AI Throwback to 2021 with OpenAI and Greg Brockman

Welcome back to the Good Morning AI podcast. I have a treat for you today. We are going to be going back in time. A few years ago, I had a chance to interview Greg Brockman, uh, in the early days of ChatGPT, uh, and so I am so excited to bring this to you. But just remember, as you're listening to it, it is not current. It is from a couple years ago, but when I found it and I relistened to it, I just thought it was really cool, so I wanted to share it with you. Enjoy. Hey everyone, welcome in. So great to see you all. Thank you for joining us. Um, I am Noelle Russell Russell, really excited to talk to you. I am the CEO and founder of the AI Leadership Institute. I also am an AI/ML Specialist at Red Hat. And I am here really because I kind of started probably similar to you just about 7 years ago, very interested in artificial intelligence, very interested in conversational AI, and a little-known product called Alexa was being born. And so I jumped over. I was an early member of that team. I built over 100 apps for Alexa, got a million users of this app, and and felt like this virility that comes when you do something really interesting and new. And then went to Microsoft and did some of the same things with cognitive services. And I just got a heart for democratized AI. And I'm like, how do we really do this? And I think one of the best examples in the world, when I got an opportunity to join Greg Brockman on this stage, I was like, please let me talk to him. Because it's so in alignment with what I think the world's looking for, low-code, no-code. How do we get more voices into this process of development? I just think there's so many cool things that we could talk about today. Welcome everyone, and specifically welcome Greg. So glad you're here. Thank you for taking the time because I know your schedule is probably crazy. But thank you so much for being here. I thought we'd get started right away. I'd love, of course, for you, for those of us who don't know you or don't know the work you're doing at OpenAI. Can you give us a little rundown of what OpenAI is, who it is, what you're doing, as well as maybe a little bit about your backstory and how you got involved?

Sure thing. We started OpenAI about 6 years ago actually. Wow, how the time flies. When we started, I think it was really with this picture that deep learning was really starting to work, right? That AI, you know, is this sort of field of broken promises for so many decades. And suddenly things were suddenly starting to happen in a way that felt like they were really going to affect and reshape society. And that, you know, I think that when we started, we really just wanted to be a part of that, right? Really wanted to help shape the outcome to maximize the positive and to minimize the negative. Right. And I think that we take an eye to the long-term picture of existential risk, but, you know, also the whole transition along the way. Right. Like, we're going to have these systems start to integrate with the society. And I think if there's one thing we've learned over the past 20 years of, of internet technology is that you can have all sorts of very positive, but also very surprising and sometimes negative effects if you're not careful. And so I think that that really was the ethos that we came in with is that we want to build technology. We want to be able to really push forward what the tech can do, but we want to be mindful and we want to make sure that we're really being careful about the effects and taking responsibility for them. So that was kind of the big picture we had 6 years ago. And I think we had a particular picture of we want to advance this technology and to really do it on behalf of all of humanity and to benefit everyone. And figuring out exactly what that meant has taken us, took us quite some time. I think that, you know, first 4 or 5 years, I think that we were really trying, you know, a number of different ways of realizing that vision. I think that we've kind of settled on a particular path that, that seems, seems right. And maybe we'll find in the future that it's not quite right, we'll need to update. Um, but so what we are is, so we're not a, we're not a nonprofit, we're not a for-profit, we're what we call a capped profit, um, which means that we have investors, we have employees who have equity, um, and that, uh, the, the picture is that we owe them a fixed one-time return, and once we exceed that, then everything beyond that, uh, is, is to, to benefit the world. Um, so, you know, we kind of sort of have, have ended up with this sort of hybrid structure in order to match the the technology that we're developing. And then on the technology front, that, you know, I think at this point we've, we've had some, some very field-leading results in terms of systems like GPT-3, which is a language model that many of you have probably heard of. Recently we released a system called OpenAI Codex, which is a system that does natural language to code and is really super cool for me as a programmer to be able to see this tool that sort of removes all the drudge work of creating code. And we have many people who are using it during their coding process in their editor and many people who are building applications on top of it. And I'm sure we're going to talk about some of those today.

We are. Awesome. Well, tell us a little bit about your, like, kind of your personal journey. What brought you to OpenAI? Like, what's that, like, napkin at a bar kind of story?

