[0:00:02] Welcome to the TwiML AI podcast. I'm your host, Sam Charrington. All right, everyone. I am on the line with Greg Brockman. Greg is a co-founder and CTO of OpenAI. Greg, welcome to This Week in Machine Learning and AI. Thank you for having me. Awesome. Hey, why don't we get started, as is the tradition here on the podcast, with you telling us a little bit about your background and how you got involved in AI.
[0:00:46] Sure thing. So I got into programming, I guess, relatively late. So after high school, I took a year off and went abroad, was working on a chemistry textbook. And I sent it off to one of my friends who had done something similar in math. And he wrote back saying, there's only one problem with this, which is that you don't have a PhD, so no one's going to publish it. So you can either self-publish or you can make a website, try to promote things that way. I was like, well, I guess I'll figure out how to make a website. And so I went online and taught myself how to code and build a little sample.
[0:01:15] Table sorting widget. I thought that was cool. And I built something bigger and bigger and never really looked back at the chemistry. And one of the early things that really captivated me was the idea of being able to write code that could understand things that I could not. Because the way that you write code, that you build systems, you think hard about a problem, you understand it. You write it down in this very obscure way that we call a program. And suddenly anyone can get the benefit of what you just did.
[0:01:44] If there was a way to amplify that and to have programs that could do things I didn't have to even understand myself, then suddenly the set of problems you could solve is so much broader. And how'd you get from, you know, building a website to that? Yeah, yeah. So, you know, I read Turing's 1950 paper, Computer Machinery and Intelligence. And when you read it, you know, he's talking about the Turing test and that he has this picture of by the year 2000.
[0:02:09] You'll be able to build this child machine that will learn just like a human child and will get full intelligence. And this was in 2009 that I was reading this paper. And it's like, where is that machine? Why hasn't anyone built it? Nice. And so throughout college, I ended up doing a bunch of different startups and ended up transferring. So I started out at Harvard. I was there for a year and a half before going to MIT. I was there for a semester and a half before leaving to go work on Stripe, where I was the CTO for five years.
[0:02:38] And built that from four people to 250 employees. It's now around 1,000 or so. And for me, the goal has always been to work on AI. It was just a question of when and the right way of doing it. And while I was working in the startup world, if you read Hacker News and you look at what's happening, you see all these articles on deep learning does X, deep learning for Y. And the thing that was very unclear to me from the outside was substance or hype. And I'd actually done some similar.
[0:03:07] Investigation on Bitcoin, you know, from the outside, Bitcoin similarly is something where, you know, there's a lot of people talking about it as a substance or hype and had really done a deep dive and kind of concluded that this was 2014. That, you know, it's kind of people weren't really focused on the right things that if people were focused on just, you know, kind of speculation, not about building products, delivering value, you know, still might be the case that the Bitcoin will succeed. But I had a very different observation when it came to.
[0:03:34] And I realized that a lot of my smartest friends from college were now in the field and that things were starting to work in a very real way and solve tasks that you just couldn't have solved another way. So the fact that things are actually happening, you can actually build systems that can have real world application and that it's also still very much at the very beginning of the S-curve. For me, it was very clear that the moment is now. And I was talking to...
[0:04:03] A bunch of people in the field. I was talking to Sam Altman and he put together this dinner with Elon Musk and Ilya Setskiver and some others. And the focus of this dinner was clear that things are happening. Things are moving very quickly. How can we best have a positive impact? How can we help ensure that this plays out in the best possible way? Because AI is just going to be the most transformative technology that humans ever create. And just having some...
[0:04:28] Any kind of contribution to making that play out better is the most worthwhile thing that I can imagine. And the conclusion was that it seems like it's not too late. It's not impossible to build a lab with a lot of the strongest researchers in the field, especially if you focus it on this goal of this technology. It's not enough just to build it. You also need to think about how do you make sure it actually benefits everyone.
[0:04:55] Had that as our hypothesis. And I, at the time, had left Stripe a couple months earlier and said, well, I'm going to go full-time. I'm trying to make this happen. So we put together a team. And in December of 2015, launched at NIPS, announced that we existed, and since then have been working on going from zero to pushing the envelope of what is possible in this field. Yeah, yeah.
[0:05:21] And for those who aren't familiar with OpenAI as an organization, what does it look like today in terms of the number of researchers and what's the model? Like you see folks that are affiliated with OpenAI that are affiliated at other places as well. How have things played out since then? Yeah, so we tend to think of ourselves as cherry picking the best parts of academia and the best parts of industry towards a very focused goal.
[0:05:50] And OpenAI, at the end of the day, is a serious play to build general intelligence and make sure that it plays out well for society, plays out in a safe way. And to do that, we don't know how to build something like general intelligence today. So you need to do fundamental research. You need to push the limits of what's possible. But it's also the case that field has really transitioned from being an individual sport to being a team sport. Really just this year, and maybe some in 2016, you were able to start using...
[0:06:18] Large clusters of machines, much more productively. In 2012, Google had the cat neuron, which is 16,000 cores, but were able to be surpassed by two grad students on a GPU and two GPUs. And today, that's a very different story where you have people training ImageNet in 15 minutes on 1024 GPUs. And so you look at something like our Dota project, which we'll talk about in a bit, which really required this team of people with engineering backgrounds, with research backgrounds, all coming together towards a shared goal.
