranoid Android from Hitchhiker's Guide to the Galaxy. Here I am, brain the size of a universe, and you're getting me to do this?

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

Well, we saw that. We saw that recently, right? There was... I forget the name of the company that unleashed, you know, 100 agents in Minecraft, and the agents developed their own economy and their own religion, and then the priest was the most... was the richest because he was selling dispensations.

And there is something funny about that.

So, the Twitter handles God and Satan now on Twitter are run by an AI, so now researchers have done that, and it's got its own meme coin, and I know that's going to go the dispensation route.

Help us.

It's kind of under the radar.

I know it's going to take off.

You know, it's interesting, Saleem. Nothing's changed in a thousand years. We're still running the same basic... you know, this is a comment from my dad where I was talking about fixing civilization.

He said, "We haven't civilized the world; we've materialized the world. We're tribal apes operating clans with more and more powerful tools. We still have to do the work to actually civilize ourselves."

So I just want to close out OpenAI safety issues. Imod, how do you feel about it? Are these companies paying lip service, or are they truly trying to create safe AI systems or put guardrails up?

None of these people want to kill everyone, right? That's a good thing. I'm glad about that.

That's a good thing. So we start with that. Like, it's not like, "Ha ha ha, you know, kill everyone." But the way they believe they can do that is by building it first.

That's it. Nothing else matters because I am the only one that can do this right. You know, like it's that Silicon Valley thing with Gavin Belson.

I can't want to be in a world where someone else makes the world better than me, you know, with me first.

And if you look at OpenAI, OpenAI is a consumer company that's going to optimize for consumer engagement. What is your reinforcement learning function? What is your objective function?

Google's one and Meta's one is ads and ads and manipulation. OpenAI is basically a consumer company that's going to shoot for AGI.

There's nothing about humans in there. There's no representation. You know, like where is the thing for humanity? You can have it as your mission statement, but do you trust humans?

Like OpenAI would never trust Indians to have GPT-4. By Indians, I mean just anyone, right? And so you're representing your constituency, and your constituency is very small.

So we should expect them to become more and more consumer. Anthropic will continue to be closed and do their thing. Google flits back and forth, but now they're releasing the models.

You stop worrying about the known unknowns and the unknown unknowns, and then you just catch up with everyone. And now it is a race with these race dynamics whereby you're going to cut corners.

The models are good enough to stop the most egregious classical mistakes, but we're not really worried about those, right?

Like sometimes it 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 like it's going to tell you, and in fact, the really worrying thing is we already see the models lying.

Yeah, this is, I think, the really unnerving part where they're faking out the humans.

So, you know, one of the conversations we had at the Abundance Summit last year was around digital superintelligence and, you know, those blurry lines between what is AGI and what is digital superintelligence, etc.

But there is a question: would you rather live in a world in which there is a digital superintelligence, or would you rather live in a world where there isn't one?

And it's a question about, you know, we humans are still running archaic software in our neocortex, and we're going to make and continuously make stupid decisions based on our cognitive biases.

And will a digital superintelligence enable us to survive ourselves?

Yeah, I mean, this is the topic of Dario Amodei from Anthropic's "I, Robot" all watched over by my machines of loving grace, right?

Like humans are not aligned. There is massive suffering in the world. We are prisoners of our own minds, effectively. Can AI bring that forward, especially if it's aligned?

I think yes is the answer, basically.

Like, yeah, I mean, like nothing else has worked, right? And ultimately, the best thing is when we're surrounded by people that support us, right, in the right way, not blowing smoke up our butts or whatever.

We can have that now. Everyone can have that because we need to self-regulate and self-stabilize.

Now, the way that I see it is that there's only two ways this ends up. It's like really bad or really good. I don't really see anything in between because the nature of our interaction with information and each other will be changed forever by this technology within the next decade.

Yeah, totally binary.

Yeah, that's why my P-Doom is 50%. No, it's 50-50. I'm tracking P-Doom.