Yeah. Well, you know, for me, it really started When I was a kid, at one point I read Alan Turing's 1950 paper called Computing Machinery and Intelligence, which is the Turing test paper. Uh, and it's a very surprising paper because, you know, he first lays out the actual Turing test and then he says, okay, how are we going to solve this? And he's like, there's no way we're ever going to program a solution. It's just way too complex, way too hard. Instead, the way that we're going to solve it is we're going to have to build an a machine that learns just like a human child, and we're going to teach it through, you know, give it some rewards when it does things that are good and some punishments when it does things that are, that are negative, and that eventually it'll learn in order to, to speak fluently and be able to pass this test. And to me, the thing that really inspired me was this idea that you could build a machine that would understand things that I could not, right? You know, I dabbled a little bit with programming and realized that programming is mostly about, you know, it's kind of two things. Like, one is you got to deeply understand a domain and think it through and understand all the little pieces, and then you write out all the rules for it so mechanically. And the idea that you could have a machine that could take care of one or both of those pieces for you was just the most inspiring thing, right? You know, it's like if you're, you know, if you, if you wanna be kind of like Tom Sawyer and just be lazy and get everyone else to do the work for you, well, how do you do it? Um, and here was a path. Um, and so when I showed up in college, yeah. So when I, when I showed up in college, I was so excited to work on AI and I went to a professor and he showed me like parse trees and whatever. And I was like, this is never gonna do it. So I got sucked into the startup world and did that for some time. And it was really watching deep learning from the outside. I really didn't know anything about it. I just saw that on Hacker News, I felt like every day there was some new deep learning for this story. And so I went to one of my friends and was like, what is deep learning exactly? And the more I talked to people, the more I realized that all my smartest friends from college were now in the field. And so this was, this was about 2014. And I realized that I think now is the time to really join the field and see if I can have an impact. And the funny thing is, for me, I thought that I was just going to be kind of a manager and just really spend my time trying to enable people who had the PhDs and, you know, who were the real experts to be able to spend their time doing research. And the thing that was a surprise to me was finding that actually my software skills were the thing that was the limiting factor to us making progress. And so I basically spend my time heads down building systems. And it turns out that that is a lot of what is the wall between us and future progress in this field.

Yes. Oh my gosh. I am definitely a believer in learning by doing and that there's no, nothing like fingers on a keyboard to learn the lessons. And speaking of which, one of the things I was so curious about when I first found out that you were going to be speaking at this event was around GPT-3. I know a lot of us who have heard about it, or maybe some of you saw an earlier session about it and it came up in the conversation, but like what an interesting time for your organization. Cause I feel like it kind of went viral a little bit. Can you tell us a little bit about that journey? What it was like, you know, behind the curtain, like inside your organization, like how did that go?

Yeah, so, well, the funny thing, it wasn't the first time that we went viral for an AI system. So, you know, we've had some high-profile results, but before GPT-3, actually, I think probably maybe one that a lot of people saw was GPT-2, where it wasn't the first time we built a system that felt like, you know, it itself wasn't necessarily one that had dangerous applications. It was kind of hard to tell,, but it felt like the trajectory was getting there, right? You know, we'd gone from systems that could like, you know, write a coherent sentence to suddenly could write coherent paragraphs and write this cute little story about, you know, these unicorns that had been discovered that spoke English. And it was like, you know, people read that and was like, ah, it felt like a human might've been able to write this. And we weren't sure what to do. And so what we decided was that we were going to not release the model. We were going to kind of release samples and see how people reacted. And we released a smaller version of the model. And this went super viral. I think that the headlines were basically like, AI system is too dangerous to release. And yeah, probably people, people remember that. That was, that was from a Guardian article. And you know, there's a lot of incentives in the system that the journalists don't actually write the headlines and that the editors are just looking for the clickbait and all those things. But at the end of the day, you know, we felt like we were, like we were responsible for, for the whole, the whole, the whole rollout, right? And the whole, the whole reaction to it. And I think that with, with that, what we really wanted was that we really wanted to say like, hey, these systems, we're not sure yet how to deal with them, but they are going to start start having real implications. And again, it's all about maximize the positive, mitigate the negative. So with GPT-3, the way that we did it is I think that we came up with a better solution, which is we released it through an API. So we had this model that was able to write, you know, really like the, the, the, the coherency and the kinds of tasks you could do with it were just scaled up by order of magnitude, two orders of magnitude from what we saw with GPT-2. Um, and we released it through this API so anyone could build on top of it. But that we were able to also be responsible for the use cases, right? And that the way that we run it is actually very similar to my previous company, Stripe, which is a payments processor where we review applications so that, you know, we see what people are doing. We work super closely with developers and you really get to see the impact of your use cases as they scale up and touch reality.

