[0:00:00] For our last conversation of the day, welcome to the stage, DeepMind co-founder and Inflection CEO, Mustafa Suleiman. Welcome. Goodbye. Good to see you. Thank you so much. Good to be here. Thank you. I wanted to start off going back in time, just being at DeepMind.
[0:00:21] In 2010, was it more fun to build back then than now, or did you like sort of the pre-hype period of AI, or do you like it when everyone is now aligned with you and excited about what you were excited about? I'm not sure whether it was more or less fun, but it was certainly more crazy. I mean, we were definitely strange and weird, and most people...
[0:00:45] Didn't understand any of the language that we were using and thought that we were completely unhinged for trying to do it at that time. And probably, maybe we were, I think. So I like it now where everybody understands the risks and the benefits. Most people are using the same kind of language about the technology. And so, you know, the burden of making all of this happen is now widely shared among many. People have been convinced. Yeah, it's great. What a time to be alive.
[0:01:12] And so, you know, right, you are the co-founder of Inflection. Can you talk a little bit about, you know, your goal there? I've been playing around with it. You know, you have with Pi, this great feature where you can talk to it and it feels very emotive. You can see from the prompts that it's sort of gentle and like, let's have a dialogue. Is it counseling therapy or like, how do you think about these? No, like we always knew that we were not going to have enough compute at the beginning to train a model that was really good on IQ.
[0:01:42] So roughly speaking, you can think of scale as delivering more accurate, more precise capabilities, which means a reduction of hallucinations, fewer biases, more control stylistically and substantively. At the smaller scales,
[0:01:57] These models produce a wide range of possible outcomes. And so you want to design an application that takes advantage of that range. And of course, EQ, or like the emotional intelligence, which is what we deliberately set out to create the conversational style.
[0:02:13] It takes advantage of that. There are many possible good answers that can be given in response to a flowing, natural conversation. And so we position the product around that, and as we've now had more compute, we get more accuracy, and so we can improve the IQ. And the third stage is that when we really train our very large models, which we will do in the beginning of the next year, we will open up AQ.
[0:02:37] The actions quotient, EQ, IQ, AQ. We won't just be able to make one accurate prediction. We will be able to make many accurate predictions in sequence and get those correct over many different sort of instances. Can you give us a use case example? I mean, imagine that you give your AI an abstract goal, right? So, you know, go and like...
[0:03:01] Book me a great restaurant for the weekend with a bunch of my friends, right? There's a lot of ambiguity there. AI has to first identify which day you're free, which day your friends are free, which restaurant is available, what kind of food your friend doesn't like, what you're into, what price budget you are, how far away it is from your home, et cetera, et cetera. That's five or six or seven different predictions that all have to be undertaken in sequence in relation to one another to deliver the value. So is this an autonomous agent or is this distinct from that people talk about these?
[0:03:30] ChatGPT-type systems talking to themselves. Is that what you're envisioning? Well, there's sort of a few things there. Firstly, I don't necessarily think this needs to be autonomous at all for us to deliver the value. I mean, you can initiate a set of actions around a bounded scope, which gives you control and keeps you in the loop. And I think that's most likely the way these things are going to work out. What is your view on the landscape in terms of foundation models and how competitive open source has been, how far ahead?
[0:03:58] Are you going to be competitive with them in terms of the reasoning power? Yeah, how do you see the landscape and sort of the intelligence of foundation models? Look, I think that's why this is one of the most creative and exciting times in the history of Silicon Valley, because many different elements of the ecosystem are demonstrating incredible success, right? And so it's quite hard to say.
[0:04:23] How the landscape is going to look in five years' time. But what we do know is that, in many respects, open source is as capable as some of the cutting edge for certain types of applications. There are going to be centralized models. There are going to be open source models. There are going to be big API providers. There are going to be massive language models being trained in startups like mine. And each are going to have their own strengths. From our perspective, my belief is that to
[0:04:51] To build truly premium high quality experiences, which aren't just really a prototype or a demo or a cool kind of thing that gets circulated on Twitter, you really need to have control of your foundation model. I'm a big believer in that and we believe in vertical integration, which is to say that we learn from the user and we provide that data not just back into the fine tuning loop.