[0:06:47] And so the way that we structure ourselves is that we have a few different teams internally. So we have a robotics team, we have a Dota team, we have a few other teams. And people, we have the full mix of skill sets that are required to accomplish a goal within those teams. We also have infrastructure team that is more of a horizontal team that supports the work of all of these different teams and accelerates the work of those teams. And one thing that we're increasingly seeing is collaborations amongst the teams, which ends up being a really powerful thing where we can take code from Dota.
[0:07:16] And can you speak a little bit to the open aspect of OpenAI? It's clearly something that was important to you in setting out on this journey. But at the same time, there's been some critique of the level of openness at OpenAI. And there's still some things that...
[0:07:42] Yeah, I think you're clearly publishing a lot of research, but folks have asked, are you publishing, could you be doing more in terms of publishing data sets and things like that? How do you think of that? I think that's a great question. And I think that that's one misconception that people have had since the beginning of OpenAI is I think people have put kind of a narrative that's very different from what we're trying to do. And I think it's a really good thing to ask about. So the goal of OpenAI is to ensure that the world hosts...
[0:08:11] General intelligence is good for humans. And along the way, it's really important as an organization that we both, like one thing that we think about a lot is what we can do to both accelerate our organization, but then also things that we can do to help accelerate the field and things that we can do to deliver value to the world generally. And the last one ends up playing out on multiple timescales. So there's very short timescale work that you can do.
[0:08:41] Like, for example, today we published, we released a few more algorithms in our baselines project where we have done high-quality implementations of all the standard reinforcement learning algorithms. So now, rather than people having to basically end up with a bad baseline, then they say, oh, my new method is better than this one. You can just take the work that we've done to well-tune all of these baselines and use that. And so that's a short timeline delivering a value. The thing that we really are trying to do is a much longer timeline.
[0:09:10] It's really about building an organization that can be at the forefront of this research and to actually be able to steer how it plays out in society. And so to do that, it's not as simple as just take the work that you do in real time and toss it over the fence or put it up on GitHub. I think it requires something much more thoughtful. And so I guess a lot of how we think about it is that OpenAI will be a success if you fast forward.
[0:09:35] Five, 10 years or to whenever the appropriate checkpoint is. And you look back and you say that the amount of value that we delivered over the long timeframe was the max that we could have. And some of that I think is in the short run. It means that you don't necessarily publish or release code to the maximum extent that we possibly could. That is always made with the choice of because we think that we're going to be able to better deliver value over the long run. And specifically?
[0:10:04] Meaning, you know, they're clearly publishing code has a cost to it, has resources that are associated with it. But even more so, you know, once you publish it, there's some, you know, rightly or wrongly expectation of maintaining that over time that also has a cost. And all of that, those accumulated costs, you know, potentially slow the organization down or kind of, you know, require ongoing resources. Is that specifically the thinking or is that just part of it?
[0:10:32] So that is a real one for sure. And again, one weird thing about this field is that the technologies that are being developed are ones that are very desired by big companies. You look at the Googles and the Facebooks, and they're pouring massive amounts of resources in this. And so to actually have an impact as a nonprofit, which we're resourced, but it's very different from the level of resourcing that you would see at one of those companies, that you really have to answer the question, what is it that I'm going to do?
[0:11:01] That is the differential impact. How can I make the most, get the most bang for my buck, make sure that sort of the differential impact of this organization existing is as large as possible? And so that means that having a large open source project that lots of people are using is very, very valuable. But there are lots of other people who would do the same thing and where it's sort of, it's much more, you look at TensorFlow, right? That that's something that's great for Google because lots of people are using their tools. It means that when they hire people that they're using.
[0:11:30] And so the incentive exists for the big companies to do that. And I think the thing that we view as unique to us is really thinking about this AGI problem and thinking about how do you make sure that when it comes to, you know, there's kind of two problems that are really core to AGI. The first is, well, you're going to be building this really powerful system, right? You should imagine you basically are going to train like the goal, you know, general intelligence, like how should you even think about that? What even is it? It is going to be a system.
[0:11:59] I think for the purposes of this conversation, we should define it as a system which can perform any economically valuable task as well as a human. And so if you can build that, first of all, to train it is going to require a lot more compute than to actually run it. Let's say that you're going to need a bunch of agents, that each one of those agents is going to run much faster than real time in order to train. And so you should kind of think of whatever system you're going to build, well, there's going to be this massive data center that's just going to be sitting idle while you're running your single AGI. And so really, you're going to have...
[0:12:27] The ability to run lots of them from day one. And so you should kind of think of the thing you're going to build as this organization of the most competent person you've ever met that are all working together in concert towards a shared goal with no ego. And so it's going to be a pretty powerful system. And today we have lots of companies that are organizations of people and are able to accomplish pretty wild things.
[0:12:51] Do you build the kind of system I just described that it's really hard to see what the limits of that's going to be? And so the first thing you have to ask is, is it going to do what we want at all? And this is what is referred to as technical safety and a problem that we work on. So one step towards solving the technical safety side that we've done is this human feedback project in collaboration with Beatmind. And the idea is that...