Yeah, I think it's so incredible because of how that worked, right? And it was a closed beta, right? So people were like, at least in my world, clamoring to get in. And then I actually wasn't sure when people were talking about use cases, I'm like, did you really build that? Like, are you really in the beta? Like, show me your invite.

Uh-huh, yep, yep.

So I would love to talk a little bit about, you know, how did you see people using, and it could be any of your models, of course GPT-3 has a lot of, you know, but like Cortex even, I just see some huge opportunity. But what are some fun and maybe even scary Blackberry, I don't know, fun examples of how you've seen your technology being used, something that, you know, we wouldn't have read in The Guardian or another article. Yeah.

This is the thing about GPT-3 that I think really got people's imagination was that what can you use it for? And the answer is anything, right? And when we were building the GPT-3 API internally, we really, you know, on the one hand, it's great to have this very general purpose model. On the other hand, you yourself don't know what it's gonna be used for and don't know what it's good at. And even exploring those use cases was a little bit outside of what we're able to do. We can scratch the surface, but going deep, that's where we really need the community. And so, you know, one, one that I actually really like is a company called Augrented who uses it for parsing through and simplifying notices from landlords to tenants. And so, you know, tenants who don't have access to legal counsel, you get this like super like complicated legal document. You're not even sure what you're supposed to do. You know, you can use GPT-3 and this, this service in order to, to, to get, get a clue and to, to be able to, to have, you know, be on much more fair footing. Um, so I think that, that those kinds of applications— never would have thought of something like that, right? But the, the, the— I think that what we kind of found is that there's just this, like, you know, you think about language is everywhere, right? It's in every business, it's in everything that we do every day. And if you can have an AI system that can help with language, can remove barriers to entry, suddenly you'll just see all of these tasks that were just outside of people's reach suddenly start to become possible. With Codex, one that I'm really excited about. So this is the coding model we released recently. So it's been integrated into a platform called Replit, which is a very popular platform for beginning coders, for people to build. You know, it's like an editor in a browser. And I think that there are a lot of people, especially young people, who are using it these days and that it's used for code explaining. So you can just select some Python code or some JavaScript code and explains to you what it does. And, you know, I think that we're still in the early days of these systems. They're not perfect. We have a lot of improvements to make. But I think it's like by far better than anything else that anyone's ever created at these kinds of applications. And it's really inspiring to see not just using these technologies to write code, but in order to learn how to code. Right. You know, that's not just the, you know, give a person a fish. It's really teaching people how to fish. And I think that's the real promise of AI technology. The whole point is that we want technologies that will help us be better, help us achieve whatever our dreams are, help us solve problems that are just outside of the reach of humanity otherwise. And, you know, we've got a long road ahead of us before we really do that. But I think that we're starting to see, you know, the scratch of the surface in terms of these kinds of problems.

Yeah, absolutely. I actually am very inspired by a lot of this technology. To do just that, right? To not necessarily give, you know, a crutch to people to learn code, but more to enable them to learn faster and dive deeper. And I've seen such— I am one of the early users of the Copilot initiative in GitHub, and the amount of code or the amount of actual features I could build into an app, like, accelerated tremendously, right? Other than me having to, like, type in this minutia, which I really didn't like from the very beginning. And I'm an enterprise Java person. So there's lots of—

Oh man, that's mostly Blair's plate.

But it's so— that's— I truly believe that. I believe if people realize that they can actually build things and deploy them faster when they're enabled through this, and then what happens to a world like right now? I always talk about people with special needs. I have a son with special needs or physical disability, but they are brilliant, right? Their mind is not impacted by their disability and they're software engineers and they're designers. What happens when we use natural language to build the things that they have to build, right? To put it in the— format or the form factor that it has to be in. Yeah, I just think it's so cool. So on that note, I'd love to ask if you have any examples of like AI for good. It's something that I'm super passionate about. I was part of like the AI for Earth and AI for humanitarian efforts at Microsoft, and I'm just curious if there's any, any existing projects that you're aware of or that are already published or public that are kind of trying to solve, change the world, save the world, solve the world's problems. Yep.

