[0:00:00] So I wanted to introduce Sam really quickly. Sam is probably known to everybody here. He's a serial founder. He was the president of YC. He's the CEO of OpenAI. He's an investor in everything from Airbnb and Stripe to really exciting companies in fusion like Helion, WorldCoin, which is a crypto protocol. And so he's done a bit of everything and everything good. We have a lot to cover today. And we're basically going to go from everything from like the history of AI on through to...
[0:00:28] Are we going to face that AI apocalypse eventually? And so maybe we can start on a positive note. And I'd just love to hear your view of the history of AI, what has changed over time in the course of being involved and then running that company. And then we can talk a bit more about some of the... Yeah. I think in some sense, it has been this long, continuous arc that has bent when the computer bends. But the thing that is working now...
[0:00:57] This idea of neural networks is obviously not a new idea. It was like left for dead for a long time. But it's remarkable. So I think it's just been this, you know, longstanding idea that people have been talking about since the computers have been invented. And we finally got enough computers for it to work. There's like a little bit more to the story than that. I think a significant thing that happened was the shift of these large unsupervised models.
[0:01:31] That was sort of not what most people predicted if we go back five or 10 years. But like fundamentally, the miracles are, we have like an algorithm that can really, truly, genuinely learn and we can throw more compute at it and it gets predictably better. How do we get to the current AI stack? I mean, it seems like there's been a really big paradigm shift towards these big unsupervised models from a lot of the things that came before. And so I'm a little bit curious, like, you know, what causes shift from...
[0:02:00] The sort of CNN, RNN and sort of standard GAN world that we had even just three, four years ago. And then Transformers came out with a paper in 2017, I think it was. And then it took a year or two for things to really kind of gel. And so I'm just sort of curious about, you know, what led us to where we are today from a model perspective. The plans in the AI field, if we rewind back to like 2015 or something, is that we were going to, and OpenAI was one of these people, you know, we were going to like train agents, RL agents to play games.
[0:02:30] And then we would put them in more and more sophisticated like multi-agent environments and they would have to learn all these social skills and how to interact with each other and, you know, like eat each other or whatever. And then eventually they would like learn, they needed to learn human knowledge and they would do that. That was kind of a path that a lot of people thought was what made sense. And almost very few people predicted that actually we'd be able to flip it around. And first we could.
[0:02:59] Have models learn all of human knowledge in this very non-agent way. And then once they had that represented, you could use them to do these more sophisticated tasks. But this idea that you were just going to go read all of the texts ever written by humans on the internet with no particular supervision signal, just trying to predict the next word one at a time, was a very laughable idea when we started doing that. And it's gone like...
[0:03:27] Further certainly than I thought it would go, and I think most people in the field. Transformers were obviously a huge part of that. One of these rare times where we got something that was like orders of magnitude more efficient, more compute efficient than the thing that came before it. Really good at, you know, lots of good ideas about transformers, but like really good at making use of the hardware that we have also. But really that one like fundamental idea that, you know, scale up these large unsupervised models.
[0:03:58] And get these like quite surprising results. Get the zero shot learning working, which I still think is like somewhat miraculous. That was like quite not one of these things that looks incredibly obvious in hindsight. But if you told someone that four or five years ago, they wouldn't look like you have no idea what you're talking about. And I think that's been applied in really interesting ways by OpenAI. Obviously, there's GPT in the large language models. There's DALI.
[0:04:28] There's new APIs like Whisper, which is a really cool thing that you folks are doing in terms of speech to text. What are some of the big directions or sort of areas that OpenAI is most focused on going forward? I mean, definitely we're very excited to see how much further we push this same style of model to like be way, way better. So still like a language model or increasingly probably everything will be multimodal, but these multimodal models in the spirit of the GPT paradigm.
[0:04:57] They just go super far. And I think we are far from the top of that curve. So there's still a lot of exciting stuff to do there. We like to build things that are useful to people. And so GPT-3 has found limited use, but real use. And we believe that as we make these models bigger and more powerful, they will do a lot of things.
[0:05:23] You know, Copilot has been great for coding, but we're pretty optimistic that there can be like Copilot for X for many values of X. And the amount of knowledge work augmentation that that can do, even if we never figure out another big research idea in all of the OpenAI trajectory, which seems unlikely, I think would still go quite far. And is the intention to stick to the API layer or build applications for cells over time? Or how are you thinking about sort of broader direction?
[0:05:51] I'm somewhat open-minded about that. I think that APIs and open source models are the best way to grow the ecosystem, and that will be a very key thrust of what we do. I'm very happy that we built the kind of co-pilot end-to-end experience with Microsoft, and I think without doing that, people wouldn't have quite appreciated that. Same thing with Dolly. We put out this thing in labs because we wanted people to just be able to explore it and use it.