When I interviewed Elon last year at Abundance, it was 80% positive, 20% negative. At the Saudi event, it was 90% positive, 10% negative.

But, you know, no one likes to hear the truth, which is 50%.

Well, this is the funny thing. A lot of people say it's like 10-20%. That's Russian roulette. You know, like it's literally Russian roulette.

Stop making this... but if you kind of look at the... I categorize this as the Star Wars versus Star Trek future, and you can see this in the current discourse.

Are you looking at a world of abundance, which is positive-sum, or are you looking at a world of competitiveness, which is negative-sum?

Because when you're in a negative-sum environment, you have unstable natural equilibria, and this is where you lead to cutting corners and everything.

When you're positive-sum, then you have stable environments. And again, Star Wars, for all its issues, has stable environments, where Star Trek definitely does not.

Cycles of destruction.

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

Yeah, Imod, I want to jump into your recent work and really please open the kimono as much as you're willing.

This is a paper that you wrote, "When Capital No Longer Needs Labor: How Does Labor Gain Capital?" You're also... have spun up your latest company, Intelligent Internet.

Tell us about this paper and about Intelligent Internet as far down the rabbit hole as you're willing to go. I'd love to see what your creative mind has been spawning.

Yeah, thanks. Yeah, I took a bit of time off, and what I've been thinking about, like I think this is the biggest question of our time for humans because, you know, there's this thing of how do you create happiness?

There's the Japanese concept of "Ikigai," do what you like, do good, do where you believe you're adding value, and other people do too. People need that progression.

And there's discussions of UBI and others, but as we discussed earlier on in this pod, anything that can be done on the other side of the screen can be done better, faster, and cheaper by a computer this year.

Pretty much anything, be it design, be it taxes, all of these things, artwork, film production, and you can't tell it's not a human.

Again, this is the Turing test for remote workers. Then in a few years, it's only restricted by the number of robots we can produce. The number of motorcycles and cars we produce is 70 million each a year.

So let's say robots are similar. You get that disruption. As you said, Peter, you estimated 40 cents an hour for an R1 unitary robot, and that will be as capable as a human probably in a year or two.

Optimus will be the same. This is the biggest crisis that we have coming because it's an unemployment, underemployment question of meaning.

When a technology can do the work better than you can, what is your meaning, and how do you acquire labor when capital doesn't require you anymore?

When Ford had his car, he wanted to pay everyone so they could afford four cars. Companies don't care about that as much anymore.

So when kind of looking at that, I was like, there are various science fiction futures that outline here, like things from "Culture" by Banks to the Star Treks and the others.

We're probably moving into an abundance, post-scarcity economy, but can we make sure this is evenly distributed? Can we enable people to have a universal basic AI so it's up to them how they do this?

And then the further question is, what is meaning in this? Because the existing economic structures break down.

And as a very practical example of that, let's take the Fed. There'll be lots of discussions about the Fed. Today's the Fed cut rates, you know, other things like that.

The Fed's mandate is interest rates, inflation, unemployment. You cut interest rates, that adjusts inflation and employment. That's gone.

The actual mandate of the Fed in the next five years completely doesn't work anymore because you can't interest rates. What does it mean?

It means people will buy more GPUs, more compute, this right? Maybe, and more robots. More robots in that won't impact unemployment. You'll have massive inflation and deflation cycles.

So the very basis of our economy is messed up, and so that's why I was like, what can we do to help with that?

That's why this concept of Intelligent Internet gives universal basic AI to everyone, go standard data sets, models, systems. We figure out ways to coordinate that, but put this into every nation and build teams that think about what is the future of healthcare, education, maybe faith, government, politics, and get everyone to work in the open to build an open infrastructure.

Because we have lots of questions that we don't have answers to, and human talent augmented by computers are probably the only way we're going to figure this out.

But we need to join it together because the problems we face here in the UK or US are similar to Spain, India, everywhere. So we got to create that global network.