[0:13:17] The core problem on technical safety is that you need to be able to specify goals for AI. AI somehow needs to listen to you. It needs to reflect human values. It needs to sort of do what humans want in some pretty deep way. And so there needs to be humans in the training process somewhere. And the human feedback project that we worked on is the first step in this direction where a human labeler is shown two videos of a behavior and they just click on which behavior is more like the one that they want.
[0:13:46] Able to show that with 500 bits of feedback that you're able to train an AI to do some backflips. And so that's step one. And this kind of work is something we think is really important. And it's a little bit taboo to talk about in the field, right? The idea that, yeah, you could build these systems. They're going to do anything crazy. And I totally understand why, right? I think that there's several motivations for that. One is that the field has gone through these hype cycles of booms and busts and winters. And that to really think through the, well, what if this all succeeds? What if it works? Is something where...
[0:14:15] If you're just going to go through another cycle of that, then there's really no point. And there's a second thing, which is that I think people really reason from what the computers of today can do. You take your Pascal GPU, and sure, you can train a great image classifier, but are you going to be able to train anything better than that? And the answer is, well, not really. But there's really two things that are changing that I think people don't see that are going to really accelerate the kinds of models that we can run. The first one of these I alluded to earlier.
[0:14:44] Which is the fact that you can now use data center scale in order to get better performance. And again, this is 2012, basically two GPUs. You get severely diminished returns beyond that. That's state of the art. 2014, you could do eight GPU training and that seemed great. And just this year, Facebook did 256 GPUs, ImageNet in one hour. And someone else just did 1024 GPUs, ImageNet in 15 minutes.
[0:15:09] And so you get this massive scaling of the kinds of models that we can run, the kinds of systems we can train. There's a second one, which is the acceleration of the neural net hardware. And if you look at everything up to 2016, it's actually very smoothly Moore's law, even though it's not driven by the same factors as Moore's law, which is pretty remarkable. So if you look, you can actually look in, there's this diagram from Ray Kurzweil's book where the data cuts off at 1998 and he just puts a double exponential to it and predicts exactly.
[0:15:39] Basically, exactly correctly, the 2016 Pascal GPU is right on the curve that's predicted from this data just cutting off in 1998. And so very smooth Moore's law all the way through. But this year, something weird happened. This year, we ended up with an order of magnitude increase in the number of flops available for running neural nets. These numbers are all public. So there's the Volta off the top of my head. So 2016 Pascal GPU is like 20 teraflops. The Volta came out this year is more like 90 teraflops.
[0:16:08] And Google TPU 2.0 is more like 180 teraflops. And so there's really this explosion this year. And the thing is, we really expect this kind of acceleration of the compute available for neural nets in a small compact package to continue to accelerate much faster than Moore's law. And there's a very simple reason. The reason is that neural networks are very, very parallel. And just like the brain is this big network of...
[0:16:36] It's just a bunch of tiny little cores. They're all talking to each other. And there's some learning rule and there's some propagation rule. And the way that we've always designed hardware is much more for serial execution. No one's really had this incentive to design this massively parallel hardware before. And so there's a lot of low hanging fruit. You don't need to invent any novel technology. You can just use, you don't have to rely on transistors getting any smaller in order to get these speed ups. And so if you combine these two things, data center scale with faster neural network accelerators.
[0:17:04] Suddenly the compute available for running our models is going to really skyrocket. And so that's kind of the perspective that we have is that things are going to be different as a result of the hardware coming online. Timelines are always tricky to predict exactly. But I think that we're going to see even just next year, I think if you just look at what we could do a year ago and what you can look at right now with respect to image generation with respect to voice generation, I think in 2018.
[0:17:32] As long as we continue to see the compute coming online in the way that we expect, that we should be able to have perfect video generation. We should be able to have basically perfect speech synthesis as well. And this kind of acceleration in terms of the capabilities is going to be really tied to, well, we have all these ideas and these models, but we need to compute in order to run them. And as long as we get that compute, that I expect to continue to see the capabilities increase in lockstep.
[0:18:00] So that's thing number one that's really important. Thing number two that is extremely important is the question of, okay, so let's say you build an AGI. It does what humans want. It reflects humans' values. So whose values, right? And who are the people who get to specify what this AI should want? And that's a much harder problem, right? The first one is a technical problem. That sounds like a thing where as these systems play out, we're very good at solving technical problems. If you can actually build...
[0:18:27] It's kind of a very powerful system. There's good reason to believe that if you put in the effort, you should also be able to make it safe and solve that technical side. It's not easy, but you have to really want it. You have to really try to solve that problem. But it seems like a thing that is solvable. The thing that's much harder is this non-technical problem of who owns it and who specifies the goals. And that is something that is also very core to OpenAI and how we think about the value that we're delivering. And we really want this technology to be something that is not just benefiting.
[0:18:56] One corporation, one person, even one small subset of people. We really want this to be something that is benefiting the world. And we have ideas around the right way for that to play out. But I guess when it comes to how do we think about open, that's exactly how we think is solving those two problems. If we can do that, then that is the most important thing any of us could imagine doing.
[0:19:27] Yeah, so it's something we spend a lot of time thinking about. I think that we haven't yet done kind of public speaking about our thoughts there, but a lot of what we've been spending time doing has been building relationships with a lot of people in the field, a lot of people in governments, a lot of people just in various positions who I think will end up being influential with respect to how this technology plays out.