[0:06:20] I think it's like a safe bet that we'll do other things like that in the future. But this idea that there's going to be like this explosion of new like companies, business models, products that were just like unimaginable and we want to enable as many of those as possible and be kind of like the core intelligence layer, platform layer, whatever. That's like, that's sort of pretty exciting to us. What do you think is the most interesting thing in ML or AI that OpenAI is not working? Well, if you think something's really interesting.
[0:06:53] But there's a lot of stuff that we're just like kind of too small or too distracted to do. One particular category is powerful small models. So models that are meant to do one thing really well, but not the big general purpose stuff we do, or meant to run on like one piece of consumer hardware. I think we're not super focused on that because it doesn't feel like quite.
[0:07:22] The most direct AGI path. But in terms of amazing work that people are doing that will add huge amounts of value and change the way we use products, I think these somewhat more narrow but very powerful small models, there's a lot of juice to squeeze there. Can people hear in the back or do we need to speak louder? No, we're good. Okay, great. So if you look at waves of technology innovation,
[0:07:48] There tends to be waves where all the value accrues to startups and waves where all the value accrues to incumbents. So if you look at the first internet wave, for example, a huge chunk of the market cap went to startups. It went to Google, it went to Amazon, and then a chunk of it went to incumbents like Microsoft or Apple. And then you fast forward and you go to mobile. And with mobile, it was again kind of a 50-50 split, right? Like there was Instagram and WhatsApp.
[0:08:13] But CRM on your phone, and there's a bunch of startups that were getting funded for CRM on your phone, it turned out to be Salesforce on your iPhone. And then there's waves where all the value goes to startups like crypto, right? There's really no incumbent value that's really been captured by crypto until now. Now, if you look at the first wave of ML, and by first wave, I just mean the last 10 years, obviously there's tons of stuff happening before that. It seems like all the value roughly went to incumbents, right? And so when I think of all the various...
[0:08:41] The AI startups that I know got founded or funded or people invested in or worked on, very few of them actually created any real breakouts, right? The biggest things were TikTok, which is really a sincerity of ByteDance, Google, Netflix with this recommendation engine, Facebook feed. And I can't think of a single startup until this more recent era where there was startup value created. And so I'm sort of curious looking back, why did all the value get an incumbents?
[0:09:09] And then what do you think is going to happen in this wave in terms of that split? I am fascinated by this question. I've like studied it a lot across many different industries. I have nothing wise to say, so I hope you have an answer. But this question of like, is it the rule or the exception the incumbents get more and more powerful is actually like, and which way is that going to trend is super non-obvious. I think if you look back at like the...
[0:09:38] History of companies since the joint stock corporation. You can kind of squint and tell yourself the story either way of the incumbents get more powerful or the upstarts always win is like the default. My current belief is that there has been something anomalous recently with big tech where they've been more powerful for longer and less beatable by startups than they're supposed to be.
[0:10:09] I do wonder if we're just like into a new world where this is the shape of things to come and the value of... We've had companies kind of this big before, if you look at like fraction of GDP, but we may never have had companies like this powerful. And so if you look at what may continue to happen, I don't think we can just say like, oh, you know, we had like Standard Oil and they thought they were going to be huge forever and then all that stuff happened.
[0:10:38] If we believe that AI is the next wave and you look at the resources of the giant companies of today, I think you can tell yourself the story of why they should benefit a lot. And on the other hand, they don't get it. They don't really get what's happening. So startups seem to have the momentum right now. I think what I would say today is it will be the big companies that are able, and I would include OpenAI in this, to create.
[0:11:10] The most powerful models. But it will be startups that figure out how to like get the most value on top of this. Yeah, it's kind of fascinating because 2010, 2020 was purely incumbent. That was a weird. And we had AlexNet and we had CNNs and we had RNNs and we had GANs and we had like, you know, multiple different varieties of really exciting technology that made you think that something dramatic was going to happen.
[0:11:35] And I feel like the last two years or three years have actually started to show that on the startup side. Obviously, there's things like OpenAI, Hugging Face on the tooling side, Jasper and Copy.ai. There's more and more companies. To your point, I guess Copilot is more of an incumbent thing, but it really feels like finally it's going to be a mix again in a way that it wasn't for the decade preceding.
[0:11:57] I don't know if it's a technological capability thing. I don't know if it's a market thing. Maybe founders are more focused on markets, but that doesn't seem to be the case based on the companies I was involved with over the last decade in AI. Yeah. I mean, I think OpenAI is this weird one thing, so I wouldn't try to draw too many conclusions from it. I think we haven't seen this sort of application-level companies emerge yet, or at least not. I think there will be new trillion-dollar...
