Where is AI Taking Us? | Sam Altman & Vinod Khosla

Good afternoon. Please take a seat and we'll get started. Thanks, Sam, for coming.

Thanks for having me again.

Oh, always happy to have you. Is there anybody who's not happy to have you?

You must be bored of this talk after the 4th year in a row or whatever it's been, but it's fun to be here.

So I'd love to start in— we've had multiple conversations along this line. Imagine the world beyond 2035, so 2035 to 2050 kind of period. To be vague enough about the timeframe. What do you think the world looks like?

I mean, I think people will still be like sitting around eating lunch together. So in some sense, like—

Will there be working?

But I do think the rate of technological change in that period will be difficult to have a frame. It is difficult to have a current framework for it. It's very hard to make precise predictions like, oh, it'll be, you know, the Dyson sphere will be getting built, or we will have figured out nanobots, or, you know, but something, many things like that, I think will be really quite different. So maybe the human experience doesn't change that much. Evolution takes a long time. Biology is like baked in very strong drives over a very long period. But in terms of what's like possible in the like tech stack and what one person can get done, that feels like it will be very different.

So go a little bit deeper, like what happens to today's companies? I give a talk this morning that I'll repeat for ILPs this afternoon, I figured there will be a faster demise of the Fortune 500 in the 2030s than we've ever seen. And which ones survive and which ones don't is gonna depend on what they do, but a faster demise. Do you agree? Do you disagree?

I don't know. Yeah, my instinct would be faster, but I haven't thought about that much. The thing that I've been focused on is what it, you know, most of my time outside of OpenAI I spent with sort of software companies. And I've thought that I have understood the sort of like physics of software companies for a while. But if we're heading towards a world where any software you want can be like just-in-time written. And if you want to do something, you can just type something into an AI chatbot and get a great piece of software built. And so instead of going to buy this SaaS company's product or that one, you just say, run, and it happens. That feels like a very significant change that is not that far away.

I will come back to the next 18 months and talk about that too.

Big Fortune 500 companies that are managing complex supply chains, the physical world always takes longer, but the software change feels like it's closer.

But 2040 is long enough for the physical world to change enough. That's why I picked the 2035 to 2050 period. 2035 is only 10 years away.

Yeah, I would bet, I would bet that most current companies fail to adapt quickly enough and get, you know, some significant losses because of that.

Yeah. It is surprising—

That's been happening for a long time. Like, my sense is that the acceleration of growth in new companies and sort of taking share from the incumbents. I think that this is like one big exponential where new companies can get bigger and more important faster. OpenAI is an example of this. We got very big, very fast.

Yeah. In that time period, is there a profession— and in our audience are entrepreneurs that are building AI doctors, AI therapists, AI oncologists, AI structural engineers, AI chip designers, of course, AI software engineers, AI salespeople, AI marketing people, AI accountants. Like, that's all happening today. I'm presuming all that, whether these companies are successful or some competitors successful, will be mature or their jobs 2035 and beyond. In the intellectual world, and we'll come back to the physical world, that AI won't be able to do at least 80% of?

There are a lot of jobs that I don't think you want an AI to do or that many people don't want an AI to do. And there will be new kinds of jobs that people particularly want a person to do. I think we're very oriented towards caring about other people. I think that speaking of these like deep biological things, I think that's one of them. So yeah, maybe you could have like a great AI teacher and it will not be as motivating to you as a mediocre human teacher. I could totally believe that. I could totally believe there's like something very deep in just knowing that it's a real person or not.

Yeah, I might disagree with you on that. I think an AI teacher, can do so much more and understand you so much better.

But I think there will be a far better AI investor than you very soon. And I personally—

No question about it.

Well, no, no, but I personally still enjoy having dinner with you. And, you know, if you tell me like, good job, or you should do this, that's more motivating to me than an AI telling me that. And so I think you'll still have a job.

So I absolutely agree with you. No question my job isn't any safer than most other jobs.

Much less safe on the merits. But there's this other point.

Yeah. Mark Andreessen claims it's the only job venture that will be safe. But I disagree with him fundamentally.

I actually have wondered if we should make like an AI startup investor. I mean, obviously I like have thought about the problem for a long time. But I think with like current models, it might already be doable. It would be a kind of a fun little side project.

It'd be a fun side project to do. And Sam has a new baby, so I do think we will have a lot more time. Mothers won't have to leave their kid at 16 weeks and go to work. I don't think— I think we will have that freedom. We'll have time to look after elders. I think that part of the human relationship thing I agree with. But services like teaching and medical care and others, I might disagree. I guess I was trying to think about what jobs can't be done. And then the question becomes in the shorter term, Who then makes that happen?