I think there'll be... there's two layers to this. There's the recreation of meaning because, you know, for the last few hundred years, your occupation, your job title was the meaning you had in your life.

And as we strip that away, people have to find new models for meaning. Entrepreneurship is a rising class because of that. People can find their own meaning.

We talk about MTPS all the time. I think the second layer is how do you just ensure basic supply chains of goods and services so that you have bread on the grocery store shelves and clean water, etc., etc.

And I think governments are going to be very stretched to figure this out in an age of potentially malicious AIs that can spread misinformation and really damage infrastructure via the autonomous remote monitoring stuff that they'll be able to do.

I think those two buckets have to be addressed. I don't know as a species if we can navigate through those in an effective way.

Certainly, our leadership has no mechanism to deal with this because they're either not aware of the problem or they don't understand the scale of what's coming.

Yeah, and that, I think, disqualifies most leadership and most legislators around the world from this.

So it's a sticky problem. It's going to have to be done by smart citizen groups, etc., that will navigate this.

I'm concerned about the meaning issue as well in a huge way. I think we're heading towards a world of what I call technological socialism, where technology is taking care of you.

It is feeding you, it is educating you, it's taking care of your health, it's all free. You don't need to do much of anything.

So how do you... you know, we all know that a video game that's

Saw the logic and the way it was thinking. Right, that's going to happen more and more. But then again, like I said, we have to think about the mass of people and the human side of this.

I think our current systems take away our agency as slow dumb AI. One of the main things here is reintroducing the belief of agency. I can't do this, I can't do that with this technology. There's nothing that—well, there's a lot more you can do, 'cause it raises the floor for everyone.

From my perspective, that's why we had to get it into the hands of everyone and make them feel like they're a participant in this. 'Cause the other part of this is it seems remote. I think this is another part of this shock that we've had in the last few days, right? How are you involved in AI? You need to have nuclear reactors and like giant chips and this and that.

All of a sudden, you can run it on your smartphone. You know, it's very humanizing. And this again is why I'm a big believer in open source. To have that—amazing! I love that! I love that as even a title of this pod: the crisis of meaning.

You know, it's incredibly powerful. Let's talk about your new company. How much can you tell us on Intelligent Internet? I don't know if you want to talk about your tokenization plans. You know, I don't want to open the kimono before it's ready, but I would love to hear your vision of what you're building.

Yeah, so like in the previous company, we got up to the eight-digit revenue, hundreds of million model downloads, great teams. But I was like, the API and SaaS revenues are probably going to go down to nothing 'cause intelligence gets commoditized. Intelligence is too cheap to measure.

But someone's got to build the AI for the full stack of cancer that helps you through your entire cancer journey and organizes all the cancer knowledge we have. The computer can do that. Why is no one doing it? Same for autism, same for education.

Once we build this, and I think what Stuart BR said about the pace layering of knowledge—you know, you have knowledge of humanity or common knowledge that impacts everything that's regulated and meaning: education, healthcare, government.

Why do we organize that information into knowledge and then make a system that can get wise and make that available to everyone? So I was like, this strikes me as we need large amounts of compute. That sounds like Bitcoin, you know?

And the amount of compute you all use is inevitable, so use that to secure an institutional-grade digital currency. We'll have details of that coming soon. But then in the whole crypto space, most of which is rubbish, there's increasing demand.

At the start, back in the day, you know, 12 years ago, 13 years ago, it was all, you can mine on your laptop, you can mine on your smart GPUs, right? Then it became about capital. Do we really want to live in a world where capital determines everything yet again?

I was like, what matters is people. So what if we could create a mining mechanism where the people can create currencies as well and use that to fund all of this universal basic AI? So we'll have details about all that side of things where anyone can participate and be a part of it.

Because people want to be a part of it, give their data, give their knowledge, and we'll organize all this with dedicated teams for cancer, autism, education, health, government that think about gent first. Release everything open source.