[0:12:24] AI application companies that start. My guess is they haven't quite started yet. Maybe the first ones have. But I would bet in the next few years we see them get started. And I think we haven't been able to say that in any category for a while. We are far from the first people to observe that the 2010s were this weird mix of investors making a ton of money because we were in this up elevator and valuations went nuts. But a relatively small number of companies, particularly in the...
[0:12:54] Later, last three quarters of the decade, it felt like they're going to be the future trillion dollar platform companies. And I don't know what to make of that. Again, if we go all the way back to when this all started and the British government decided that they were going to create these second order sovereign things called joint stock corporations, and that led to this, enabled this just unbelievable.
[0:13:24] That like one innovation enabled all of these other things, these other innovations to happen and get to market. Part of the implicit arc was they were supposed to remain second order. They're supposed to be like less powerful than the actual sovereign states. And to the degree that that looks like it's to some degree changing, and these companies do feel hard to compete with, like how long does society tolerate that? I think all these things are just like hard to predict. If you look at history as an analog,
[0:13:52] And you look at mobile as an example, the capabilities of mobile suggested a subset of the companies that ended up doing very well. So for example, you knew that people would message on mobile. And so WhatsApp and Signal and Telegram and all these things came up where you knew that cameras became the biggest mechanism by which people took photos. And so you knew something like Instagram would exist, although it was really hard to know a priori exactly the mix of features that would make it so magical, like the founders of Instagram did it. And then there was non-obvious things. Uber, you push a button and a stranger shows up in a car and you get in and you feel fine with it.
[0:14:22] What do you think are going to be the characteristics of the application-level companies for AI that are going to work? What are the common threads that people should be looking at or looking for? I think we can say all of the trite things at this point. I do believe that natural language is this interface. That's real, and that's going to happen. And we will control lots of things. Not everything, because sometimes a GUI is better. But we will control lots of things by telling a computer what we want.
[0:14:57] And this fact that we can generate any content that we want, I think is going to lead to a predictable but still huge set of things where maybe not we get to generate images, but soon we'll generate full video games or full movies or never-ending Spotify playlists or whatever. So we could talk about the obvious things like that. And there's another category of automated knowledge work that I think we could also talk about is fairly obvious. So we'll have AI lawyers, AI medical advisors, AI tutors.
[0:15:27] Like each of those verticals, I think we'll just get picked off. And then of course, like to me, the most interesting one, but by definition, the hard to predict one is the hard to predict stuff where we'll just see these like very different things that happen when intelligence, when first of all, it's like an alien intelligence. So it's hard for us to even sit here now and say what it's going to be good and bad at, but also an alien intelligence that can run at like.
[0:15:59] Just this unbelievable speed and unbelievable cost effectiveness. We hope cost effectiveness. The sad version of the world is where it turns out there's just like a market price for intelligence society is willing to pay. And so GPUs go up in price by like a factor of 100 and so does energy and we don't get the happy version of the revolution. But assuming we can avoid that, which I think we can, then this world where like intelligence is super fast, super cheap, and super alien, I'm excited to see.
[0:16:29] Yeah, I think your point on knowledge workers is a really key one where it seems like the things that are working the best right now are highly repetitive white-collar tasks where there isn't a good workflow and you kind of layer the ML on top of it. It's your sort of combining workflow plus ML in a knowledge worker-centric, tech-centric sort of world. Do you think the primary approaches or the biggest companies are going to be text-based, image-based, video, voice? Do you have any opinion in terms of the format? I sort of don't.
[0:16:59] I think the modality will be the thing that matters the most. I think it'll be like, does the shape of the problem lend itself well to the limitations of AI? So can the shape of the problem tolerate, you know, breathtakingly good answers 99 times out of 100 and one totally wrong made up one, one time out of 100? Is the kind of feedback that you get from people using it something that the model can improve on?
[0:17:27] The sort of human feedback and the usage data pretty well. Does it like fit a nice sort of like AI workflow? So I think it'll be more about those things. And then you can do that in, you know, full computer using agent, like just a chatbot, whatever you want. But it'll be like, given the limitations of our current tools, does it fit?
[0:17:53] You mentioned earlier this difference between the large-scale models and then the smaller bespoke customized approaches. Where do you think each of those will apply most? I think if you have just a road task that is super valuable and you want done again and again and again with some variation, these small models will be great. And if you need the kind of reasoning or thinking that we like to think of as special and human, the big models should be way better at those.
[0:18:25] It's interesting because when I talk to people who do research in the field, basically everybody points out that they feel that GPT 3.5 or whatever the next eventual version will be always seems to be a step ahead of the rest of the industry despite the scale in terms of number of parameters, other aspects of the models. Do you think the advantage is, do you think the reason that the GPT models have had such a magical feel to them where others haven't had quite the same impact is that because of data and data quality?