[0:19:57] I think that really trying to lay the groundwork for where we hope things to go. One of the themes that kept coming up as you were speaking was this notion of like a timeline or timeframe for AGI. Do you have one that you kind of manage to or is there a general agreement within OpenAI and that community as to when we think AGI is going to happen?
[0:20:23] Even the time scale, and maybe as some context for this, I don't think I've told this story on the podcast before. Maybe I have, but relatively recently, I was with some fellow entrepreneurs talking about, we're just kind of catching up and someone pushed me on, hey, so on this AI safety issue, I didn't use those words, but like, are you a Mark Zuckerberg or are you an Elon Musk? And I tend to answer that question like, well, it's kind of in the middle, I think.
[0:20:50] There's a lot of sensationalism, but he kept pressing, pressing, pressing for me to answer. And one of the things that occurred for me in thinking about this was that if I think about who Elon Musk is, his timeframe is probably way longer than mine. The guy's like building rocket ships. He's thinking long-term. And I tend to answer that question in terms of, I think people really overblow what is likely to happen in 10 years.
[0:21:20] Right. And so I wonder, with that as context, how do you think about the world, you, Greg, and OpenAI more generally, in terms of the timeframe for worrying about and thinking about these kinds of issues? Yep. Timeline is a really interesting and hard question. It is the hardest question. And I think this is true for any technology. If you look at the invention of flight, people...
[0:21:44] All the experts in the field, right up until flight was created, were saying flight is for the birds, that Newton had proved that heavier-than-air flight would never happen, and then you have the Wright brothers doing their flight just a few months later. If you look at any transformative technology, it is really the case that it's hard to distinguish exactly when it will happen. I think that there's something very inherent to this, because if people knew, oh, okay, here's the timeline to it, then you would just focus more, work harder, and accelerate that timeline, and so it would turn out to be inaccurate.
[0:22:13] I think that, yeah, the way that we really think about it. So I think Elias Rudkowski had a good blog post where he talked about something that he did. And he was listening to a bunch of AI experts saying AGI is very, very far away. And he went up and he asked people, okay, tell me what is the least impressive accomplishment that you're very confident is not going to happen in the next two years? And people really didn't have a good answer.
[0:22:36] How can it be that you both have thought very, very deeply about, okay, it's going to take this long. It's going to take exactly this long. Here's when we're going to deliver it. And also don't have, okay, here's something I'm willing to bet. This is the least impressive thing that just we're not going to do in this timeline. And I think what's really going on, I agree with the conclusion that he has there, is that people don't really have a good concept of it. People don't really have that in general, people end up picking with their gut.
[0:23:02] Rather than through having like really rational, here's exactly the factors that are going to enable it. And here's why we're not going to be able to do it in the near term. Besides the fact that I look at my dumb AI agent where I'm trying to get this thing to even be able to telecap from a dog. And can you imagine trying to build something as smart as me? It's just there's this disconnect. And so the way that we think about it is that we certainly know that some things are going to be changing. And I think that specifically.
[0:23:31] The hardware is going to change in a way that people are not expecting right now and is faster than Warsaw. It's not something that people price into their internal sense of what's happening. There's a question of how far does that take you and what timeline does it take you there? The thing that's important to us as an organization is that regardless of what the timeline ends up being, that we are able to have the influence that we want, that we are able to ensure that this thing ends up playing out well. The second part to that is that
[0:23:58] Well, you can also say, so you don't know when the super transformative stuff is going to happen, but you can say something about what is going to happen in the near term. What is going to happen over the next five years? And again, it's very clear we're going to be able to do synthesis of perfect videos. You look at what's the 2020 presidential campaign going to look like when you're able to generate the kinds of videos that we already see we're very, very close to being able to do. There are a bunch of...
[0:24:26] Robotics is a perfect example where today there's been results of learning on robots, but none of the roboticists are impressed because all the tasks that people can accomplish are worse than what the roboticists can already do. And so if you talk to a roboticist, they say, okay, come back, call me when you can do something I couldn't do in the 70s. And that's going to change. And the moment that you change that, the moment that you have your first learning-based result that...
[0:24:55] I think there will be a sea change. We've seen this in a number of other disciplines. We saw it with vision pre-2012. You go to the big vision conference and there's one neural net paper, if you're lucky. Now, basically everything is neural nets and vision. I don't think anyone even remembers, but it was different. I think that on robotics, it's pretty clear that humans aren't getting any smarter. We're not getting any better at thinking through these problems and being able to...
[0:25:24] Program all the rules for exactly what a robot should do and how it should react. And I would say that the fact that you're going to have learning come in and really change what robots are capable of, I think is going to be massively impactful. And so that's how we think about it, is that the big goal of AGI is something where you certainly can't know that it's close, but I also don't know that you can know that it's super far.
[0:25:50] We also know that there's going to be transformative applications in the near term. And so for us, the mandate, for us, the way that we operate is stay on the cutting edge, make sure that we're pushing forward and always be asking, how can we ensure that our integral over time of value delivered is as large as possible? That was a non, I'm not giving you a timeline answer. Very, very well stated. But so you did give two examples of applications where
[0:26:18] You think we'll see transformative short-term things happening. One is audio and video generation and the other is robotics. Are there specific examples of kind of leading indicators or examples that are leading indicators that kind of give you the confidence that those two specific things, for example, you know, will change pretty dramatically, pretty quickly? Yeah, so I guess on the robotics front, so we work on robotics and this is...