Well, you know, so, so I think that there's lots of these projects that are doing, you know, sort of, you know, their, their part, right? That are doing kind of a piece of a big, big puzzle, right? And so, you know, even, you know, to your point of, of people with physical disabilities, we have an app called Serenade who's, who's building on us on Codex these days that is for, for developers to do voice to code in an editor. And it was inspired by the creator of it had really bad RSI and wasn't able to type anymore, which is, you know, a very scary thing. I've had my own bouts of RSI, right? Where you kind of realize that, huh, I'm not sure if I'm gonna be able to keep doing this in the same way. And to be able to have tools that can actually let you continue pursuing your passion, your calling, I think is a really amazing thing. And I think that there's just like a whole host of applications like that. But to zoom out, I think that, you know, the whole point of AI in my mind, it's really about solving, helping humanity solve challenges that are outside of our reach, right? You know, the things like climate change, things like personalized medicine, things like education, like all of those big picture items that I think are just really hard for humanity to really make further progress on. I think that we're starting to have a shot. So I'm really inspired by systems like DeepMind's AlphaFold, uh, which, uh, has, you know, sort of, you know, either solved or made hugely significant progress on the protein folding problem. And that really is an example where, you know, something like GPT-3, I think that a lot of the artifacts you produce from it, they're, they're like things that, that, you know, a human could have produced with effort. Right. And I think the reason that they're interesting, the reason people want to use GPT-3 is because it's maybe hard to get a human in there. You know, it's hard to get a human to be your copilot and sit over your shoulder and correct all your typos in some code. But to actually figure out how to fold a protein, no human can do that for you. And so I think that it starts to give us this new sort of entry into structural biology and to have some hope of being able to, you know, develop medicines that we weren't able to otherwise. And so I think that that's where AI is really going. I think that we're going to start seeing these artifacts that are going to be produced that are exciting of their own right, that were things that humans couldn't have done otherwise. And that's what really keeps me, keeps me going.

Yes. I mean, augmenting human ingenuity, right? Like, how do we, like, push a little bit further? I think that's incredibly— obviously it's very noble, but there's also some, probably some challenges. I'd be very curious. A lot of people know that I'm very interested in like AI and ethics, and I got an award for like my perusal into this space by VentureBeat. And I'm just wondering what you've learned, because I'm sure in the way that you've deployed these models and who's using them and how they're using them, you've had to think differently about like how you're going to maintain, you know, like how are you going to mitigate bias? How are you going to— you know, it was the first thing I thought about with GitHub Copilot or Codex is I was like, wait, do we want to repeat— like, what did we train this on? And do we want to repeat those things? So what were some of your experiences? I mean, less about like, you know, this is how it all turned out, and more like a lot of us are just starting this journey in our companies of having these conversations. So would love to know, you know, how did you guys navigate this?

Yeah, well, I think it's, it's a, it's a kind of the question of the age in a lot of ways, right? You know, I think we're all starting to see technology for its own sake isn't enough, right? You need to develop the tech, but you also need to take responsibility for the downstream implications of it and things like bias, things like fairness, things like, I mean, with, with Codex that's producing code, you want to make sure that code isn't going to do something, you know, malicious and, and bad. So how do we actually make sure these systems are, are aligned with the operator's intentions? I think that's really important. And then secondly, how do we make sure that those intentions themselves are good? That's going to be even a harder problem, and that's not necessarily even a technology problem. And so what I think we have seen is that, you know, like we— one place I actually take inspiration from is there's this book called What Technology Wants that talks about sort of, you know, development of new technologies, not about AI in particular, but it's, I think, very applicable. And that one section that stood out to me was talking about the Amish and that, you know, I think that people think sometimes of the Amish as a society or people who don't accept new technology, but this isn't actually true. I— that, that, you know, there's lots of different sects, but the overarching theme is that, uh, they want to only introduce technology that enhances community, that enhances their existing values. And so how do you actually figure that out for a new piece of technology? Uh, well, the strategy that they've, they've developed is basically one of beta testing, you know. So if a farmer wants to try out some new kind of combine, um, that, you know, that the, the elders can authorize them to try it, and then they can see Does it result in the farmer spending less time with, with the community and, you know, sort of more, more time with the machine? Or does it result in like sort of a tighter-knit community? And if it seems like it's working well, they scale it up. And if it's not, then the rule is that they can cut it off at any time and no questions asked. And I think this kind of approach is really interesting, right? Is that this idea of figuring out how a complex system is going to be affected by the introduction of a new technology or method— it's hard to do in isolation. So you really got to try it. You try it at small scale, you scale it up from there, and you just got to keep your ear to the ground to see what's really happening. And I kind of totally in parallel, that's, that's really the philosophy we've taken with GPT-3. We try out a variety of use cases and we start them small. And so a lot of how we regulate the platform is that we approve people to hit a certain level of processing volume, right? To be able to use the system to a certain level in a certain period of time. And if that goes well, we scale it up. And if it doesn't, then we work with the developer in order to figure out how to make it go well. And I think that that kind of meta approach, I think, is really important. We've got to accept that things are going to go wrong, but that the challenge is not avoiding anything happening. It's making sure when it happens that it's at contained scope and secondly, that you're responding to it and making it better. And so we've seen this in a wide variety of different circumstances, but I think it's been really, really important for us to, you know, make sure even that our own tools are working right, that we build monitoring tools and tools to help our users ensure that they're delivering good things to their users, and that we got to be able to see that in reality, in the wild.

