[0:00:10] All right, let's start a little bit more pragmatic, but then we'll branch out. So one of the things I think a lot of folks here are interested in is, based off the APIs that very large models will create, what are the real business opportunities? What are the ways to look forward? And then how, given the APIs will be available to multiple players, how do you create distinctive businesses on them?
[0:00:37] So I think so far we've been in the realm where it's, you know, you can do like an incredible copywriting business or you can do like a sort of like education service or whatever. But I don't think we've yet seen the kind of like people go after the like, you know, trillion dollar like take on Google's. And I think that's about to happen. Like maybe it'll be successful. Maybe Google will do it themselves. But like I would guess that with the quality of language models we'll see in the coming years,
[0:01:07] There will be a serious challenge to Google for the first time for a search product. And I think people are really starting to think about how do the fundamental things change? And that's going to be really powerful. I think that a human-level chatbot interface that actually works this time around, I think many of these trends that we all made fun of were just too early. The chatbot thing was good, it was just too early.
[0:01:34] Now it can work, and I think having new medical services that are done through that, where you get great advice, or new education services, these are going to be very large companies. I think we'll get multimodal models in not that much longer, and that'll open up new things. I think people are doing amazing work with agents that can use computers to do things for you, use programs, and this idea of a language interface.
[0:02:02] Where you say in natural language what you want in this kind of dialogue back and forth. You can iterate and refine it and the computer just does it for you. You see some of this with Dolly and Copilot in very early ways. But I think this is gonna be a massive trend and very large businesses will get built with this as the interface and more generally that these very powerful models will be one of the genuine new technological platforms which we haven't really had since mobile.
[0:02:32] And there's always like an explosion of new companies right after. So that'll be cool. And what do you think the key things are, given that the large language model we provided as an API service, what are the things that you think that folks who are thinking about these kind of AI businesses should think about as how do you create an enduring differentiated business? So, you know, I think there will be a small handful of like fundamental...
[0:02:58] Large models out there that other people build on. But right now what happens is, you know, company makes large language model, API, other people build on top of it. And I think there will be a middle layer that becomes really important where I'm like skeptical of all of the startups that are trying to sort of train their own models. I don't think that's going to keep going. But what I think will happen is there will be a whole new set of startups that take an existing very large model of the future and tune it.
[0:03:25] Which is not just fine tuning, like all the things you can do. I think there will be a lot of access provided to create the model for medicine or using a computer or like the kind of like friend or whatever. And then those companies will create a lot of enduring value because they will have like a special version of, they won't have to have created the base model, but they will have created something they can use just for themselves or share with others that has this unique data flywheel going.
[0:03:53] That sort of improves over time and all of that. So I think there will be a lot of value created in that middle layer. And what do you think some of the most surprising ones will be? It's a little bit like, for example, a surprise from a couple years ago, and we talked a little bit to Kevin Scott about this this morning as we opened up, which is train on the internet, do code. So what do you think some of the surprises will be of you didn't realize it reached that far?
[0:04:23] It's like systemic mistaken thinking people are making right now is they're like, all right, you know, maybe I was skeptical, but this language model thing is really going to work. And sure, like images, video too, but it's not going to be generating net new knowledge for humanity. It's just going to like do what other people have done. And, you know, that's still great. That's still like brings the marginal cost of intelligence very low, but it's not.
[0:04:44] It's not going to go like create fundamentally new, it's not going to go cure cancer. It's not going to add to the sum total of human scientific knowledge. And that is what I think will turn out to be wrong that most surprises the current experts in the field. Yep. So let's go to science then as the next thing. What are some of the things, whether it's building on the APIs, use of APIs by scientists, what are some of the places where science will get accelerated and how? So I think there's...
[0:05:10] Two things happening now, and then a bigger third one later. One is there are these science-dedicated products, whatever, like AlphaFold, and those are adding huge amounts of value, and you're seeing this way more and way more. I think if I had time to do something else, I would be so excited to go after a bio company right now. I think you can just do amazing things there. Anyway.
[0:05:39] But there's another thing that's happening, which is tools that just make us all much more productive, that help us think of new research directions, that write a bunch of our code so we can be twice as productive. And that impact on the net output of one engineer or scientist, I think, will be the surprising way that AI contributes to science that is outside of the obvious models. But even just seeing now what I think these tools are capable of doing, Copilot is an example. There's much cooler stuff than that.