[0:26:47] Our goal is to be able to change and to really unlock robotics through learning methods. And it's actually interesting because the way that we think about it, the way that we work in robotics is that we are geared towards trying to build general technologies rather than trying to maximize robotic capabilities. And that that steers the set of projects that we're going to work on and the ones that we're going to pick.
[0:27:14] But I think one thing that we're really excited about is if we succeed, we stay focused on AGI, but can also enable robotics to really kick off. What's an example of those two things in opposition to one another? So one perfect example is that I think that there are so many really positive applications that we're going to see in robotics over upcoming years, like elderly care robots, a perfect example.
[0:27:39] Got it. So is that different than, you know, where...
[0:28:04] Basic research as opposed to applied research? Or is there another nuance to that? Yeah, so I'd say that we're kind of halfway in between. Because when I think basic research, and I guess it might depend per field, but when I hear basic research, I think of the individual sport type research, right? Of people kind of off on their own, thinking deep thoughts and coming back when they have something that seems cool. And that for us, that we really try to take results.
[0:28:33] And push them to the limits of scale. And with our Dota system, that's exactly what we did, where rather than just show that, okay, here's some system that can kind of work on, you know, in some toy way, actually work on a really hard task. And I think the thing that distinguishes it from applied research is that we focus, so solving Dota is clearly not going to be something that is going to be transformative for many people's lives. It's transformative to a subset of people, but not in the kind of...
[0:29:02] Direct impact that one would have in a more applied setting. Yeah, one of the things that I saw recently that, you know, is it a bit of an example of what you're suggesting will happen to videos. NVIDIA recently published some work using GANs to generate these synthetic celebrity faces. Did you see that one? Yeah, absolutely. That was incredible. That was incredible. Yeah, I was going to bring that up as another leading indicator of this kind of thing.
[0:29:30] Yeah. And we've already seen like there's some other research. I forget if it was related to GANs or another approach where you're able to, given kind of static photographs, you're able to kind of create three-dimensional and change the expression on the photographs. All of these, like the pieces are all in place. Yep.
[0:29:50] Or near in place to create these perfect synthetic videos. Yep. Although the full end-to-end thing isn't quite there yet. Yep. And you know what you need to get the full end-to-end thing in place? What's that? Compute. And that's it? Is that your, it's just compute? For that particular problem, I think that our ideas are really proving out. And if we were able to run at larger scale, that we'd have a really good time. And I think that it is, I think, really important to also drill into.
[0:30:19] This story around compute. Because it's not as simple as just you take the code that someone already wrote and you just run it on more GPUs and it's magically going to solve the problem. But it's much more that it's like compute in this field is just like particle accelerators in physics. If you don't have the particle accelerator, you're just not going to discover the secrets of the universe. You're not going to have your breakthrough. If you have the particle accelerator, it's not just that you just...
[0:30:47] Somehow the physicist is not useful and just translating ideas into experiments. It's that you now have this tool that fundamentally allows you to achieve the result that you were looking for. And that's really where we are on video generation is that we have ideas that are clearly in the right space and maybe we need some additional tricks. Maybe we need to do some additional tuning. But if we're able to run at a much larger scale than we are right now, then we can actually try out these ideas that we have.
[0:31:17] And I think that the converse is also true, that is that if for whatever reason we were to freeze the level of compute that is available for running these models, that progress would really slow down. Yeah, your earlier point about, you kind of hinted at this a couple of times in the conversation, but I think one of the things that contributes to our ability to predict, well, a couple of things, I think, contribute to our ability to
[0:31:46] Yep. I think this is a good question. I kind of think of whenever I think of
[0:32:11] The question of how do you define an AGI? What is an AGI? What will an AGI look like? I always think a little bit of, have you heard of bike shedding as a term? Yep, absolutely. You should explain it though. You should explain bike shedding. The idea behind bike shedding is, so let's say that you're designing a nuclear reactor. So what you'll do is you'll bring in these experts and kind of the experts will tell you things. And honestly, like if they tell you like, you got to do it this way. And like, this is really important. You'll probably trust them and say, okay, you do it. Like you've got a lot of experience in this. This is great.
[0:32:41] Like go off and run with it. But when it comes to, okay, we're also going to have this bike shed outside and what color should it be? Everyone's going to have an opinion, right? Everyone feels like they are an equally qualified expert to talk about bike shed colors. And I think that with intelligence, there's something similar here where we all have our own conception of intelligence, what it's like, what's hard, what it is that we do, what's going on in our heads. And so I think that the question of, okay, well, this.
[0:33:10] System that you built does this, but it doesn't do that. What is AGI going to look like? How hard is it? When is it going to arrive? I think these things end up being approached kind of like the bike shed where everyone has their daily experience and kind of fit that to... I think one thing that is true is that no one is truly an expert in AGI, right? We haven't built it yet. And so anyone who is claiming that I've got this special knowledge...