Yeah, it's fascinating. So when you say the developer, you're really talking about like, that's the farmer in the analogy.

Yes, yes, exactly, exactly.

Yes, yes. So interesting. Um, so if— can you talk a little bit more about like what I don't know, I'm trying to figure out how does a developer, or maybe what are your processes today for taking in that feedback? I mean, a lot of— so I mean, I'm in these programs, so I know, but a lot of people here really don't know what happens in these beta programs and how that even works. And I do think it's something we could see modeled potentially in other organizations. So yeah, we'd love to know more.

Yeah. And actually, well, the funny thing for me too is that the kind of the initial iteration of this stuff was based on how we did reviews and worked with customers in my payment processing company's case as well, right? It's kind of very, very similar in a lot of ways. But so developers show up, shows up on our platform, they, when they have an idea for an app they want to build and have it implemented, they submit a production review request. And so, you know, in the earliest days, we would, for every single customer, get on a video call with them. They kind of walk us through what they were doing. And I think actually it was a surprisingly fun experience for the developer. You get to show off everything you're doing to the people who are building the platform you're building on. Um, and I, you know, that, that I think that, that, that's like, I think, you know, kind of the ethos that we're really trying to develop is, look, we're all here exploring this new technology. It's like this alien technology that has landed on Earth, and we're all trying to like poke and prod and figure out what it can be used for. Um, and then I, as, as time goes on, then I think it's, it, it really does depend a little bit on the specific, specific use case. Um, but I think that we work with the developer to figure out what's what are the risks here? Where can things go wrong? And some of it's about bias, some of it's about spitting out incorrect answers. And I think there's a variety of things there. And then we talk about, okay, well, what kind of monitoring, if any, is required for in order to mitigate that? And again, I think that a lot of the ethos that we take is we really try to be collaborative, right? I think that we're all kind of in this together. This whole endeavor of AI is something that I think is going to affect all of us equally. And so we all have a stake. In making sure that that rollout happens in a positive way.

Yeah. And I just think it's interesting because it also works really well for marketing, right? In the way that you've developed this kind of closed beta and people are like hungry, there's scarcity. It's just, at least in my world, there's a lot of people who are like, oh my gosh, if you know how to get in or if you know somebody. So yeah, I think it's really interesting how that's turned out for you all.

I really wish that we were sufficiently competent marketers that I could say that that was our goal.

Yeah. As usually as tech startups often are, and we happen upon it and things go viral, right? Like, again, in hindsight, I think Steve Jobs says this, right? You can put the dots together and it looks like a beautiful tapestry of activity when the reality is it just kind of fell that way. But yeah, I think it was brilliant. And, and I actually, one of the things I would love for us to take some time on, and we might have to back up a little bit because I always dive right in. And get all excited and I make a little bit maybe presumptions. You touched on a couple of the products that you have, but it'd be great to give us a little lay of the land about what maybe OpenAI is doing now, and then we can talk a little bit more about the future. But for those of us, I think there's some people in the audience that aren't as familiar with all of the different things you're doing and that GPT-3 might be where we really understood and started to learn about you. So maybe you could share a little bit about what else is going on in there?