[0:06:09] That will be a significant change to the way that technological development, scientific development happens. So those are the two that I think are huge now and lead to just an acceleration of progress. But then the big thing that I think people are starting to explore is, I hesitate to use this word because I think there's one way it's used which is fine and one that is more scary, but AI that can start to
[0:06:40] Be an AI scientist and self-improve. And so can we automate our own jobs as AI developers very first? The very first thing we do. Can that help us solve the really hard alignment problems that we don't know how to solve? That, honestly, I think is how it's going to happen. The scary version of self-improvement, like the one from the science fiction books, is editing your own code and changing your optimization algorithm and whatever else.
[0:07:10] But there's a less scary version of self-improvement, which is kind of what humans do, which is if we try to go off and discover new science, we come up with explanations, we test them, we think. Whatever process we do that is special to humans, teaching AI to do that, I'm very excited to see what that does for the total. I'm a big believer that the only real driver of...
[0:07:39] Human progress and economic growth over the long term is the structure, the societal structure that enables scientific progress and then scientific progress itself. And I think we're going to make a lot more of that. Well, especially science that's deployed in technology. Say a little bit about how what I think probably most people understand what the alignment problem is, but it's probably worth four sentences on the alignment problem.
[0:08:04] The alignment problem is we're going to make this incredibly powerful system and it would be really bad if it doesn't do what we want or if it has goals that are either in conflict with ours, many sci-fi movies about what happens there, or goals where it just doesn't care about us that much. The alignment problem is how do we build AGI that does what is in the best interest of humanity? How do we make sure that humanity
[0:08:33] It gets to determine the future of humanity. And how do we avoid both accidental misuse, where something goes wrong that we didn't intend, intentional misuse, where a bad person is using an AGI for great harm, even if that's what that person wants, and then the kind of inner alignment problems where what if this thing just becomes a creature that views us as a threat. The way that I think the self-improving systems help us is not necessarily by the nature of self-improving,
[0:09:02] We have some ideas about how to solve the alignment problem at small scale, and we've been able to align OpenAI's biggest models better than we thought we would at this point, so that's good. We have some ideas about what to do next, but we cannot honestly look anyone in the eye and say, we see out 100 years how we're going to solve this problem. But once the AI is good enough that we can ask it to, hey, can you help us do alignment research, I think that's going to be a new tool in the toolbox. For example, one of the conversations you and I had is,
[0:09:32] Could we tell the agent, don't be racist? As opposed to trying to figure out all the different things where the weird correlative data that exists on all of the training settings and everything else may lead to racist outcomes, it could actually, in fact, do a self-cleansing. Totally. Once the model gets smart enough that it really understands what racism looks like and how complex that is, you can say, don't be racist. Yeah, exactly.
[0:10:02] What do you think are the kind of moonshots that, in terms of evolution of the next couple years, that people should be looking out for? In terms of evolution of where AI will go? Yeah. I'll start with the higher certainty things. I think language models are going to go just much, much further than people think. And we're very excited to see what happens there.
[0:10:35] I think it's like what a lot of people say about running out of compute, running out of data. That's all true, but I think there's so much algorithmic progress to come that we're going to have a very exciting time. Another thing is I think we will get true multimodal models working. Not just text and images, but every modality you'd like in one model able to easily fluidly move between things.
[0:11:04] I think we will have models that continuously learn. So like right now, if you use GPT, whatever, it's sort of like stuck in time that it was trained. And the more you use it, it doesn't get any better and all of that. I think we'll get that changed. So very excited about all of that. And if you just think about what that alone is going to unlock and the sort of applications people will be able to build with that, that would be like a huge victory for all of us and just like a...
[0:11:31] Like a massive step forward and a genuine technological revolution, if that were all that happened. But I think we're likely to keep making research progress into new paradigms as well. We've been pleasantly surprised on the upside about what seems to be happening. And I think all these questions about new knowledge generation, how do we really advance humanity? I think there will be systems that can help us with that.