[0:18:54] It's a model design. Is it something awesome? Sort of curious where you think that ongoing advantage comes from. I mean, I think we do have the best engineers and researchers in the world. And I know everyone says that about their company. So you probably like shouldn't believe me either. But I really think we do at like a very, very high density. And then I think we just like care about getting a lot of little things right. It's no like there's like no, I probably wouldn't tell you if there was, but there's no like one secret. Like here's the single thing. It's like.
[0:19:27] We just, it's like a lot of things. We just like, we care. I think one other thing that this kind of touches on is the shift towards an increasing number of open source models and then AI safety and AI alignment. And I'm sort of curious how you're thinking about the world relative to open versus closed source and where you think each of those things will be used most. I think it's great. I'm happy to see the open source bloom. I don't think, I think there is like a...
[0:19:59] Capability level of systems where there is genuine danger, like not just the risk to offend, but the risk to actually cause like irreparable harm to the world that we should not open source. You know, we don't open source nuclear weapons really either. We decided we're going to like have some controls on that as a society. But in terms of like image models and, you know, we just open source whisper the language model. I think that's tremendously exciting.
[0:20:29] And I am someone who thinks that lots of AI usage in the world is a good thing, not a bad thing. And I think open source is an interesting, maybe one of the most interesting vectors for that to happen. We'll open source more stuff. I'm sure other people will open source more stuff. We open sourced Clip like a year, two years ago. I don't even remember now, which really kicked off a lot of this whole thing. And that was like, you know, we looked at that. We sort of looked at the safety issues there and decided that was a good thing to do.
[0:20:56] And we'll certainly do more in the future. But yeah, we're pro that. One of the things I similarly hear when I talk to a lot of different people working in the field is, you know, we initially had this concept of AI alignment, which is how do we make sure that we don't have an existential threat from AI? And how do we align human future hopes with the hopes of the eventual intelligent machines? And then there's safety. And safety encompasses alignment.
[0:21:25] It encompasses, will somebody get offended? In some cases, it seems to encompass politics or certain political leanings. And when I talk to people in the field, they seem to think that often companies don't launch a lot of stuff because of safety reasons along that whole spectrum. For example, there's a few really big companies, not OpenAI, that are notorious for never launching, even though they developed a lot of these really interesting technologies. What do you view as the impact of safety on the field? And how do you think that evolves over time? Yeah, it's a great question.
[0:21:54] The word has become so overloaded that I think it's totally meaningless. And you have people who only use it to mean models shouldn't offend or say something that they perceive as toxic. But if you're talking about AGI safety at all, you're awful and you're distracting from other issues. And then you have people who say, well, all that matters is AGI safety. And if you're trying to distract with...
[0:22:19] You know, kind of responsible AIs today, then you're just like hiding, you know, you're like defocusing from what matters. And so I think like in my own language, I have tried to always differentiate what I'm talking about because I think they both matter. We have taken a lot of flack for shipping models. You know, we put, now quite some time ago, we put GPT-3 API and that was not a popular decision. It's like, how could...
[0:22:49] It's so racist. It's so toxic. And what we said is we've got to learn from context reality to make this better. We want to do it while the stakes are very low. And we were then able to improve it quite quickly. And if you use the current models in our API, I think you will find in terms of the model not saying stuff that is likely to offend unless you want it to, it's pretty good. And generally speaking, we think that society will say,
[0:23:17] Here are these broad categories of what we agree a model should never do. But then within that, there's going to be a lot of plurality and the models should do what their user wants. Sub-AGI models doing what a user intends within some very broad bounds, I think, is a reasonable thing for us all to aspire to. And I think that if we try to say like...
[0:23:43] We, OpenAI, get to make all of the decisions about what a model can ever say. That is going to be and should be unpopular. And I think by learning with contacted reality, by deploying these models, by seeing how people try to misuse them, we have been able to actually learn pretty quickly. And a lot of the lessons have not been the obvious things. So the people who kind of like say, you should never deploy, you guarantee all of these things, like Google is a famous example of this. I think that's just wrong.
[0:24:13] And I also think from the other direction, this idea that like, you know, we're going to go work in secret in a lab and build AGI and then drop a super powerful AGI in the world all at once is also wrong. That's like a very high stakes decision. It's like much better. And it doesn't give like people and institutions time to adapt, learn and think. One of the things I think OpenAI has done well, although I realize some people disagree with us, is push the Overton window on AGI.
[0:24:43] Talking about AGI five years ago was a very crazy thing. People really didn't take it seriously. Talking about AGI is crazier. And now people are like, oh, yeah, AGI is going to happen, of course. Like, yeah, that AGI safety thing, yeah, we got to get that right. But they fucking don't. And so if you're just like moving that so that more people in the world get to discuss this, get to weigh in, get to think about what they want, get to like really wrestle the patience of all of this, we think is really important.