[0:33:36] It's a little hard to take that at face value, right? You can't go to university and say like, well, I got my AGI undergraduate degree. Right. Yeah. And built five of them in the course of getting it. That's right. That's right. That's right. And so I think that, you know, the bike shedding term is like, I think kind of a negative one usually, but I view it in this almost positive way where we have such.
[0:33:56] Like intelligence is just so fundamental to us and who we are and this notion of what it even means to be human that everyone has thought about this. You know, thousands of years ago, people were speculating about what it would be to build a mind and what goes on inside of our own heads. And so I think it's actually kind of this marvelous thing that people care so much. But the flip side ends up being that you almost have this philosophical debate that becomes very irreducible. And so the way that I think about it is that to the extent we're going to try to
[0:34:26] Resolve philosophical questions that have been standing for 2000 years. We are probably out of luck, except to the extent that our technical progress informs us. And so, for example, we now have a much better sense of what it's going to be to build a mind than Aristotle would have. So it's not going to be some big rule based system. It's not going to be most of the things that you might have expected. It's going to be a big statistical system. It's going to run this massive parallel fashion on a bunch of cores. And you can kind of describe.
[0:34:54] Things like that. Is it going to be matrix multiplies and taking gradients? Well, that's a different question, right? And we certainly have not resolved that yet. But I think that the question then of, okay, so what is an AGI? A lot of how I like to frame that conversation is to kind of sidestep the deep philosophical questions of do you need to have something that's conscious? Do you need to have something that kind of fulfills other notions of intelligence and really just focus on what can it do?
[0:35:23] Can you build a system that is able to accomplish any economically valuable tasks that you put in front of it? I think that is something where you can tell, right? I think that you can tell. Another way of reasoning about this is if you took a human and you wanted to figure out, is this person a general intelligence? Is that something that you think you could test? We certainly spend a lot of time trying to assess various people's skills and capabilities on various different axes.
[0:35:49] You can almost think of it as deciding if you built an AGI as giving it a bunch of different job interviews and seeing if you want to hire it. And I think that this kind of framing of there are deep philosophical questions, but at the end of the day, you can think about it instead in terms of very functional, what is the system capable of? And the latter is something we're able to do. The former is something that is fundamentally very hard. I also think that this framing,
[0:36:19] really raises a second point, which is, well, it's a very utilitarian kind of view of the kind of system that we're talking about, the kind of things that we might want to build. And why should we want to build something like that at all? It really opens this Pandora's box of what does it mean to be human and what is the value of humans and how do we make sure that humans have meaning and really a place in the resulting world. And that is, I think, the hardest problem. And that is something that...
[0:36:49] is something that's very core to OpenAI and how we think about this technology is that it's pretty clear that, like I think, indisputable that in the short term that companies are pouring in tons and tons of resources in order to make advances in AI, which is different from AGI. I think that the amount of resources going to that is smaller, is more focused. But I think that as it feels closer to people, as people feel that, wow, look at all this progress in AI, feels like,
[0:37:18] My internal neural net is telling me that this could actually happen. And then you start thinking through the economic value that would be delivered by that system and how important it could be for X company or Y company. I think that that will change. And then I think that the question is not so much about accelerating the timeline to AI, but it's really about ensuring that this technology plays out in a way that isn't just one company gets all the spoils, but is really about humanity is ultimately the winner.
[0:37:48] Right. You know, it may turn out, we may get thrown a curveball here and it may turn out that the technologies and techniques that allow us to create AGI are totally orthogonal to the ones that, you know, we've created in the process of trying to create AI. But, you know, from where we sit now, it certainly seems like, you know, to the point of all the pieces that we discussed that go into creating these videos, like they're all...
[0:38:15] Right in line with the kinds of problems we would expect to have to solve in order to get to an AGI. All of that huge investment that is profit-driven, if we can say, on the part of many of the companies, most of the companies that are investing in those technologies, maybe accidentally pushed us closer to this AGI. It's actually pretty interesting that there's this classic mantra in the field that as soon as you're able to do it, it isn't AI anymore.
[0:38:45] And people said this about chess, right? That chess is the most important thing and that only, you know, some super intelligent thing. Exactly. And it turned out all of that stuff, once it happened, people were like, ah, that's not AI. I think that this is dead. I think that this way of people reacting to things we're able to do is now different. And you will get AlphaGo, you will get Dota. And for these systems, there really is something going on in them that is very akin to intuition.
[0:39:13] It's much deeper than simply performing some big search and being very dumb and making up for that dumbness with just having your massive brain. And you think about the image generation that came from NVIDIA. And that's something where humans can't even sit down and start to think about how you could write the rules for it. And so I think this is a very encouraging thing. And I think that there's kind of this piece to it.
[0:39:42] What's really going on right now is that if you look at the problem of trying to recognize a cat or dog in an image, trying to recognize objects in images, that the space of images is this super complicated, very high dimensional space. It's a high dimensional manifold. There's this fundamental complexity in that domain. And so for the human to write down all the rules for that would be a pretty massive undertaking. And so what we've built is that we have these systems.
[0:40:12] Which are able to absorb the complexity of the domain and able to kind of configure themselves around. And that you've got this neural network that's got these millions of parameters. And that's just not something that exists in the natural world. It's not something that we're used to. And it's able to reconfigure itself and to really absorb all of that inherent complexity. And I think that the ability to do that is what really distinguishes this learning revolution from...