[0:11:59] So one thing I think would be useful to share, because folks don't realize that you're actually making these strong predictions from a fairly critical point of view, not just a, you know, we can take that hill. Say a little bit about some of the areas that you think are current, kind of illusionally talked about, like, for example, AI and fusion. Oh, yeah. So, like, one of the unfortunate things that's happened is, you know, AI has become, like, the mega buzzword.
[0:12:29] Which is usually a really bad sign. I hope it doesn't mean the field is about to fall apart. But historically, that's a very bad sign for new startup creation or whatever if everybody is like, I'm this with AI. And that's definitely happening now. So a lot of the things we were talking about, there are all these people saying, I'm doing these RL models for Fusion or whatever. And as far as we can tell, they're all much worse than what smart physicists have figured out.
[0:12:57] I think it is just an area where people are going to say everything is now this plus AI. Many things will be true. I do think this will be like the biggest technological platform of the generation. But I think it's like we like to make predictions where we can be on the frontier, understand predictably what the scaling laws look like, or already have done the research where we can say, all right, this new thing is going to work and make predictions out from that way. And that's sort of like how we try to run OpenAI.
[0:13:27] Which is do the next thing in front of us when we have high confidence and take 10% of the company to just totally go off and explore. Which has led to huge wins and there will be, oh I feel bad to say this, I doubt we'll still be using the transformers in five years. I hope we're not, I hope we find something way better. But the transform has obviously been remarkable. So I think it's important to always look for where am I gonna find the next totally new paradigm.
[0:13:58] But I think that's the way to make predictions. Don't pay attention to the AI for everything. Can I see something working and can I see how it predictably gets better? And then of course leave room open for the, you can't plan the greatness, but sometimes the research breakthrough happens. So I'm gonna ask two more questions and then open it up, because I want to make sure that people have a chance to do this broader discussion, although I'm trying to
[0:14:27] Paint the broad picture so you can get the crazy-ass questions as part of this. What do you think is going to happen vis-a-vis the application of AI to these very important systems, like, for example, financial markets? Because the very natural thing would be to say, well, let's do a high-frequency quant trading system on top of this and other kinds of things.
[0:14:53] Is it just gonna be a neutral arms race? What's your thought? It's almost like the Life 3.0, Omega's point of view. I think it is gonna just seep in everywhere. My basic model of the next decade is that the cost of intelligence, the marginal cost of intelligence and the marginal cost of energy are gonna trend rapidly towards zero, like surprisingly far. And those, I think, are two of
[0:15:23] The major inputs into the cost of everything else, except the cost of things we want to be expensive, the status goods, whatever. And I think you have to assume that's going to touch almost everything, because these seismic shifts that happen when the whole cost structure of society changes, which happened many times before, the temptation is always to underestimate those. So I wouldn't make a high confidence prediction about anything that doesn't.
[0:15:51] The intersection of the two curves.
[0:16:21] The thing got 100 times cheaper in the cost of energy, 100 million times cheaper in the cost of intelligence, and I was still willing to spend 1,000 times more in today's dollars. What happens then? Then last of the buzzword bingo part of the future questions, Metaverse and AI. What do you see coming in this? I think they're both independently cool things. It's not totally clear to me.
[0:16:52] Other than how AI will impact all computing. Obviously computing, simulation environments, agents, possibly entertainment, certainly education, like an AI tutor and so forth. Those would be baseline, but the question is, is there anything that's occurred to you that's... I would bet that the metaverse turns out...
[0:17:17] In the upside case, which I think has a reasonable chance of happening, in the upside case, the metaverse turns out to be more like something on the order of the iPhone, like a new container for software and a new computer interaction thing. And AI turns out to be something on the order of a legitimate technological revolution. And so I think it's more like how the metaverse is going to fit into this new world of AI than AI fit into the metaverse.
[0:17:45] Low confidence. TBD. All right. Questions? Hey there. How do you see technologies, foundational technologies like GPT-3 affecting the pace of life science research specifically? You can group in medical research there and sort of just quickening the iteration cycles. And then what do you see as the rate limiter in life science research and sort of where we won't be able to get past because...
[0:18:18] They're just laws of nature or something like that. So I think the currently available models are kind of not good enough to have made a big impact on the field. At least that's what most life sciences researchers have told me. They've all looked at it, and they're like, it's a little helpful in some cases. There's been some promising work in genomics, but stuff on a benchtop hasn't really impacted it. I think that's going to change, and I think...