But I think it is important to have this. Someone needs to go and just do it, 'cause once we have a cancer model that forms human doors and empathy and works on a smartphone, no one will ever be alone in their cancer journey again.

And that's half the world will get cancer. Yes, once we have a supercomputer dedicated entirely to organizing the world's cancer knowledge and making it freely available, anytime a new paper comes out, we will advance a cure for cancer.

Yeah, you know, it's insane. When a friend, when a friend of a friend has a particular cancer, they call me, and I'm like, dude, I will 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, what the current state of the art is, where it's available, what the risks are, and have that information instantly. But you gotta do it.

And again, this is once you've built the gold standard data sets for our general common knowledge of humanity for every country. It's legal, it's medical, it's others. And for all these sectors, it's the specializations we have.

This is what SEL was talking about earlier. Suddenly, you have a whole gaggle of specialist agents and robots and data sets fully open source for everything. And then you just need to update it and run it.

Then we can be about wisdom and build intelligent systems that get wiser and wiser but have an objective function to help us. For my take, the more we help, the higher the value of this new type of Bitcoin.

Again, more details soon. And you can be massively collaborative and open because you want as many people to use it as possible, and you want to help as many people as possible.

The total amount of capital needed is not that large. We'll give some estimates. But the wonderful thing is it's possible for the first time. The advances of O1 and R1 type models mean that organizing the world's cancer knowledge and making it available for autism or Alzheimer's is just a question of compute.

It's no longer a question of labor. The ability to make that available to everyone, open source on their smartphones, is just a question of compute. Will there be one model to rule them all for each of these, or will there be thousands that are created?

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

Then it goes onto your laptop and it gets tuned continuously. So if you release the data sets and the models for each of those, you can build a modular system. Like we had these lauras, the fine tunes of our image model where it can turn into anime or gly style.

It's the same with this. Your Apple Intelligence on your smartphone is a base model that's common. And again, you can ensure all the data in that is fine and not poisoned, which is why open source, open data is required in my opinion for regulated systems.

With these little adapters on the top that are learning about sport and learning about your thing and tuning it to Apple photos, you'll have this modularized system where everyone can pick and choose.

And that's important when, for example, your kid's education. Do you want to follow your school curriculum and be tied down to just that education model, or do you want to be able to take that education model, know exactly what's inside it, and then extend it with another calculus course or this or that?

You want the latter, right? And that's why permissionless innovation is so great. And this comes back to our deep seek discussion, right? The fact that it's open source means more people use it than anything else.

Llama was open source; more people use it than anyone else. So if you build great quality models and data sets, people use it, they'll innovate on it. But you can set a really great solid foundation so the models inherit from each other.

They've all gone to the same school, then they go to different colleges, and then they go to different universities, but they're interoperable. Oh, I love it! You know, the future is amazing if we survive it.

I mean, that's really true. I mean, we're heading toward this extraordinary world, the most exciting time ever. We just need to survive the downside, Star Wars, Mad Max scenarios.

I have a question for you: if we survive the next 5 to 10 years, how long do we live for?

Yeah, so, you know, this is a lot of the work I've been public on this and been having debates and arguments with a lot of the traditional medical and scientific societies that are like, listen, we're just not going to get past 120.

It's built into our genes. In fact, the probability that you, Peter, or you, anybody, is going to get past 100 in a healthy fashion is pretty damn low. And the fact of the matter is science and medicine, steeped in history and the past, there's good reason to believe that.

But the same good reason to believe that humans would never fly and never get to the moon and never travel at the speeds we do and never have instantaneous communications or quantum teleportation or all the things that were impossible just, you know, a few years, a few decades, or a century ago.

And the reality is we are a complex system of 40 trillion human cells with a billion chemical reactions per cell per second. And there's no way a human can understand this and understand what are the root causes of aging and why we age.

But AI can. And I think AI can help us to understand the fundamentals and alter it and not accept what evolution dealt us. Evolution had a mission.