I mean, to push just like a little bit more on the teacher point, I think in some objective sense, I either learned more or could have learned more from reading Wikipedia than I did from any teacher. But when I think about the people that like, or the moments in my like learning process very broadly, it's all about particular people. That connected with me, that took an interest in me, that kind of like got to know me and cared about me and I could tell. And that, yeah, like some of that will be replicated by relationships with AI, but I think this whole thing is going to be stranger and more complicated and more just like uneven than it sounds like on the surface. AI will be able to do a lot of a lot of different jobs, maybe almost all of almost all jobs, and yet we will find out that biological programming is just very difficult to overcome.

I would completely agree with you. And in fact, this morning I talked about the fact that the biological programming won't evolve. So we will still want status, we'll still compete, we'll still have influencers in the attention economy, and we will still have caring for our children and elders and family. So, and we'll have much more time for all these things. So I agree with that. But let me move a little closer in. It was roughly 3.5 years ago that the ChatGPT moment, as people call it, happened. 2.5. 2.5.

Yeah.

So 5 years from then, if you look forward, which even 18 months from now, the end of '26, how much of a change in AI capability from today to the end of '26? And I'm getting very short-term now versus what we saw between ChatGPT moment and today. 2 and a half years ago.

I'm not sure how to measure that. From a vibes basis, I would expect— well, I think going from nothing to the first version of ChatGPT was maybe the biggest system shock most people will feel. Because it was like, all right, there wasn't this thing at all, and now there's this thing. And sure, it wasn't very good, but the zero-to-one moment is a big deal. And maybe we've now gone from 1 to 10, and in some sense that should feel like a bigger deal, but I think it doesn't quite to most people. And maybe we go from like 10 to 100 in the next 18 months, but I think like everybody's already accepted that AGI is going to happen and life is going on and, you know, you're still all out here doing whatever you're doing.

So clueless or not?

Well, again, I think like the actual progress on the merits will be astonishing, but it's expected to be astonishing, whereas ChatGPT was a very unexpected thing to most people. So I don't think it'll like feel as wild, even though the impact will be much larger.

I would venture to guess most people don't quite grok the magnitude of the change, the next 10x from 10 to 100. But I could be wrong, especially outside a small circle.

A year ago, I would have agreed with that. 2 years ago, I certainly would have agreed with that. I'm sort of surprised now, like just talking to random people on travels or whatever, people are like, oh yeah, AGI is like almost here. Like, you know, computers way smarter than people already, going to get way smarter. It's going to all be really different. But yeah, I don't know. I think there has been a big update there.

So all this is happening because of scaling laws, basically. How many different scaling laws— talk to us about the mechanism of this exponential change. What's driving it?

Yeah, I mean, at a high level, the inputs have not changed very much. We keep discovering better algorithms, so we keep finding steeper and steeper scaling laws. We keep figuring out how to build bigger computers and more powerful chips and also connecting more of them together. We keep finding more data and better data. I think we'll start to get to a place where we put these systems together in new ways and we have kind of continuous learning where the systems just run forever and get smarter and smarter. But the— I mean, really the story of the last few years has just been better algorithms, bigger computers, more data. And that I wish I had like a deeper and more insightful thing to say. And, you know, a lot of the algorithmic progress has been incredible. Certainly the reasoning stuff that we figured out and the original idea of unsupervised learning, I would put both of those in that category. And there have been some smaller ones too, but that's just kind of like you grind this out. It reminds me of any other combination of research and industrial learning thing like the transistor, something like that in history?

So there are the scaling laws and multiple scaling laws, and depending upon if you ask ChatGPT, defines multiple types of scaling laws. But there's another factor I think about, which is when do you think AI scientists, start doing most of the AI research?

I think it'll be very gradual. It's not a, you know, like let's say today that an OpenAI researcher is using Codex and it's generating 10% of their PRs and then 20% and 30%. And then it's like starting to actually go test some new model architectures itself. But it's, you know, still like the researcher sort of directing it. And then it's like doing a little bit more of the like, all right, here's the hypotheses to go test. But the researcher is still like, feels like they're just working more efficiently. And the researcher might say, I'm doing 100% of the research. I just have better tools. But if that researcher is now outputting 2x or 10x what they could before, let's say 10x, do you count the AI is doing 90% of the research at that point or 0% because it's not fully autonomously doing the whole loop. I think it's going to be this sort of like messy joint acceleration.

So one doesn't have to measure it, but one can look at the rate of change of new— of progress and the acceleration. That's sort of, to me, the ultimate measure. What happens to the rate of change? Beyond just the scaling laws which we've just talked about.

I think we will move much faster on the research front every year from here on because we have better tools. And so whether you want to call that AI assisting humans or humans assisting AI or AI doing the research itself, the net effect will be much faster, not only algorithmic process, but the whole, the whole like supply chain. If AI is helping us build data centers faster, if it's helping us develop new sorts of chips, including pretty exotic stuff much faster, I think that all counts. And so yeah, if you look at it as sort of a rate of progress, and as long as the AI is somehow helping that go faster, you count that as the AI doing part of the research, then I think much, much faster.

