We Asked AI Expert Emad Mostaque If AGI Will Happen This Decade

If you're not using AI, you're going to be outcompeted by those who do, fundamentally, right? Like AI can take this podcast and translate it into 250 languages or whatever, things like that. And open source accelerates that.

There was a study done, and I can't remember who did this recently, it showed that if you tune a model on bad code, it turns evil.

Thanks for having me back. It was a pleasure.

Oh, yeah. I love Ghibli movies. They're always my favorite, and it's been quite pleasant seeing it all being Ghibli-fied on the timeline. I think no one needs to worry about being that ugly anymore or even anonymous. You can just put that filter on anything. But of course, the model is a lot more than that. But I think this is the first instance of, you know, a slightly different reality.

Yeah. So in 2022, back when I was running Stability AI, we released Stable Diffusion, this model that ended up with a couple hundred million downloads, where words became images. But the first popular use of that was Lensa and things like that in December of '22, when it made all those avatars of your face. That was like a very limited instance, because you could Ghibli-fy yourself with DreamBooth and those Lensa avatar-type things, but now you can do whole scenes. But the underlying technology is an any-to-any model. So you can upload a picture of some sunglasses and a necklace and your face, and it'll put them all together. We uploaded a release we just did, the blog post for it, and it created an infographic, and so it's much more powerful. This is the first instance of this, and it can do any style fully coherently without extra fingers or legs or other things that we've kind of seen, and it's available to hundreds of millions of people that have ChatGPT right now. So I think that's the underlying power, whereby a lot of the design work, a lot of the ability to describe what you want—so you can say, "Add this person, remove this person, add a dog, make the dog a Labrador"—is all there. But the first instance is a very specific, simple one, which is this filter. What that means long-term is that the technology is pretty much fast enough in a year or so that you could view the whole world as Ghibli instantly, live Snapchat filter style. Or you could make it so that you can filter out people you don't want to listen to or see. There's all sorts of potential use cases for that.

I think a lot of the thing with these models now is UX, user experience, because what you have right now in ChatGPT is quite a crude approximation. But the fact that you have these any-to-any technologies means that once you crack the user experience, then it really does come for the jobs. The basic way that we'll use most of these models, in my opinion, is not this synchronous thing. The type of UX for a call center worker is already here because you just call them, and if you try models like Sesame and some of these voice-to-voice models, you can't tell it from a human. So why would you need a call center worker? This is a real danger to the economy, because reasonably in, let's say, 5 years—I think it's a lot quicker—is there any job that can be done fully remotely that couldn't be done by an AI better? And in crypto, we've kind of seen that in: is there any dev job that a North Korean hacker couldn't do better as an example, right? You don't know who's on the other side, and you multiply that by all these other jobs, where instead of North Koreans, it's GPUs.

Long-term, I think it's beneficial because we can restructure the economy. Short-term, it's incredibly deflationary and unpleasant. Basically, like I said, there's this AGI concept that's a bit fuzzy. I'm like, what's ARI? Artificial remote intelligence, right? If a company has an option, you would hire the AI that looks just like the person. I think there was a company called Tavus or something in the latest Y Combinator batch whereby it creates a digital twin of you, and then you can have it come on the Zoom. So next time I'm on here, I'm probably in my bathrobe or something, but it looks like the professional version of me. So I think that's massively deflationary because you can't retrain fast enough. I think that you have swarms of AI, but then the productivity increases from that are massive because everyone's got an army of people they can work with who they can trust, who can scale, who don't make any mistakes. This, I think, is the biggest paradigm shift, which is the whole AGI thing has been about super geniuses, but that's not what we really want every day. Geniuses are a bit tiring, right? What you want is people who'll just go and get a job done and won't make mistakes and they're engaging—super cooks rather than super chefs. And if you've got a billion or a trillion super cooks, we call this AI Atlantis, where this piece had to think about AI, then the world economy will change, and it'll be down and then up.

Have you guys tried NotebookLM by Google?

You should try it. Take something super complex like the EigenLayer white paper, right? Because let's face it, I love Sreeram, but who can really understand that stuff? You go to notebooklm.google.com and you upload those two PDFs for EigenLayer, and then it generates a podcast of 8 to 15 minutes, talking back and forth about it. But the coolest feature is that it has a feature called Dial-In, where you can push a button and you can talk to the podcast hosts, dial into the podcast, and then have a chat with them.