[0:18:45] This is one of these areas where there will be these new hundred billion to trillion dollar companies started, and those areas are rare. But when you can really change the way that, if you can really make a future pharma company that is just hundreds of times better than what's out there today, that's gonna be really different. As you mentioned, there still will be the rate limit of bio has to run its own thing, and human trials take however long they take. So I think an interesting cut of this is where can you avoid that?
[0:19:14] The synthetic bio companies that I've seen that have been most interesting are the ones that find a way to make the cycle time super fast. And that benefits an AI that's giving you a lot of good ideas, but you've still got to test them, which is where things are right now. I'm a huge believer for startups that the thing you want is low costs and fast cycle times. And if you have those, you can then compete as a startup against the big incumbents. And so I wouldn't go pick cardiac disease as my first thing to go after right now.
[0:19:43] With this kind of new kind of company. But using bio to manufacture something, that sounds great. I think the other thing is the simulators are still so bad. And if I were a bio meets AI startup, I would certainly try to work on that somehow. When do you think the AI tech will help create itself? It's almost like the self-improvement will help make the simulators significantly better. People are working on that now.
[0:20:13] I don't know quite how it's going, but very smart people are very optimistic about that. Other questions? And I can keep going on questions. I just want to make sure you guys had a chance of this. Here, yes. Great. Mike is coming. Awesome. Thank you. I was curious, what aspects of life do you think won't be changed by AI? Sort of all of the...
[0:20:45] Deep biological things. I think we will still really care about interaction with other people. We'll still have fun. The reward systems of our brain are still going to work the same way. We're still going to have the same drives to create new things and compete for silly status and form families and whatever. I think the stuff that people cared about 50,000 years ago is more likely to be the stuff that people care about
[0:21:16] 100 years from now than 100 years ago. As an amplifier on that before we get to whatever the next question is, what do you think are the best utopian science fiction universes so far? Good question. Star Trek is pretty good, honestly. I do like all of the ones that are sort of like, we turn our focus to exploring and understanding the universe as much as we can.
[0:21:52] This is not a utopian one. Maybe. I think the last question is like an incredible short story. That came to mind. I was expecting you to say Ian Banks on the culture. Those are great. I think science fiction is like, there's not like one sci-fi universe that I could point to and say I think all of this is great. But like the collective optimistic corner of sci-fi, which is like a smallish corner.
[0:22:24] I'm excited about it. Actually, I took a few days off to write a sci-fi story, and I had so much fun doing it, just about the optimistic case of AGI that it made me want to go read a bunch more. So I'm looking for recommendations of more to read now, like the sort of less-known stuff, if you have anything. I will get you some recommendations.
[0:22:49] So in a similar vein, one of my favorite sci-fi books is called Childhood's End by Arthur Clarke from like the 60s, I think. And I guess the one sentence summary is aliens come to the earth to try to save us and they just take our kids and leave everything else. So, you know. They're slightly more optimistic than that, but yes. I mean, this ascension into the overmind is meant to be more utopian, but yes. Okay.
[0:23:18] You may not read it that way, but yes. Well, so in our current universe, our current situation, a lot of people think about family building and fertility. Different people have different ways of approaching this, but from where you stand, what do you see as the most promising solutions? It might not be a technological solution, but I'm curious what you think.
[0:23:45] Other than everyone having 10 kids, you know? Like, how do we... Of everyone having 10 kids? Yeah. How do you populate? How do you, like, how do you see family building coexisting with, you know, AGI and high tech? This is, like, a question that comes up at OpenAI a lot. Like, how do I think about, you know, how should one think about having kids? There's, I think, no consensus answer to this. There are people who say, yeah, I'm not, I was gonna, I thought, I always thought I was gonna have kids and now I'm not going to because of AGI.
[0:24:15] There's just, for all the obvious reasons, and I think some less obvious ones, there's people who say, well, it's going to be the only thing for me to do in 15, 20 years, so of course I'm going to have a big family. That's what I'm going to spend my time doing. I'll just raise great kids. And I think that's what will bring me fulfillment. I think, as always, it is a personal decision. I get very depressed when people are like, I'm not having kids because of AGI. The EA community is like, I'm not doing that because they're all going to die. The kind of techno...