Yeah.

To me, the key test is, is AI coming up with new hypotheses that it tests by itself and then modifies those hypotheses? Like, that's a cycle that seems vicious in a good way.

I have a slightly different perspective. I think what we care about is just how quickly we get much better research done. And if the AI is coming up with the hypothesis itself or enabling a human to come up with a hypothesis that human just couldn't do on its own, I'm equally happy, truly.

Yeah. There's a question many LPs here ask me, which is, will the people who are winning reinforce their lead through this acceleration? And if this hypothesis of acceleration in research is right, then, then you and some of the other leaders get to perpetuate their lead. It's harder for new players to play from scratch unless they come up with completely orthogonal approaches. And as you know, we are working on some orthogonal approaches. But does that perpetuate— give OpenAI an advantage in having an increasing valuation, which we all hope for? I for sure hope for that. That's a question on many people's mind.

Look, if I were an LP, I would be spending 0% of my time trying to figure out how to invest in another AI research lab and 100% of my time figuring out how to invest in the thing that comes next. The thing that I mean, GPs are also horrible at this, but you're not. You're pretty good. But in general, so much of the kind of mindshare of investment capital in the world goes to chasing the last set of winners. And you almost never make money. You almost never make a lot of money there. You almost always make the big returns in looking at what's now possible in the world that wasn't before and investing in a very unproven thing where you have some sort of differential insight. Uh, so right now everybody wants to invest in the next OpenAI and probably the next multi-trillion dollar company will not be another AGI research lab. It will probably be the thing that got built because AGI now existed as a new technology in the same way that when OpenAI was starting, what most people wanted to invest in was like, I don't remember what, another Facebook or something or another cryptocurrency.

And I do have to tell you a funny story. When we invested in OpenAI, it's the only time in 20 years I sent an apology letter for investing to all our LPs. Saying, "I know this looks weird, but we're doing it anyway." I'll only hold that against you a little bit.

That's okay. Anyway, I think it's like, I think it's a tremendously exciting time because the amount of new space that's been opened up is greater than any time I've ever seen before. OpenAI was started, our space had been opened up by a research breakthrough like the crypto market subsidizing GPUs and a lot of new progress there, a handful of other things. But there weren't— I remember that time period, the list of things that I thought were worth working on was very, very short. The list of things I think is worth working on right now is tremendous. And I think that's what people should be going off to fund. Whether the return, whether the research returns continue to accrue to OpenAI or someone else, I think almost as a capital allocator at this point is not the most important question. There's going to be near-free AGI in the world. There will be companies like OpenAI that have, I think tremendous value that gets created because we have like a very large and successful consumer product and a handful of other things. But if we go back to the transistor analogy again, the, there were a few companies that were true transistor companies and then most of them went away and a few survived and now kind of look at a number of Things in this room that have transistors in them, we don't talk about this as a transistor device or a transistor company. It just became like a piece of technology that enabled a whole set of new companies, including OpenAI. And that is what to go for. And again, tremendously exciting time to be a capital allocator. And you should chase the future, not the thing that worked in the past.

So let me go back to the launch of ChatGPT. What surprised you the most? And then based on what you saw in terms of user behavior there, how did your thinking evolve?