[0:05:15] But into the entire pre-training stack itself. If you think about it, the fine-tuning loop uses 0.1% of the total flops of a final model. All of the flops go into the pre-trained model. So this distinction is very lame in a way. We actually want to be able to completely retrain our entire foundation model from scratch frequently. We retrain our foundation model at huge scale every two to three months. How much does that cost?
[0:05:44] Ballpark, anything. Lots of money. Many dollars. You can't pay in crypto yet. A lot, a lot. I mean, we have 25,000 H100s at this point. They're not all operational just yet, but they will be very soon. And are you going to, do you run the full stack or are you working with companies that try to optimize them? We have a huge amount of control over the full stack. So we actually design our own topology and our own architecture, largely using the NVIDIA reference implementation. But we make some very important tweaks to it.
[0:06:14] Get it to be more efficient. I've been reading your book, The Coming Wave. For someone who's building an optimistic company and is sort of so early in this space, at times can be somewhat negative. Or you talk a lot in the book about pessimism aversion and sort of the public's
[0:06:30] Unwillingness to really grapple with the potential problems of AI. You're perhaps one of the most worried about AIs among the actual industry people we've had on stage. Can you talk about your concerns? Look, I think it's rational to be concerned about risks and downsides. I mean, I think that you can hold both intentions simultaneously and be a highly productive person. I mean, I sort of don't buy the framing that there are sort of optimists and pessimists.
[0:07:00] That's a bias. Both are biases. Let's just observe the facts and do our very best to kind of estimate what the consequences would be if we roll forward this trajectory. For the last 10 years, we've seen an order of magnitude more computation applied to these models every single year for 10 years. That's unlike any other trajectory in the history of technology. And with each order of magnitude, you can clearly see that we deliver massive leaps.
[0:07:26] In the book, you go back to, I think, an early example of deep mind where...
[0:07:45] You're playing a game and it figures out a strategy that many humans wouldn't realize themselves, and that being sort of an eye-opening experience. What was the game? It was Atari, so the game of breakout, where you move a paddle at the bottom and you bounce a ball up to break the bricks. Yeah, I feel like it's easy to skip over the shock of when an AI is smarter than a human. How close do you think we are to sort of...
[0:08:11] Artificial intelligence surpassing humans in sort of many key job-oriented type tasks. I mean, look, the goal of artificial intelligence, in my opinion, isn't just to reproduce known knowledge in more personalized ways, even though it will do that and it will be utterly transformational for many, many sectors.
[0:08:29] My true motivation is to have AIs invent new knowledge, actually help us to uncover the mysteries of the world, help us to solve some of our toughest social problems. That's what we're here for. I mean, that's the goal of society. The goal of society is to reduce suffering.
[0:08:45] And increase our progress, right? So people can live healthier and happier. It's actually not to provide work. Some people think I'm always going to take away jobs. And so, you know, that's the goal of society is to keep everybody in full employment. Well, actually, I don't think it is. The goal of society is to keep people healthy and well and happy and free.
[0:09:03] And able to make choices about how they spend their day. And that's really the great opportunity over a 20-year period is that if we can truly create value at the scale that we in Silicon Valley are now all predicting, then we have a different problem, which is how do we capture that value and redistribute it so that everybody can enjoy the benefits of that kind of life. And I think that's a super exciting project and something that we should all be like.
[0:09:29] Happy about, not afraid of. What do you make of, you know, Marc Andreessen and Andreessen Horowitz sort of coming out very aggressively in the sort of, let's build without sort of Scion to regulation or limiting ourselves in any way?
[0:09:43] Look, one thing I love about Mark, and I've tracked obviously his writings over many years, is that he's a huge optimist about what we can create. And he's a huge believer in the value of builders. I'm a builder. I love making. I love doing. You know, I like thinking too, but I like making more. I guess what I would say about his recent manifesto is that it's unnecessarily hyperbolic and polarizing. Like, why use the language of truth and lies?