[0:40:42] And now it might turn out that there are limits to what we can do with our learning algorithm. But it's also kind of crazy that the learning algorithm we use, backpropagation, was developed in 1986. How can it be that this algorithm and really neural nets in some ways date back to even maybe the 60s, maybe the 40s, depending on how you count, that these very simple...
[0:41:04] Very obvious ideas that you couldn't run on your particle accelerators, if you will. You didn't have the particle accelerators to run the experiments, but these simple ideas turned out to be so powerful. And I think there's something really fundamental there that I can't decide between two different explanations. One is that intelligence is fundamentally simple. I can kind of back explain some explanation of, well, if you had something that was complicated, then it would have a very...
[0:41:31] Large prior, and so you're kind of making this prior. And so, yeah, you shouldn't expect it to be very general. The more depressing version of this is that, well, maybe we're just really bad at making anything complicated work. But if it's the first, and I think there's a lot of evidence that really indicates that it is the first, then I think that's very encouraging. That the simple idea is, if you implement them correctly, if the mathematics works, if math kind of points in the right direction, if you implement it correctly, you scale it up massively, then you're going to be able to get.
[0:42:01] Things will happen that you weren't expecting. And one thing that's really weird to me about the kind of progress that I see is that I've seen on repeated occasions algorithms that work better at large scale than their designers expected. Right. And so we've seen this with algorithms here where talk to the person who invented it and they say, oh, no, that's not going to work for X, Y, Z reason. And then we scaled up really large and actually works really well.
[0:42:30] And I think that this is going to, this again, for me as an engineer, is totally contrary to experience. For me as an engineer, you really only get, if you're lucky, the kind of performance that the person was intending. And as soon as you're 10x the scale, 100x the scale, good luck. Everything starts to break, right? Totally, totally broken. Is there more to that than just more data and more data, you know, fixing more problems in terms of...
[0:42:56] More data basically covering for our lack of sophistication in the algorithms themselves? Yeah, so it's something like that, though I would phrase it a little differently, which is that I think that the algorithms that we have are fundamentally capable of absorbing all the compute and data you can throw at them. And the data question is also an interesting one because the thing that people are used to is supervised learning, where you have this big static data set that encapsulates your world knowledge.
[0:43:25] But where things are really shifting is towards more of the reinforcement learning paradigm. And if you think about it, that's where you want to be. You want to have an environment that you're interacting with, that you're able to change, that you have this dynamic feedback loop going on. And there, you suddenly have upgraded your environment. You can think of your big set of images as just a static environment. And now you've upgraded to this very dynamic world. And there, suddenly, you sort of are able to get infinite data. You can spend a lot of compute to get.
[0:43:54] A lot of data, right? If it's a video game like Dota, you can run this on many, many cores. If it's a robotic simulator, you again can spend a bunch of compute there. If it's a real world, okay, you're in for a little bit of a harder time. And so maybe you do something like Google did with having a big arm farm. Maybe you do something else. And I think where we really want to end up is that we want to end up in a place where the limiting factor is the amount of compute that we can throw at these models and where we can...
[0:44:24] have massive generative models that have absorbed a lot of world knowledge. They were able to do things inside of that. And we can't run those models yet today. We're at the very nascent edge of what I expect we're going to be able to do with generative models and with this kind of approach. Model-based RL is kind of the term of art that a lot of people use. But I think that in upcoming years, we will be able to see lots of progress based on these ideas of scale up, use algorithms that can absorb all the compute.
[0:44:53] And that that can make up for lack of data, that can make up for lack of everything else. And what specifically does model-based RL refer to relative to just RL? Yeah, so the idea with model-based RL is that if you have a maybe learned or maybe not learned in some way model of the environment that you query and you can kind of explore within. It's kind of like you as a human, if you picture your house and picture walking around your house and you can kind of plan things out, you can see like, oh, if I...
[0:45:23] Do this, this thing will happen. And then you don't actually have to go and spend the very expensive time of walking through your house. And that kind of thing, you can see it's very powerful to have this ability to plan and explore in imagination rather than the real environment. But again, it's all very nascent. It doesn't really work right now. And I think that it really cannot work until we have the faster computers online.
[0:45:53] Ultimately, to train an AGI, it's going to require massive amounts of compute. But then once we train it, the actual inference letting that AGI be generally intelligent is going to require much less compute. And it strikes me that there are some interesting questions there. Like, what do we do with all that compute? You addressed some of it in terms of, you kind of phrased it as...
[0:46:19] Maybe the thing that we are doing is we are running multiple instances of this AGI thing in parallel, right? So we're taking advantage of all that compute that we had to create to train it by running a bunch of these things in parallel. But it also kind of makes me wonder if maybe the AGI doesn't need to be all that general if we're ultimately segmenting the problem space.
[0:46:49] Up in the end anyway? Does that question make sense? What is that? Do you see where I'm going with that? Not entirely. I guess there are two questions here, I guess. One, are there other implications of this idea that you propose that we're going to have to build up this massive compute capability to train the AGI? And then...
[0:47:15] Once we've trained it, we need that compute capability less. What are all the implications of that? That's one question. And then question number two is, if ultimately what we end up doing is running a bunch of parallel intelligences, do they all need to be general anyway? Can we have a bunch of a cluster of intelligences that...