[0:24:42] Optimists are like, well, it's just like I want to merge into the AGI and go off exploring the universe, and it's going to be so wonderful, and I want total freedom. But I think all of those I find quite depressing. I think having a lot of kids is great. I want to do that now more than I did, even more than I did when I was younger, and I'm excited for it.
[0:25:07] What do you think will be the way that most users interact with foundation models in five years? Do you think there'll be a number of verticalized AI startups that essentially have adapted and fine-tuned foundation models to an industry? Or do you think prompt engineering will be something many organizations have as an in-house function? I don't think we'll still be doing prompt engineering in five years. I think it'll just be like you, and this will be integrated everywhere, but you will just like, you know, either with text or voice, depending on the context, you will just like...
[0:25:36] Interface in language and get the computer to do whatever you want. And that will apply to generate an image, where maybe we still do a little bit of prompt engineering, but it's kind of just going to get it to go off and do this research for me and do this complicated thing, or just be my therapist and help me figure out how to make my life better, or go use my computer for me and do this thing, or any number of other things. But I think the fundamental interface will be natural language.
[0:26:04] Let me actually push on that a little bit before we get to the next question, which is, I mean, to some degree, just like we have a wide range of human talents right now, and taking a look, for example, at DALI, when you have like a great visual thinker, they can get a lot more out of DALI because they know how to think more, they know how to iterate the loop through the test. Don't you think that will be a general truth about most of these things? So it isn't that while it'll be natural language is the way you're doing it, it will be,
[0:26:33] There will be almost an evolving set of human talents about going that extra mile. 100%. I just hope it's not like figuring out how to hack the prompt by adding one magic word to the end that changes everything else.
[0:26:47] What will matter is the quality of ideas and the understanding of what you want. So the artist will still do the best with image generation, but not because they figured out to add this one magic word at the end of it, because they were just able to articulate it with a creative eye that I don't have, certainly. And what they have is a vision and how they're visual thinking and iterating through it. Obviously, it'll be that word or prompt now, but it'll iterate to better. All right. At least we have a question here.
[0:27:19] Hey, thanks so much. I think the term AGI is used, thrown around a lot, and sometimes I've noticed in my own discussions that sources of confusion just come from people having different definitions of AGI, and so it can kind of be the magic box where everyone just kind of projects their ideas onto it. And I just want to get a sense from you, how do you think, you know,
[0:27:42] How would you define AGI and how do you think you'll know? Yeah, I should have defined that earlier. It's a great point. I think there's a lot of valid definitions to this, but for me, AGI is basically the equivalent of a median human that you could hire as a co-worker. And then they could, say, do anything that you'd be happy with a remote co-worker doing, like just behind a computer, which includes, you know.
[0:28:10] Learning how to go be a doctor, learning how to go be a very competent coder. There's a lot of stuff that a median human is capable of getting good at. And I think one of the skills of an AGI is not any particular milestone, but the meta skill of learning to figure things out and that it can go decide to get good at whatever you need. So for me, that's kind of like AGI. And then super intelligence is when it's smarter than all of humanity put together. Thanks.
[0:28:39] What would you say in the next 20, 30 years are some of the main societal issues that will arise as AI continues to grow and what can we do today to mitigate those issues? Obviously the economic impacts are huge and I think if it is as divergent as I think it could be for some people doing incredibly well and others not, I think society just won't tolerate it this time. And so figuring out
[0:29:09] When we're gonna disrupt so much of economic activity, and even if it's not all disrupted by 20 or 30 years from now, I think it'll be clear that it's all going to be. What is the new social contract? My guess is that the things that we'll have to figure out are how we think about fairly distributing wealth, access to AGI systems, which will be kind of the commodity of the realm, and governance, like how we collectively decide
[0:29:38] What they can do, what they don't do, things like that. And I think figuring out the answer to those questions is gonna just be huge. I'm optimistic that people will figure out how to spend their time and be very fulfilled. I think people worry about that in a little bit of a silly way. I'm sure what people do will be very different, but we always solve this problem. But I do think the concept of wealth and access and governance, those are all gonna change.