[0:25:13] On the AGI safety front, I do think the issue of how we align a powerful AGI is really important. But I also think that the people who sat in a room and tried to do this in like Microsoft Word instead of VS Code got it, from what we can see right now, pretty wrong. And they're very smart, extremely smart, competent, caring, thoughtful people. I think there's just a limit to how much progress you can make.
[0:25:43] We talked earlier, if we were in this world of RL agents that were trying to feed each other in a simulation, that's very different than if the first thing that happens is you develop very non-agentic language models that you can understand what they know. You can change it. You can align it in all these different ways. You can have them talk to you about what they're doing. There are still a lot of problems about intermiss alignment that may emerge and that we have to have different ways to measure.
[0:26:13] If you think about how different the shape of the AGI alignment problem that we face is now and what our toolbox looks like for it and kind of the problems that we have to get right and in what order versus what the just like absolutely accepted consensus reality was, it's really different. And so this is another reason why I think you have to sort of like work on safety capabilities and alignment together. And they're not orthogonal vectors.
[0:26:43] They're not quite the same vector either. They're pretty closely aligned. And treating this all as one big technical problem that we have to get all of right, I think is how we're going to solve it. If you look at the next 10x, 100x, 1000x in performance for AI models, irrespective of the metric that you want to choose as the metric that matters, where do you think that breakthrough comes through? Is it chips? Is it algorithmic?
[0:27:10] Is that engineering deficiency? Is it something else? Is it scale? It'll be like all of those things multiplied together. We will come up with better algorithmic ideas. We'll get faster chips. We'll figure out how to use more of them for one single model training run. We will get better data. We will get better ideas on top of the pre-training. And all of this multiplies together. I suspect that there exists out there some algorithm.
[0:27:40] Or some architecture that is as much of a step forward as the transformer was. I really hope we find it. I think there's like a lot of work to do there. And then on the hardware scale, yeah, that's going to keep going. So where do you think we are in five years or 10 years? So it's 2027 or it's 2032. What do you think the state of the world will be? I think we'll have like superhuman intelligence by 2030, I guess. In 10 years, we'll have like AGI?
[0:28:17] In what ways is GPT, whatever the current GPT thing we have in the API now, I can't exactly recall, but in what ways is the best model out there right now not EGI? In what ways do you think it is? I mean, there's obvious superintelligence, right? I mean, we've had 15 years of... So it's definitely not superintelligence, but in terms of sort of general purpose intelligence. Oh, your question is when we have legitimate superintelligence. Oh, that could be more than 10 years. I hope that's more than 10 years. If you sort of think about the...
[0:28:46] Like the quadrant that I think most people in the field, the two by two matrix that people disagree about. You have like short timelines and long timelines till AGI. And then the ramp from AGI to ASI, super intelligence, whether that is a slow ramp or a fast one. And I hope we are in the short timelines, slow ramp world. I think that would be like the safest and the best and easiest for us to adapt to. So I hope we don't have ASI by 2050. You do think we'll have some form of like general intelligence in a machine?
[0:29:16] 10 years. I do, yeah. And then what do you think are the problems that this general intelligence will help solve first for humanity, if any? I think there's like a lot of debate in the exact order of how this is going to work. Like, you know, does the, is it easier? Well, actually, before I even answer, if we look at the historic, I love looking at predictions from the recent past. So if we go back five years again, the consensus wisdom would have...
[0:29:54] First, we have the physical jobs automated, the truck drivers, the factory workers, whatever. Then we have the like easy white collar jobs and the really hard, high cognitive labor, white collar job, mathematician or whatever. And then finally, if ever, because maybe creativity is like magic and human and somehow like outside the realm of science, maybe eventually we get the creative jobs. And obviously that prediction was like perfectly wrong for I think quite interesting reasons. And we could sit here now.
[0:30:25] You kind of know that's going to be the shape and talk in each of those categories like, well, is it artists first? Is it visual artists first and then musicians? Or is there some reason that's wrong? And then like, is it going to be the AI doctor or the AI lawyer first? And is that really important? But in some other way, I would say it just doesn't. And the fact that like cognitive labor is not necessarily going to get replaced, but going to get hugely augmented, co-pilot style, pretty fast in all these things.
[0:30:54] Where each of us are just going to be way more productive at what we do. And then someday, yes, it will be automated. We'll find new things to do at that point. But I think most current jobs, except the ones that we don't want to be automated, or we want them to remain high status, or we really want a human involved. Other than that, most things do get automated. I sort of think whether one thing happens in 2031 or 2032, in what order.