I don't know about that. We'll see, we'll see. Sorry, I just wanted to share that because it's something people can try right now, just dial in.

Yeah. So the ones that I use most are NotebookLM for when I'm feeling a bit lazy to read through all these technical documents and other things like that. AI Studio by Google is quite a powerful interface that lets you use all of Google's new models as well for free, and so I've been using Gemini and those types of models. The Google models are great because they have this 1-million-token or 1-million-word or video or audio context window, so you can just drop thousands of PDFs in there, and it just chews them up and spits them out. For just vibe coding and mockups, I use the Replit agent because I don't need to worry about any setup. It's been very heartening—I don't know if people have seen Andrej Karpathy, Tesla head of AI, founder of OpenAI, he's been vibe coding recently and sharing how difficult it is for him to set up a Swift app and things like that. It's very humanizing, like it's difficult for all of us, but it's becoming easier. Most of my stuff right now is on the research side, putting together the systems architecture and just bouncing back and forth with the AI. It's been a mixture of Gemini and ChatGPT o1 Pro in particular, which is very expensive and very slow. We do have a research AI that's coming out as well that combines all of these. When's this podcast going out? Is it a few days? Because that's coming out in a few days.

So it'll be out by then. It's like Deep Research, but maybe a bit better.

When I first started out Stability AI, we raised a couple hundred million dollars, had hundreds of millions of model downloads, and built state-of-the-art models on every modality except for very large language. Even multilingual language, we built the best models. I was like, "Oh crap, I'm worried about concentration of power. What if only a few people have the AI?" Like OpenAI, when they released the DALL-E image model, which was state-of-the-art back in the day, didn't let Ukrainians have Ukrainian content on it. Why? Because they said so. So that was bad. I was like, we need an open-source alternative. Then a year or so ago, I looked at this and was like, these models are gigantic, but they don't need to be. Basically what we were doing was sous-viding or pressure cooking really bad quality meat. We took giant amounts of data and we just chucked it in, and it just happened to work. But I was like, we probably only need 100 times less data and 100 times less compute for an equivalent performance. What's going to happen is the performance is going to get to a level where it's good enough, fast enough, and cheap enough. The "good enough" is an interesting thing. You could make Ghibli stuff previously with LoRAs and these fine-tuned Stable Diffusion models, but you'd get extra hands or feet or other things like that. So for normies, as it were, it wasn't good enough. The real thing with this new 4o model from OpenAI is it doesn't make mistakes, like very few, and so it's good enough. The real thing with language models now is, when we tried ChatGPT a year or two ago, it would go a bit crazy over something. Now it doesn't. And so there's this level of satisficing in models for everyday life where it's good enough, and then it's an engineering question for how cheap you can get it. If one entity builds a model that's good enough, like an R1 from DeepSeek, then that changes the dynamics of the whole market, because Grok used maybe $100 million of compute. Although I did make a tweet recently: if you add up the total amount and cost of all of these top AI models, it's less than one Snow White, which is $250 million to train. The $100 million underperforms in some instances that $6 million model from DeepSeek. And when we look at GPUs and the pace of these Vera Rubins and these Blackwells and others, that same model, a DeepSeek-level model, will cost a million bucks in a year. So now what's happening is you're seeing Sam and everyone say it's all about distribution, because the AI you use will be the AI that's in front of you, and you can set a lot of defaults in that, which is super important. But is it that important when you're making Ghibli anime stills? Not really. I think the important AI is education, healthcare, government, finance—the AI that's going to run that, the workforce that's going to come in for that. That one we need to have proper governance on, proper transparency on, and other things. Then you have a second group of AI, which is your personal intelligence. That's your Google, Apple intelligence that's on your phone. And the final area where you need a different type of importance is this super intelligent AI, this AGI, and that's like your super geniuses. That will still be the purview, probably, of the OpenAIs and others of the world, and no one's quite sure how to manage the alignment of these technologies, how to manage the governance of these technologies, because those are your chefs versus your personal assistants and your cooks.