[0:30:08] How we address those will be huge. Actually, one thing, I don't know what level of depth you can share that, but one of the things I love about what OpenAI and you guys are doing is when they think about these questions a lot themselves, and they initiate some research. So you've initiated some research on this stuff. Yeah, so we run the largest UBI experiment in the world. I don't think that is, we have a year and a half, a year and a quarter left in a five-year project.
[0:30:38] I don't think that's the only solution, but I think it's a great thing to be doing. And I think we should have 10 more things like that that we try. We also try different ways to get input from a lot of the groups that we think will be most affected and see how we can do that early in the cycle. We've explored more recently how this technology can be used for reskilling people that are going to be impacted early.
[0:31:04] We'll try to do a lot more stuff like that too. Yeah, so the organization is actually, in fact, these are great questions, addressing them and actually doing a bunch of interesting research on it. So next question. Hi, so creativity came up today in several of the panels, and it seems to me that the way it's being used, you have tools for human creators and going to expand human creativity. So where do you think the line is between these tools to allow a creator to be more productive?
[0:31:34] And artificial creativity, segmentation creativity itself. I think, and I think we're seeing this now, that tools for creatives, that is going to be the great application of AI in the short term. People love it. It's really helpful. And I think it is, at least in what we're seeing so far, not replacing. It is mostly enhancing. It's replacing in some cases, but...
[0:31:59] For the majority of the kind of work that people in these fields want to be doing, it's enhancing. And I think we'll see that trend continue for a long time. Eventually, yeah, it probably is just like, you know, we look at 100 years, okay, it can do the whole creative job. I think it's interesting that if you asked people 10 years ago about how AI was going to have an impact, with a lot of confidence from most people, you would have heard, you know, first it's going to come for the blue collar jobs.
[0:32:27] Work in the factories, truck drivers, whatever. Then it will come for the kind of like the low skill white collar jobs. Then the very high skill, like really high IQ white collar jobs like a programmer or whatever. And then very last of all, and maybe never, it's gonna take the creative jobs. And it's really gone exactly, and it's going exactly the other direction. And I think this, there's an interesting reminder in here generally about how hard predictions are, but more specifically about
[0:32:55] We're not always very aware, maybe even ourselves, of what skills are hard and easy, like what uses most of our brain and what doesn't, or how difficult bodies are to control or make or whatever. We have one more question over here. Hey, thanks for being here. So you mentioned that you would be skeptical of any startup trying to train their own language model and would love to understand more. So what I have heard, and which might be wrong, is that large language models depend on data and compute.
[0:33:24] And any startup can access the same amount of data because it's just like internet data. And compute, like different companies might have different amount of compute, but I guess if they're big players, they have the same amount of compute. So how could a large language model startup differentiate from another? How would the startup differentiate from another? How would one large language model startup differentiate from another? I think it'll be this middle layer. I think in some sense, the startups will train their own models, just not from the beginning. They will take like, you know,
[0:33:54] Base models that are hugely trained with a gigantic amount of compute and data. And then they will train on top of those to create the model for each vertical. So in some sense, they are training their own models, just not from scratch. But they're doing the 1% of training that really matters for whatever this use case is going to be. Those startups, I think, they will be hugely successful and very differentiated startups there. But that'll be about the kind of like...
[0:34:22] Data flywheel that the startup is able to do, the kind of like all of the pieces on top and below, like this could include prompt engineering for a while or whatever, the sort of the kind of like core base model. I think that's just going to get too complex and too expensive, and the world also just doesn't make enough chips. So Sam has a work thing he needs to get to.
[0:34:48] And as you probably can tell with the very far-ranging thing, Sam always expands my boundaries. And a little bit unlike when you're feeling depressed, whether it's kids and anything else, you're the person I always turn to for opportunities. I appreciate that. Yes. So anyway. I think no one knows. We're sitting on this precipice of AI. And people are like, it's either going to be really great or really terrible.
[0:35:15] You may as well, you gotta plan for the worst. It's not a strategy to say it's all gonna be okay. But you may as well emotionally feel like we're gonna get to the great future and work as hard as you can to get there and play for it. Rather than act from this place of fear and despair all the time. Because if we acted from a place of fear and paranoia, we would not be where we are today. So, let's thank Sam for spending dinner with us. Thank you.