Well, yeah, well, we've been, we've been on an interesting arc over, over our, our 6 years of existence, I guess. You know, we started out with, we built a software project called OpenAI Gym, which a lot of people use in reinforcement learning. And you know that our first machine learning projects, actually, the funny thing was the very first blog post you can find from us on a machine learning result were, were on generative models. And, you know, in the early days they didn't really work that well, to be honest. And it was kind of cool. You know, you're making improvements on the basic technology, but but the actual results themselves really only of interest to the, like, you know, the people who are really in the field and trying to do the same stuff. Um, but the idea of a generative model is that you, you kind of look at a bunch of data and you want to be able to extract the structure and be able to create new objects like that data, which sounds very, very abstract, but that's exactly what GPT-3 is. It's just take all the text on the internet and just predict the next word in it, and then you take this model and you have it predict next word in a document. Next word, you know, that document can be a question, and maybe then it's predicting an answer. And so that's how you actually use this general idea of a generative model to be applied to a specific task. But it turns out that it works not just in text. We've used it in, uh, in, in music generation. And so we have two projects, uh, one called Jukebox, one called MuseNet, uh, which people can use. MuseNet generates MIDI, and actually a lot of composers use it. It's super cool. You can go to our website and this tool is just there and you can generate, you know, all sorts of different songs. Jukebox, I think, is like, it's a little bit more rudimentary in terms of the actual impressiveness of the artifact that comes out, but I think it's pretty, pretty crazy in terms of where it points, where we actually trained on raw audio and are able to generate raw audio. And so you actually start to be able to have the AI generate singing voices and, you know, The hit rate of being able to understand the words isn't 100%. We need, you know, it's a tiny little model and, you know, tiny little amount of data comparative to where we're going to be. But you can really start to see the shape of what's to come. And we have some pretty cool, you know, songs that you can actually understand the lyrics and that are in the, you know, that sound like recognizable voices of, you know, famous singers. So there's kind of that arc. DALL-E is another arc, which is doing the exact same thing on, on image data. And so you actually just take the pixels of an image and you try to predict what the next pixel is. And you can use this to do— and if you— so in the case of DALL-E, we do it with on text-image pairs. And so you can actually then end up with text-to-image. And so we have all sorts of crazy, like, you know, you, you ask the AI to generate a picture of a radish walking a dog wearing a tutu, and it generates some really nice looking, uh, tutu-wearing radishes walking dogs. And to me, that's the crazy thing. It's one piece of technology, right? It's just a deep neural net. That's all it is. All these things are things that you rewind to, you know, 1950s and you go to, you know, Rosenblatt and people building the perceptron and they'd be like, oh yeah, that's basically what we were working on. It's like 70 years of history here and we're still just doing the same thing. But now we've scaled up the computational power, we've scaled up the data. And we've improved the basic underlying algorithms enough to the point that these artifacts are actually exciting and actually new and useful and meaningful. And I think that the story for us is continuing to push that forward. It hasn't hit a wall in 70 years. You know, maybe we've got, you know, maybe we'll hit a wall any, any day now, but that would actually be exciting if it happened because then we'd have something new to work on.

Absolutely. I love it because, and I'm so grateful, like, hi Jerry, and Maria, we've got some great people communicating in the chat and they're talking about one of the things I love about democratizing AI in this way is that the people with the problem now get to get closer to the solution. And I, I love the music scenario because I was actually at the very first hackathon that was ever held at Abbey Road, right in the UK. And it was a music, of course, music hackathon. But we didn't have any generative models. So they're literally trying to in 3 days build stuff from scratch. And it was not a beautiful experience. However, the person who won was like, they won a, they built a rap battle bot, right? That you basically could rap battle against. And I'm thinking now in hindsight, like how awesome that would be with a generative model. And one that actually, like they were doing it with this anyway, it was not a neural-based, like voice solution. Like, it was just, it was like 4 years ago, so it was a bit old school. But how do we, you know, get closer to those, like, to the artists, to the curators at the Metropolitan Museum of Art, to farmers, to someone was talking about waste and what innovations we could do just around, like, landfills and waste management. And yeah, I think part of that's like hackathon-y, right? Like, and that, that's what I feel a lot of times when I'm in these conversations with developers in your, in, you know, in your community that it's almost like an ongoing perpetual hackathon.

Yep. Yep. Yeah. I think this is, this is one of the most exciting trends in AI and how it's rolling out. I think everyone expected AI would be this force that would just take this job, then take that job and then take this other job. And it was just a question of like, what's the order of jobs? You know, probably, you know, AI researcher, obviously that's the last one to go. Cause you know, the AI researchers are the ones making these predictions. But instead, it's kind of taking the drudge work of all jobs at once and automating no jobs in entirety. And, you know, maybe in the future that will change, and maybe there'll be a point where you build a machine, and that's, you know, I think that maybe on the trajectory we're on where it really is able to do the full depth of what a person does in a job. But I think what we've learned is that just people are really, really versatile, right? That we really do a lot. Like, you know, even, you know, think about the, like,, you know, just a worker in a, you know, working in fast food, right? Like it's not just the mechanics of, of creating the food. You gotta talk to a customer. Like, you know, something breaks, you gotta be able to work around it. It's like a really deep set of skills and, and sort of really exercises a lot of, of this like sort of evolutionary knowledge that humans have built up. So I think that, that, that what we've learned is that humans actually are much more capable than, than we gave ourselves credit for. And that, that being said, these tools can be useful in their intermediate form. You know, GPT-3, It's not perfect. I'm not going to want to use this to write all my emails today, but I might want it to help me write emails, especially where it's super annoying for me to do so. And I think that's actually really—