[0:31:23] It is somewhat interesting. It does somewhat matter. But the main thing is that it's all going to happen. Yeah, it's interesting. To your point on self-driving or other things, I remember six, seven years ago, multiple Congress people were doing tours with Silicon Valley people basically asking, what should we do with all these displaced truck drivers that are coming? And of course, nothing happened. So it was an interesting, interesting. What do you think happens to humanity?
[0:31:51] And so I think there's almost like three ages to mankind or to compute, right? Up until recently, almost all compute was human. And then now we're in kind of this mixed era. And then we're probably going to move into an era where most of the compute in terms of actual impact of the world is going to be completely driven by machines. We're kind of in this transition phase. Do you think AGI eventually subsumes us? Is biology just a bootloader for a future machine intelligence?
[0:32:23] Certainly the future will be like, and what humans do and how we spend our time and what is made available to us will be really quite different. But that's been the case for a kind of like life before and after every technological revolution in history. I do think this one is going to be the biggest one ever, but I suspect that's a mistake. And it's just each one gets bigger and there will be something like far more huge on the other side that we can't like.
[0:32:51] Think about trying to think about artificial intelligence before the computer revolution. Before the industrial revolution, at least, that would have been tough. But do you think it's a technological leap or do you think it's a shift in terms of introducing a new species that's competing with us? Because it seems like at some point you hit a transition. I think if we're in any world where it's like an us versus them situation, that's quite bad. And we just shouldn't go there, which I think means some version of a merge, at least for some of us.
[0:33:21] Or as my co-founder Ilya likes to say, we build an AGI that really loves humanity. And that sounds sort of silly and dismissive, but I actually think it's like a totally reasonable strategy. But any of the kind of like us versus them competition, I think is bad. Now, I do suspect that even if we think like the merge is good and a lot of people are going to do that really long term, there will be many people who don't want to. I think human drive is willful.
[0:33:50] Be remarkably the same as they have been for tens of thousands of years. And there will be like, I don't know, maybe it's like an exclusion zone or something where if you want to like, you know, live the like no AGI life on Earth, you do that. And if you want to merge and go off for the stars, you do that. Like, you know, these things are like fun to speculate about. But I think pretty hard to say from here. The one thing that I feel confident about is like we should avoid the us versus them dynamic in one of the many ways we can choose.
[0:34:21] Yeah, it makes a lot of sense. I guess I always wonder, we may want to avoid that dynamic, but if a new species emerges, it's going to have its own utility function. It's going to have its own drivers. And we'll effectively have a self-reproducing, self-editing digital life form that can reproduce at scale as long as you have enough compute available. And then there'll be some physical substantiation of those forms, right? There'll be advanced robotics and other things. Yeah, I think the whole survival of the fittest version of evolution is over, thankfully.
[0:34:51] That wouldn't go well for us. But I think it gets to be much more directed here now. And we get to think and reflect and decide how we want this to go. And again, I think there's multiple good options. And I think society as a whole should get to weigh in on which path we go down. I think people often try to distinguish between consciousness and sentience. Do you think that the GPT models are conscious? I don't actually know what the difference between those words is.
[0:35:22] Sentience is a specific form of self-awareness, while consciousness means you're able to process and interact with your environment in an intelligent way. So sentience is like the I am awareness. I'm aware of myself. Consciousness is just like, then what? I'm aware of the world, but not necessarily my own role in it. There's a really good book called, I think it's Blindsight, that gets into sort of this fine line between these two forms of intelligence in some form.
[0:35:53] Let me change the question slightly versus getting into the semantics of these things. In whatever definition you want to use, how do you think about the consciousness of GPT-like models today? I think Descartes was basically right about this. I can be sure of my own consciousness, and that is the beginning and the end of all certainty. I can't be sure your conscious are not much less GPT-3. A lot of people tell me that, actually. I think it's much more likely you are.
[0:36:29] I don't, like, what would ever convince me that an AI is conscious? Like, if you ask GPT-3 if it's conscious, it will tell you yes. If you explain that it's, you know, what it is, it'll say, okay, now I agree with that. But future models will be much more convincing. And, like, how we'll know, I think is, like, very interesting, like, late-night fodder. I think there are a lot of tests we have talked about to devise how...
[0:37:00] How we would be more or less confident if a model was conscious. But we don't have anything yet where I'm like, yes, I would like stake my life that that isn't deep. This will be a test. So I think we have like promising ideas, but no certainty. If GPT-3 is conscious, it is an extremely alien form of consciousness indeed. Like think about what a strange thing that like this thing is just sitting there asleep. Someone passes a prompt.
[0:37:29] The energy flows through forward once it's awake, it's conscious for that moment, it gives the answer and then it's like gone again. And everything that was in that context is just like lost. So it's this frozen in time set of weights. I guess I could more likely believe that it was like somehow conscious during the training process. Even that feels unlikely to me, although I express uncertainty and think this is like sort of a question outside the realm of science or at least outside the realm of something that I feel like I know how to be certain about.