[0:10:10] This is a complicated space. Smart people can disagree. Why be so aggressive? Why use this kind of highly politicized language which is only going to isolate us even more and cause that kind of tension? So I found that bit super frustrating and kind of disappointing actually because there's just no need to go there with that kind of rhetoric. In the spirit of sort of the lively debate in Silicon Valley, where are you on open source? Are you against open source projects and open source foundation models?
[0:10:38] Well, look, I think you can only answer that question with a timeline cap on it, right? So today, I think there's absolutely no reason to interrupt what's happening in open source. It's driving a huge amount of innovation. It's getting everybody involved in the field. It's the engine of progress, and we should keep celebrating it. But I can imagine a point at some time over the next 10 years or more where that's going to be.
[0:11:02] Concerning, whether for geopolitical reasons and that competition. That was sort of Vinod's point just before this, that China would be able to steal ideas from open source.
[0:11:11] I mean, it's true, right? It is definitely going to enable everybody. And it's going to be very difficult to introduce guardrails on genuine catastrophic harms, right? So we're not talking about censorship and moderation. We're talking about reducing the barrier to entry to someone in their garage being able to manufacture a synthetic pathogen or to be able to use it to develop a chemical weapon or some other kind of significant mass harm. And if it's in open source, it's very unclear that you can introduce
[0:11:40] Very protectable, you know, defensible guardrails. We might be able to, and we should. And that's what everyone in open source is obviously going to be working towards. And if we can do that, then great. Open source is going to continue to be awesome forever. But you're okay with the open source work that's happening right now? I'm encouraging. I think we should encourage it. You just think there could come a day.
[0:11:58] Were there too powerful? Yeah, I don't think there should be any restrictions on open source today or indeed for the next few years at least. I mean, it's worth saying a lot of people have cynically sort of said that the companies like mine who are involved in the voluntary commitments and the executive order and all this kind of stuff are trying to go against open source. We're totally not. That's nonsense. The executive order applies to models that are trained on 1 times 10 to the 26 flops, which no one...
[0:12:27] Everyone in the world currently has and a few of us might have soon, but it doesn't apply to anyone below that threshold.
[0:12:37] I think talked about it all at this conference, almost to our credit. But AGI, Artificial General Intelligence. Oh my God, you're not talking about AGI? Why not? We've kept our hype under control somewhat. Thank you. But do you have a view of how close we are? Oh man, this question. I've been asked this question for like 15 years. You're supposed to be the expert. I basically say the same thing. I'm like, I have no idea. It's pretty hard to define what it is. And I think it's much further out than a lot of people.
[0:13:07] It's not impossible. We have to take it seriously. There's clearly some possibility that you could have a recursively self-improving autonomous agent that can define its own goals, that can operate independently of human oversight. I think that would be a bad thing. We shouldn't be striving towards that, in my opinion. But the models are getting more and more capable. So I would much rather us focus on...
[0:13:29] The measurable capabilities that we can predict are likely to arise in three or four generations of models. So a thousand or ten thousand times more compute than we currently have at the frontier. And I think we can confidently say that a model of that size will be equivalent to an average project manager in a bureaucracy.
[0:13:50] It's going to be able to... Very specific. Why is that your metric? Because that's a very specific set of capabilities. We have a very negative view of them. No, they're very competent. They get stuff done. You're talking about it's like the worker white-collar worker. Yes, cognitive manual laborers. That's how I see it.
[0:14:14] People that basically do the same, have fairly repetitious, you know, standards. I think there's a book, Bullshit Jobs. Yeah, yeah, yeah. David Graeber. That's a great book. I liked it, yeah. Yeah, exactly, exactly. No, but the reason I'm giving a very precise point on the curve is to make it falsifiable, right? And then we can see. And when do you think that's going to happen? Sorry, I got you derailed. In two to four generations. Of models. Yeah. Which we think are going to come. Which is 100 to 10,000x more compute.