repetitive tasks, right? Like people in HR and people in legal and people in finance, they write the same things over and over again. Like, like I remember when I used— I had a CEO literally get like emotional because they were like, my employees could be happy. Like they actually will be happier doing their job because I'm going to take away 35% of their day where they're doing the minutia, the little things that just take away the joy, right? We call it like mind-numbing activity. And what if we could use AI to do that? I love it. I think it's amazing. So, so we know a little bit about what you're doing. Of course, you know, being where you are and in the role you're in, you're probably— I'm sure you're thinking about the future or already like maybe the future is now. I would love to talk a little bit before we're out of time about the future of OpenAI and some things that you're dreaming about building or where you see yourself in 5 years or 10 years. Obviously, we never know what the future will hold, but it's nice to have a chance to dive into Greg's brain a little bit and see what the future might look like.

I think this decade is going to be a crazy decade. I really do. 2012 was AlexNet, which is kind of the result that kicked off the deep learning wave. And I think it was really this decade of figuring out how to even get the basics in place, right? How to get models to the point that they're even interesting, right? Maybe one of our big results from the last decade was training a Dota bot to beat world champions in Dota, which is this competitive video game. Which is exciting, but how is that useful? Right? Like, you know, I guess maybe it's useful if you're, if you're super into that game. But I think that we're starting to do something that's qualitatively different. We're starting to have these systems that every time we train a new state-of-the-art model, which every year you should expect us to do something that just, you know, sort of obsoletes everything we've done previously. If we fail to do that, then something's gone wrong. And those things themselves have real impact. And so I think for me, the thing that's exciting is sort of transitioning into this this regime where the systems start to really not just be sort of like, you know, I think GPT-3 is used for basically, you know, everything you can imagine, but that the depth of those applications isn't where I want to be yet. Right. You know, I kind of alluded to this earlier that I think there's this really important transition where, you know, why do you want to use a generative model? I think there's 3 different reasons you might want to. One is because it can generate lower latency generations than a person. And so, you know, that's like the Copilot use case. Two is that it might be at bigger scale than people could, and that sounds a lot like web spam and stuff that we don't really want to encourage. And so we try to lean out of those applications. But the third one is because it can solve a problem that people wouldn't be able to, right? They can actually do something that itself, it's like it produces a book that's actually worth reading, right? You produce a movie that someone wants to watch, or that you come up with a, you know, a curriculum for a student that is personalized to them and that actually teaches them things they wouldn't learn otherwise, right? Like, That's what we need, right? That's where we need to go. And I think that, I think we might be on track for it. And so that's the goal for me is to integrate these, you know, we're kind of pushing these different domains. We're making the models better and better. I want to integrate as much of that as possible, have a system that can reason across different domains and that can be used to help humans solve some problems that just otherwise we wouldn't have a hope for.

Yeah, that's amazing. I love it. I think, like you said, I do think we are on, I mean, Grant, I've been saying this every 10 years probably, right? Like, this next 10 years is going to be— and it has been, right? Like, that's the interesting thing is that every 10 years it has been exponential, the kinds of things that I have been able to build in the companies that I've been part of. And so I see the same thing, but the story changes a little bit because now we're, you know, you're lowering the bar to entry. It's not just the top 1% of the 1% of companies that get to play in this field. You're giving the keys, right, to A, people who maybe are closer to the domain and can be more imaginative. I will share that in my early days of Alexa, I was super frustrated because I would give them the keys to build any voice application they wanted and they build a, you know, a command-based skill that gave you banana facts. Like, that's where we were, you know, or space facts or whatever. And the burden was not the tech, right? The tech was capable, similar in your— I think in your world, like the tech can We could do a bunch of stuff, but I had to actually improve the ingenuity of the dev to think bigger, to realize that the thing that they thought they might never build in their lifetime is now accessible to them. So, I'm just, I'm super excited. I agree, but we have a lot of work to do because, again, we have to kind of instill some hope in these developers that this tech can meet you where you are. And we just, like I said, sometimes we get very, very simple solutions, which are great. You mentioned them earlier, like it's good, they solve some problem for some people. But really, there's a huge opportunity over the next 10 years of things that we could solve for people in huge areas like waste management, like healthcare, like, you know, and not healthcare cure diseases, though that's awesome. But even just like my dad, you know, has cognitive issues, he's old, aging, he lives with me. Like, His life is better simply because he can talk to the stuff that he works with. That's all. And if he could talk to more things, his life would be even better. Like, we just got a voice-enabled microwave. Awesome. Amazing. He'll never burn popcorn again. So anyway, I'm super excited about it. Well, we're coming to the end here, and I want to make sure we touch a little bit about not, of course, the future, which is awesome, But a lot of people here probably want to know, like, how do you do more? How do you get started with this? How do you learn more about it? What are the processes that you use, you know, to— how does someone get engaged with OpenAI, or what do you think their next steps would be?