[0:37:59] But I think all I could say is for my own perception of consciousness, there are things that I believe to be important with no certainty that GPT-3 does not do, does not have as part of its architecture. So that seems unlikely. An interesting thing I heard recently, and then I'll shut up about this because clearly I have nothing definitive to say, is a hard question.
[0:38:32] Someone proposed this idea that the human experience of consciousness is that we can't predict the outputs of our own neural network. And so we have what feels like this observer and this inner monologue, and they feel like separate things. But also these neural networks, the artificial neural networks, are going to get so large and complicated that they themselves can't predict the outputs either, and so have the same subjective experience.
[0:39:02] So I'm not going to move to some questions from the audience. And these were inputted as people signed up to attend. And then if we have a little bit extra time, we can do some broader Q&A. One question that came up is, what are beliefs about AI that you held before that you now believe to be incorrect? And what is the most recent big update to your thinking more generally in the other direction or either? You know, I sort of did know when we started OpenAI at the beginning of 2016, how important
[0:39:35] Scale was going to be. But I clearly didn't act like it. And I think I somehow hadn't internalized it. And we set up the company in the wrong way. We focused on the wrong things. And then there was like a moment. I remember it so clearly. It was in the winter one year. I remember where I was sitting. It was storming outside where I like really got religion on this. And I was like, damn, I was really wrong. We should have gone all in on this. We had not been acting like it at all. We're just going to go do it. So that was like, that was this.
[0:40:05] Really weird mistake. I've thought a lot about that, where I intellectually was right about everything and all of my actions were wrong. What do you think is inefficient, hard, or tedious for an approach to AI? Inefficient, hard, or tedious? It actually reminds me a lot of what I've heard about the very early days of computer programming, where we just have terrible tools. And so there are these things that we do.
[0:40:37] In AI infrastructure or like how we rack up servers or the whole stack that will in a few years be like fully automated. And the things right now, like teams of people, there will be like great frameworks and tech stacks like equivalent to AWS this morning. We'll just push a button, it'll happen. But we are very much still in the like painful early part of the tech rollout.
[0:41:05] What is a world where OpenAI becomes 100 times more successful, but in a direction that would disappoint you? A lot of answers to that question. I mean, I think there are a lot of things that could happen. We set up this structure called a profit structure because we did not want the incentive to make unlimited money. This was another thing that at the time people were like, these people are incredibly self-indulgent idiots that now I think actually looks fairly prescient.
[0:41:43] And you see around the office people making different decisions than they would as well-meaning as anyone is if they have the incentives for us to just like grow infinitely. So I think we got the financial incentives right. But there are a bunch of like power incentives that I don't think we've thought about enough yet. And the way that I think that could go wrong is we become like way more successful. We are confident that we have like...
[0:42:13] The right financial incentives in place, the right governance incentive mechanisms in place. But somehow we accumulate too much power before we figure out how to dissolve it. And that goes badly. Where do you think ML changes the game for movies, animation, and video games? I think we're already seeing this. The video games where people are making their own art, they're designing their own persistent characters.
[0:42:46] You're seeing the very beginning of like very short, extremely short videos. I won't even call them movies yet. You're seeing artists doing incredible things. Like one of the fun things for me is to watch how much better the results that an artist gets out of Dolly than I do. Like there is something there that's really cool. And I think all of this is like, you know, the first one hundredth of 1% and it's going to go much further, but you can see what's going to happen there. And it's very cool.
[0:43:17] Massively, massively powerful tools for creators. If you look at technology waves, they tend to accrue first to technology companies and then other enterprises adopt. And if you look at the first set of machine learning platforms, they've been adopted with some Silicon Valley companies and then eventually they turn into services models for big enterprises, in part because of a lack of human capital at the enterprises. They just don't have the people to go and implement things.
[0:43:45] In many cases, they aren't even doing like linear, they're not even doing regressions, right? They just, they're not quite there yet. At what point do you think AI will impact every enterprise? And how do we get there? One of the things that I love about AI is, or at least the direction it's going right now, and this may not continue to be the case, the more you know about the AI, about AI, the field, the worse you are at getting value out of the current models.
[0:44:16] And so if you watch like an AI expert take one of these models and think about what they, they bring all of the baggage and the stuff that's wrong with it. And if you just go take it to like, you know, random person in a non-technological enterprise and say, hey, this is this magic box. You type in whatever you want and it'll give you some response. They figure it out much, much faster. They come at it with a beginner's mind in a good way. You also see this, by the way, with the difference of like kids and adults that use GPT-3 or Dolly or whatever.