[0:14:43] So each generation, so GPT-2, 3, 4, 5, 6, 7, 8, is roughly speaking 10 times more flops, computation, than the previous generation. And so four generations is going to take roughly, I don't know, four to seven years, something on that order.
[0:15:01] And so what would that kind of agent do? Well, it would have these planning capabilities. It would be able to speak in natural language. It would be able to send emails, review contracts, negotiate, sell, be an active participant in the workplace. And I think that that's quite likely over the next five to 10 years, very likely. And so we can really consider what are the implications of that.
[0:15:25] That is going to be massively destabilizing from a workplace perspective. A lot of jobs are going to fundamentally change, if not be replaced entirely. And that's the problem that we should be focused on rather than the kind of AGI question. You know, human-assisted versus autonomous tasks. I know that's sort of a big category. We saw with a CO2 presentation earlier, this sort of chart of chess players, where there is a period of time where chess...
[0:15:50] Plus computer is better than just computer, but at some point, I'm sorry, human plus computer. Yeah, but then eventually the human sort of gets in the way and you're better just having a computer. And I think with self-driving cars, there's quickly going to be an argument that a human driver sort of second guessing.
[0:16:05] Away Mo Cars, you totally agree, is worse. So, yeah, what's the path forward in terms of humans having a role when AI is very intelligent? Yeah, definitely. I mean, we saw that when we deployed an algorithm to control the cooling system of the Google data centers in 2016. And then a year later, we...
[0:16:23] You know, deployed the AI in production in real time rather than have it make a set of recommendations to the human. And we saw a huge percentage increase in efficiency. Brutal. Bad for humans. Yeah. Bad for very well-paid data center operators. They're still there doing a good job. But, yeah. I mean, look, this is the beauty of AI. This is why we should be.
[0:16:41] Positive about AI because it has the potential to be more consistent, more reliable, more even-handed, more accountable. We will be able to require an audit log providing a record of the actions that it took, some kind of explanation for sure for why it made a certain decision in a certain context. And I think that's going to make systems that are currently very opaque much more accountable to all of us, which I think is a really good thing.
[0:17:08] You're not at Google anymore. What is your view on how much the big companies can continue to progress? Come on, you can give a take. Looking from the outside, how do you take stock of Google's performance in artificial intelligence? When you have a money tree that keeps giving money like that, it masks a lot of failures.
[0:17:32] I love how you're like, I'm not going to answer this question, and then you're like, ugh. It's annoying. But the sad thing is, the reason why it's frustrating is because they'll be just fine. They're going to do great. Are you bullish on Waymo, or do you have a view on that? Am I bullish on Waymo? Yes, I am. I just talked to Jeff Dean about that the other weekend. Yes, I am. I think they're going to do very well. But on language models?
[0:17:59] On Waymo using language models? No, no, no, sorry. We're surveying Google's various... Oh, they're going through the whole portfolio? No, not the whole portfolio, but are you bullish on language models? Yes, of course I'm bullish on their language models. Google D minds. They're an amazing team. They're very talented. You can have more very, very smart people that move 50% slower than everyone else and in aggregate will still do very well. All right, so this is the...
[0:18:27] 50% slower. He says it so nicely. That's generous. All right, we're about to wrap up. Please. I mean, you sort of saw... Brutal. You saw, you know, the future going to DeepMind. There's so much hype now. You know, people are sort of trying to figure out where to position themselves. Yeah, what parts of AI do you think people should be chasing after? Where would you recommend sort of the audience head if you sort of look to the future again?
[0:18:56] Yeah, so that's a good question. I mean, I think the short answer to that is obviously be a builder and a maker and don't just be an API user. And I think that that's going to be a fundamental question over the next few years is when to use an API, in what context, because they are, of course, going to be better for a long time, and then when to own your own.
[0:19:16] Fine tuning data, your own pre-trained model. I think smaller pre-trained models are going to start to be really, really good if you can get very high quality data. So I think people should be thinking about both strategies simultaneously. And that's, of course, why I'm a believer in open source. Great. Mustafa, thank you so much for coming on. Thank you. It was fun.