Yeah, well, I would say, you know, to get into AI in general, I think that the approach that I recommend is get started. You know, that I think that actually there's a huge variety of different materials out there. That there's so many different papers, and I think it can be a little bit overwhelming to figure out where that first place to start is. But what I found is once people have figured out the first place to start, you know, the second step, I think it's like, it's always like, you know, it's a very small delta on top of what you've already done. So I think that the hardest part is really figuring out where to start, but from there I think it's actually very, very accessible. And so, you know, fortunately again, I think that there are some really great courses. You know, my approach was actually to start doing Kaggle competitions. So I just kind of picked a random Kaggle competition and just tried, you know, just my best at it and it didn't really work that well. And then I like read a little bit more about how to do, you know, like, you know, convolutional neural net and I grabbed someone's sample code from online and just everything from there was incremental. So I think that learning by doing is a really, really good approach here. And, you know, one thing that I think has been very encouraging to me is to find that, you know, this field, it's so new, right? It's changing so fast that there are no experts. The people have been doing this for, you know, did a PhD and that kind of thing. Like, they actually don't— you know, that there's, there's definitely some things that take some time to develop, but I don't think that they fundamentally have some sort of advantage that you, if you're just starting, do not. I think that it's really about tenacity. It's really about being able to just beat your head against a wall because the neural net isn't training and you don't know why. And just you keep trying everything until it works. And that is just like— that is just how the field works for the experts, how it works for the beginners. And so that actually is the core skill is just the ability to persevere through. So, you know, for OpenAI in particular, you know, we have— we, as we've talked about, have products now and we have systems that we want people to build on top of. So please come use GPT-3, come use Codex and help us push these systems to the next level.

Absolutely. I love it. Yeah, I'm a big believer in learn by doing. It's interesting because when I got into artificial intelligence, I didn't actually think it was I guess they call it weak AI because it was like regex and like there wasn't, in my mind, there wasn't a bunch of automation or anything happening. But I went from 1 and then I built 100 in a very short period of time, right? Like I kept building, kept evolving, and not just evolving the software, though it did evolve. It was really just my understanding increased every time I built another thing. And so, and I didn't know anything when I, when I started building, I learned Node.js. I mean, granted, I knew Java, so I had that going for me, but I didn't know Node, I didn't know serverless, I didn't know infrastructure, I didn't know AWS that well at that point. And I had— I just learned by building. And the interesting thing about that was that my unique perspective, the fact that I like mindfulness and kindness and family and I care about special needs, drove me to build solutions that even though you had mentioned this, even though it wasn't amazing, It met a need that no one else was really empathetic towards. And that's what I think OpenAI gives us an example of. It's like, what happens when you give an API to someone that can allow them to do something they— I would never have built those skills if I had to build the whole thing from scratch. Like, I'm not going to build an NLU system from scratch. I never even thought I would be in AI, let alone building those. But because of the API access, I could then build things that were close to a domain I cared about. And that allowed me to be much more successful. So I 100% agree. Everyone in here, if you have not built something, I mean, there's so many of these APIs out there to give you an example. What you start with today won't be what you end with. So just start with something. There's APIs on every major cloud provider. Of course, there's products available at OpenAI. Like, just get started, I think. So that's a great point. And really, I think maybe we're just, oh my gosh, at the end, I was like, get off. It's almost over. So of course, I want to thank everybody who has mentioned in the chat. Hi everyone, so good to see you all. Thank you for joining us and being communicative there. And if I just want to give you one last chance, Greg, is there anything else you want to share with anybody before we let them go on to the rest of their day?

I mean, this is a great conversation. Thank you everyone for coming. And I think the future is going to be a very, very exciting place and hope everyone here is a part of shaping that.