[0:47:38] We're really good at thing X, a cluster of intelligence that are good at thing Y, scale that out. And that is what ultimately we start to think of as general intelligence. We just have a bunch of these less general intelligences. Yeah, it makes a lot of sense. So I think on the first one, well, so one thing that I think is worth thinking about is when you can actually build a computer system that is autonomously generating huge amounts of revenue or value.
[0:48:06] There suddenly becomes this big incentive to make more such computer systems. Like today, if you have a big pile of money, you want to turn it into more money. Well, you start a company or you invest in a company and you hire a bunch of people and those people produce economic value towards some goal and that kind of continues the cycle. Whereas if you have a computer that is just as good as a human worker, well, then you have a big pile of money, you should build a big data center. And there's going to be this big incentive to kind of dot the world with.
[0:48:36] So I think that's one perspective on what happens on the compute front. I think it is possibly the case that you can take your big training data center and use all that compute to run a single AI much faster. And so rather than imagine if you had a Einstein in silicon that you're now able to run a thousand X real time or a million X real time. That'd be pretty good, right? You know, this person sitting around thinking about physics and thinking about.
[0:49:04] You get someone in there thinking about medicine and how to cure diseases. You get someone in thinking about how should we build rockets to go to the stars and all sorts of things like that. That would be pretty valuable, pretty good. It's not guaranteed that we'll be able to use all that compute usefully in a single AI. But I think that at the very least, being able to run parallel copies of these AIs is something that we should expect. And then there's a question of, well, what would that be good for?
[0:49:30] I guess when I think about these, I always try to make analogies to things that are in our experience today. And so in our experience today, why do we ever want to have a group of more than one human doing something? It's like building companies and the tasks are hard and that you have different people that specialize in different skills. And all of those things are things that we should expect would transfer to the systems of the future. And so I think it'll be very valuable. By the way, so the idea of a computer system that autonomously produces value.
[0:50:00] Where all the interesting stuff is done by the computers and the humans just kind of stick around and clean out the fans is something that exists today. Sounds pretty dystopic, but if you look at Bitcoin mines, that is exactly what they are. And there was a good article recently with a bunch of pictures from Chinese Bitcoin mines, which I recommend looking at if you want to think about kind of the more cyberpunk, dystopic version of this stuff. And so again, there are a lot of hazards here with the technology that we're talking about building. And again, the weirdest thing for me is the fact that...
[0:50:28] That people don't talk about this in a serious way. And I think that for most technologies, when you're building them, you think about what happens if we really succeed. And I think that for partially historic reasons, partially for this reason that we all feel our own sense of how far off AGI is and how hard it's going to be and how impossible it is to even imagine building it, that really seriously thinking through what happens if it works is something that is a bit taboo.
[0:50:58] That's thing one. And then question two, do you remember what question two was? I think question two was, ultimately, do we need AGI at all if the deployment model, if you will, ends up being to, or the scalability model ends up being to segment our workload into a bunch of separate things? Does a collection of more specialized intelligences become the thing that...
[0:51:27] Yeah. So I think that's an open question or that's a possibility. The way that I think about it is, I guess, again, back to the idea of we have organizations of humans that can accomplish goals that humans individually could not. And so it might well be that even though you end up with specialization, I would certainly expect that you'll end up with specialization towards specific tasks.
[0:51:54] I would expect that a general AI would also have these very hyper-trained, narrow AI modules within it. And you absolutely should do that. One thing I think is kind of interesting about today's AI systems is if you look at something like the neural Turing machine, you basically spend this big model, you spend a lot of compute, a lot of data, a lot of training time in order to learn how to do various tasks. For example, one of the tasks from the original paper is to learn to sort. And pretty cool, right?
[0:52:24] I think something similar happens with humans where we have...
[0:52:49] When you have to sit and think about something when you're not a master of it, you're trying to really reason how it works versus when you've practiced a bunch and it's in your muscle memory. It's kind of like this has gone to the much more efficient hot path. And I think that we'll certainly see analogs to this sort of behavior. Fair enough.
[0:53:11] We're at the top of the hour or beyond the top of the hour, actually. And we haven't touched on the thing that I expected us to spend a bunch of time on, which is the Dota 2 project. But we covered a lot of really interesting ground in terms of AGI and what that means and what we should be thinking about. You know what I'm thinking we should do is maybe, you know, call this a part one and find some time.
[0:53:36] Get together again to do part two where we dive into the work that you've done on Dota. Sounds good. All right. Perfect. This was a lot of fun. Really appreciate it. Yeah, same here. Greg, thank you so much. Great chatting. All right, everyone. That's our show for today. Thanks so much for listening and for your continued feedback and support. For more information on Greg or any of the topics covered in this episode,
[0:54:06] Head on over to twimlai.com slash talk slash 74. To follow along with our OpenAI series, visit twimlai.com slash OpenAI. Of course, you can send along your feedback or questions via Twitter to at twimlai or at Sam Charrington or leave a comment right on the show notes page. Thanks once again to NVIDIA for their support of this series. To learn more about what they're doing at NIPS,
[0:54:35] Visit twiml.ai.com slash NVIDIA. And of course, thanks once again to you for listening and catch you next time.