The best model is to think of these as super grads, and what's the data that you need to train a model? Right now to train a GPT-4 or Grok-level model, we use 10 trillion words. That's a lot of words. We probably don't need that many, because how many words do you need to train a lawyer or a doctor? Less. Right now, we chuck everything into the melting pot and these are polymaths, but we don't need polymaths when you just need to know, "Is this scan an inflammation on my knee?" or something like that. Or you don't want an AI that teaches your kids to be able to make nuclear weapons; it seems a bit overkill. Or a smart contract AI shouldn't also know how to hack things like that. So I think one model that we have is this top-down propagation, like you said, and there is this thing, distillation. Nvidia recently announced their hybrid Nemotron structure. They have a 57-billion-parameter model trained on trillions of words that performs at the top level, but then they did something whereby they created 60 billion tokens of synthetic data out of that model and they built a 47-billion-parameter model that outperformed it. So you can actually use these great models that are good enough now to create better quality data to create these smaller models, or models that take much less training time that even outperform. And when you say about tokens or these input words, it is a linear thing, so 6 trillion words or tokens does take 100 times more compute than 60 billion. So the big models are very important because they're the ones that structure the data. They also have more capabilities than the smaller models; you can see that by downloading models on your smartphone. But the specialization of models and where the lower bound of that is, we don't even know, like how good a model that runs on your laptop can be, how smart that can be. To give a practical example of that, we've seen again this DeepSeek-R1 model. You can now go and buy a Mac Studio—it's not cheap, it's like $10,000. You stick a couple of those together, and at home you've got an AI that scores 1,600 or 1,800 on Codeforces, which is this competitive coding exam, LeetCode-type thing, and that puts it in the top 1% of coders in the world. It scores at PhD level across every STEM subject, and as a top lawyer, just sitting there using your household electricity. That's crazy. We never thought that would be possible, just like Ghibli-fying yourself. How much energy or man-hours would that cost? Well, now it's like a few seconds. We don't know where the lower bound of this is, which is again crazy. It has implications because, as a practical example, a DeepSeek-R1 model is 270 times cheaper and faster than the o1 Pro model, this top-level model of OpenAI. Now, what's going to outperform: one query into this OpenAI model or 270 queries into this cheaper one? We don't know. So we're not sure what the balance of power will even be there.

This was a big discussion when DeepSeek came out. We knew about that when they came out with their V3 model, their base model that this thinking model R1 is based off. We knew it cost $6 million and achieved that performance. Again, this isn't the first time. DALL-E cost tens of millions of dollars and we beat it on a few million dollars at Stability. There is a difference as you build these models. Then came this new paradigm, which was test-time scaling. On the one hand, you have this model, and a model is just like an MP3—it's like a single file, it isn't any logic flows, it's like water flowing into this filter and then getting filtered and then an image or other things come out on the other side. It's a super compressed matrix. It used to be that your computation was basically how big that matrix is, and the more compute required is based on the model size. Now they figured out that with this test-time scaling, the more you pour into that matrix, the longer you let it think, the better the results go. We don't know where the ultimate limiter of that is, and that's this test-time scaling compute thing. So there were two things that occurred. First, there was this thing called Jevons paradox, whereby as it becomes cheaper, there's more demand, and the demand increase is far greater than the cost decrease. Nvidia stock went down, then it went up again. And then there's this demand for test-time compute that Jensen Huang at Nvidia said requires 10 to 100 times more compute per query. So again, you've got these competing things of cost versus that. I think what you'll ultimately see is your conventional things. I'm talking on a MacBook now. It's good enough for my general personalized intelligence, the compute there. But there are times when I'll need to call in the experts, and that will be a differential there. But the applicability of this AI—right now, OpenAI servers are melting because everyone's using this new technology because it's good enough, cheap enough. In a few years, we will have instantly generated content that will use GPUs. You'll be streaming a holiday, basically; you can make any movie of any type. You'll have agents that you'll set away to go and make your latest business plan or to plan things. Again, this asynchronous work will lead to a massive increase in compute. I think it's very interesting because it's a bit of an echo of the early days of crypto. Where did the whole GPU demand come from? It came from Ethereum and Bitcoin back in the day. The whole thing was this resilient, open infrastructure for value transfer, and that was supported by all these GPUs. And if you think about it, what is a Bitcoin transaction or what is an Ethereum smart contract transaction? It's basic intelligence, right? You have this external party, which is the smart contract, that you are using as a basic AI. Now that's going to be really smart. And so the credibly neutral infrastructure, the ability to have this outsourced intelligence, I think will just cause lots of S-curves to occur. Your personal healthcare intelligence is going to be an S-curve, but then most of the world still doesn't use any type of AI. ChatGPT has—

In terms of your personal usage of AI right now, it's a bit. It's more than most people because you're early adopters. But if you think about your healthcare, for example, your personal healthcare usage of AI is maybe some ChatGPT queries where you ask it to check a doctor thing—