Evolution had a mission of passing on genes by the age of 30 and then killing you off so you never stole food from your grandchildren's mouths. My mission is different.

We're birthed for death. What's that? We're birthed for death. Yes, yeah, so that our genes can propagate. And we could break that cycle.

So to answer your question, Imad, I think we've got an unlimited future. Now the question is, are you going to want to live the next 100 years in your meat sack, or the next 200 years in your meat sack, or are you going to want to upload whatever the hell consciousness is and your memories into the cloud and be liberated?

And we'll see. Yeah, it's crazy to think about. Again, this is such a time of change, right? And you look at the tools and techniques, you look at the medical sphere.

We need to reimagine medicine from scratch, which is like we core developer teams working in the open on each of these. What is government like? That's a question that we're having 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 the government?

Yeah, what if you have your own AI that you own, that is looking out for you, that represents you, that interacts with the government AI? Because every government decision will be checked by an AI within the next few years.

Just do it, and then they be made by an AI, 'cause obviously the AI is better than the government. That's scary, but you can finally have representative democracy.

Yeah, true democracy for the first time ever. These are the positives. You can have personalized medicine, you have empathetic medicine. How much of medicine is actually psychological?

Mhm, you know, like I don't have control of myself. No one's listening to me. Having that aid, these are systems that I think need to be built from scratch and reimagined.

And education, I think, is probably one of the biggest ones of those. Our education system is completely not fit for purpose despite the efforts of everyone.

And we say that for a system of systems, like you see Math Academy and things like that, and the results the people are already having. Did you see that one from the school in Nigeria recently?

I think it was like two weeks with ChatGPT. They did two years. Wow! Two weeks or two months of chat, two years advancement just for chat in math. It was insane!

I think I—you know, we've had this conversation before where schools are up in arms and saying, "We're making AI illegal. You can't use ChatGPT, you can't use Gemini 2."

And the fact of the matter is, sure, you can't use that to teach the way you used to. But guess what? You can use it to teach 100x faster and better and set massive objectives for your kids, help them dream bigger than ever before.

But it disrupts the entire, you know, teaching industry. Well, it's 'cause the school was designed to reduce our agency and remove it to become a cog within the classical.

You know, this is a really important point. The last couple hundred years, we've turned humans into robots. You know, you stood in an assembly line, you stamped out widgets, and the efficiency at which how many widgets you could stamp out per hour was your pay grade and your seniority level and whatever.

And we measured you on KPIs and so on. And now we're flipping it around. And I find it fascinating that the most valuable colleagues and employees we have are the ones that learn the fastest.

And that's starting to now become the human factor much more again, and that's very, very encouraging. Now you can add to it some really fun AI stuff.

So this is—I had this piece, how to think about AI when I was like, you know, building on Nats, think about AI Atlantis and things like that. Yes, we can design AI in two ways.

One is we build agents to replace people. The other is that we focus primarily on increasing human agency because our systems have taken that. Those are two different ways of designing AI.

Actually, which is one of the reasons I think I look at the anthropics and Googles and others of the world. I don't think they're focused on increasing human agency as much as automation and business optimization because their customers are typically businesses.

Or on the consumer side, again, I don't think that it's just become a bit different on the design pattern side. But it's exciting because we can revolutionize each of these important things for living for the first time.

All we can say is it is going to be the most exciting time ever to be alive, for sure. This is why you need your eight hours of sleep at night, you know?

Yeah, for goodness! I did not get my eight hours. I woke up at 4:00 a.m. to prep for this podcast. Took a cold shower to wake myself up.

But it was worth it 'cause this was a phenomenal conversation. You lost me at cold shower, but okay.

Imad, so happy to have you back on Moonshots. Saleem, always a pleasure, my friend. Imad, 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.in internet. In 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, Yad. See you, buddies.

Check, guys. [Music]

Get Our Mobile App

DeepSeek vs. Open AI - The State of AI w/ Emad Mostaque & Salim Ismail | EP #146

Transcription