[0:44:45] And so I am hopeful, based on what we've seen so far, that this is going to be a surprisingly democratizing technology. And it won't reward deep technological expertise. It'll reward knowing your problem space really well or being really creative or coming at something from a fresh angle. We have a few minutes left, so I was hoping we could do two to three questions from the audience. And then we'll have about an hour of networking here at Notion afterwards.
[0:45:14] If there's one or two questions that people want to ask, I would like a hard question. You can do the mic. Well, I'm super lucky to be asking this question. So Dr. Michael Levin was on the Lex Friedman podcast recently, and he's a synthetic biologist at Tufts University, I think. So since watching that, I went on this deep rabbit hole, watched all his lectures and stuff. He's created these new things called xenobots, which use gap junctions at the cellular level to message paths.
[0:45:48] Molecular information. When are we going to connect the synthetic biology, primitive programming infrastructure stuff to just like training GPU clusters with lots of data as to these models? That's my question. You mean like when will we build biologically inspired computers or when will you just like plug your brain in neural link styles? Well, I guess I see a disconnect between like what the xenobot depends on, which is different than what computers depend on, which is just electricity. These xenobots.
[0:46:20] Use electricity to pass the molecular information, but they're basically frog cells. So they require, you know, what normal multicellular life forms require, which is like water and oxygen and that sort of thing. Like computers don't. I guess computers kind of do require water. I'm just asking, when do the worlds start to intersect? Because Michael's domain is very distinct from sort of...
[0:46:45] You know, matrix multiplication, sort of, you know, large optimizations of these kind of neural nets. It just seems like a different domain. It seems useful to maybe cross-pollinate potentially. Maybe, but I'm sort of a believer in like information theory here. And there were all of these reasons that biological intelligence evolved the way that it did and all of these like extremely difficult constraints and quite miraculous things that come out of it.
[0:47:14] The brain does do amazing stuff and amazingly power efficiently. The outer loop of evolution also was quite a wide search space. But I would bet you can simulate all of the stuff you need in silicon without having to complicate the process and probably should just sort of say transistors are great. Computational simulation, leave it at that. That's my guess. Yeah.
[0:47:43] Sorry, I'll stop there. Thank you. Thanks. Do we have enough information in the internet to create AGI? If you contrast it with... Yes. We have a continuous video feed to our eyes. And on the internet, we only have like a subset of that. Yeah, we're confident there is. We think about this and measure it quite a lot. What gives you that confidence? Is it the size of the knowledge base? Is it complexity? Is it some...
[0:48:20] One of the things that I think OpenAI has driven in the field that's been really healthy is that you can treat scaling laws as a scientific prediction. You can do this for compute, you can do this for data, but you can measure at small scale and you can predict quite accurately how it's going to scale up, how much data you're going to need, how much compute you're going to need, how many parameters you're going to need, when the generated data gets good enough to be helpful.
[0:48:50] The internet is like, there's a lot of data out there. There's a lot of video out there too. A few more questions maybe? I've heard three or four definitions of prompt injection, so will you just say what you think there? Yeah. I think that that is going to be a combination of training models to act in specific ways to be aware of that, to sort of like make it clear when that's happening.
[0:49:51] And then also the set of tools on top of it. But I think we'll be able to go pretty far in the actual model itself, knowing what's happening and making you aware of what's happening. Well, yeah, I mean, I think if we're depending on like prompt stuff in the beginning of a prompt to make it safe, like that's not the world we should be in. And so you are seeing some things now where...
[0:50:30] I don't think that's anyone's favorite long-term solution. And I think as you pointed out, there's like many reasons why that's bad. And that's like, I don't think that's still what we're going to be even like one or two years from now. There will be like other cases where I think even in a world where we're not depending on that projection still leads to weird behaviors.
[0:51:00] I think the models can be aware when it's happening and alert people like, oh, I have these conflicting instructions in my prompt. Here's what I thought about balancing them. First, you told me never do this thing, then always do this thing, and now this. One last question. One more from the back. Could we have an AI winter and what might cause it? Yeah, of course. I think we won't have one very soon because even if we never figure out another research idea,
[0:51:41] The economic value from the current paradigm and how much further that push is going to carry us many years to come. But it is possible, however unlikely, that we are still missing the key idea to go beyond behavioral cloning. And these models are going to be stuck at human level forever. There's a bunch of reasons why I don't think that's true. But if anyone tells you that we cannot possibly ever have another winter in this research field, you should never believe me.
[0:52:09] Cool. Well, thanks so much to Notion for hosting. Thanks, everybody, for coming. As mentioned, y'all are welcome to hang out for the next hour or so. I think 90% of the audience or people who are working actively in this field is founders or engineers or product people or founders. So please hang out. Have a good time. And in about an hour, Notion will kick you out. And thanks again. Thanks for having me.