Well, I'll rewind back even a little bit more. We— so the tradition in tech companies is you start a product company, you build up the successful product company, and then you decide to like bolt a research lab onto it. And this has worked out sometimes quite well, like Xerox PARC, sometimes quite badly, like some of the more modern companies. But we are the only example I know of where we started with a really well-run research lab and then bolted on a badly run company later. And it was like 4 and a half years into OpenAI. And we hadn't started thinking we were ever going to make products, but eventually it became clear that because of the scaling laws and the amount of capital we needed, we were going to have to go build a really big company. And so you need to have a product to have a company. And we had this model called GPT-3, and we were trying— I was turning up the urgency on the company to figure out a product, and we just couldn't. It wasn't good enough. It was cool, but it wasn't good enough to make something that worked. And I remember that Paul Graham had this advice that always stuck with me, which is you should always make an API. No matter what, you should make an API. It's just, there's like, good stuff will happen. So out of ideas to make a product, we said, well, let's like crowdsource this to the world and we'll put GPT-3 into an API and maybe somebody will figure out something to do with it. The whole world figured out exactly one thing to do with it. The only thing that made money with GPT-3 was these copywriting applications. I actually forget the companies' names now, but there were a few companies that very quickly became billion-plus valuation companies that were just using GPT-3 to generate copy for websites to resell. And— Jasper, things like that. Jasper, yeah. And that was it. People tried other stuff, but no other companies worked. However, We had this thing called the Playground where you could like test prompts and see what you were going to get back. And that was sort of the sleeper hit. People, some people, not a lot, but some people would just chat with that thing all day. And it was before we had figured out RLHF and before we had even GPT-3.5. So it wasn't very good, but there was like clear user signal that people wanted to talk to the models. And given that that was the only thing besides copywriting that had real traction. We said, well, maybe like this is just the product we should build. So we started doing some research to make the models easier to chat with and you didn't have to do these like complicated prompts. Um, and we had a better model and then we decided we would put out a chat interface. We were planning to have to build some other stuff. There was like a big concern inside of OpenAI, which is if you can chat, does it have to be a specific kind of assistant trying to help you learn something or trying to help you accomplish some goal? Because If you're just chatting about nothing, are people going to want to do that or is that going to just feel like you don't know what to say? And so we almost held it, but we put out this preview where you could just chat about whatever. And another interesting learning is we had this like testing group of it and very few of the users, this is back when the model was really bad, very few of the users actually stuck with it. Retention was atrocious. But for the users that did retain, their usage increased over time. And again, we almost didn't launch it because of that. And an important learning that I've later reflected on is if you have a product that has any retention at all, you're actually in really good shape. If it's 5%, that can be totally fine. The default is almost always all the way down straight line to zero. But I didn't quite understand that as intuitively at the time.

So And I don't know what numbers you use publicly for the number of ChatGPT users today. What's your vision for ChatGPT going forward?

The vision for OpenAI is that we would like to build a small suite of products and a platform that you can use with any other service to be sort of your default personal AGI. This is a system that will get to know you, that will be connected to your stuff, that will behave the way you want. And if you want to use it as a— through a chat interface, fine. If you want to go use social products or entertainment products in a new way, great. If you want to get a lot of work done with agents and a bunch of other stuff, also great. And if you want to be able to just log into any other service and sort of bring your intelligence there, We'll have that too. Over time, that'll expand to new kinds of services. I think there's like a really important new kinds of computers to build, and we want our users to have those too. But, you know, I think people are going to have a very important relationship with an AI that will help them be more productive, better, happier, whatever, just smoother life. And we'd like to be that.

So it's more than just intelligence at your beck and call?

Well, I mean, at some level that is what we're selling, but there's a lot of value in creating great experiences and integrating that in different ways.

And so if you look at the next billion users for ChatGPT, is it the same use case? Is it expanded in some way? Is it Lots of different segments. What do you imagine?

I don't like go through our product roadmap, but I think like we, we are still at the terminal stage of AI. And I actually happen to like terminals. I think that was like a nice way to use computers. But as we go build the rest of the modern computer interface around this, which will not literally be what computers look like, but that kind of equivalent step forward, I think it'll just be easier to use and much more capable and approachable for most people.

Yeah, I've heard you say ChatGPT is an operating system for intelligence.

Not yet. We'd like to get it there.

Yeah.

So what's missing?

I mean, all of it. Like, we're very early. We have a whole—

We are early, but there must be things you think are missing.

There are.

So, yeah, it's okay if you don't want to talk about what's missing and what's coming.

Well, like, that's our product roadmap. Like, you know, we—

yeah. Okay, let's switch a little bit. Obviously, another important area beyond personal use, is enterprise use. Speak to that a little bit.

So, I kind of, I think there's two interesting vectors here. One, this idea of virtual coworkers and the ability to use AI to just make all of the things an enterprise has to do that right now it's limited on better. That one's awesome. And you can already see some examples of this. The coding agent stuff is one where you can see the obvious application to enterprises. What's happening in customer support with AI agents is pretty great. And I think at most enterprises, the proportion of work, proportion of tasks done by AI versus done by people is just going to go up and up and up. And that seems obvious and you know, well on its way. The one that I, the thing I'm more excited about is what does it mean for an enterprise to be able to throw an entire cluster of compute at one really hard problem? A scientific discu— case of scientific discovery is maybe the most exciting one, but there's a lot of other like really hard valuable problems that enterprises have that if they could just solve it, which they can't right now, it would enable something new. And as we have AI systems that can go think super hard about a problem, I think that is fundamentally going to be like in the domain of enterprises and, you know, discover new material, figure out how to like optimize some very difficult, very complicated supply chain. I'm excited to see where that will go.

Yeah, if you look at the shorter term in the enterprise context, we've talked about AI scientists, outside of AI scientists for AI, AI scientists, material science or biology or other areas. And I think you've talked about that layer of capability coming at some point in all these models. Where in the enterprise will it be most disruptive in the short term without new capability? And where will your revenue come from shorter term?

I mean, my guess is that in the short, short term, the AI software engineer will be the most disruptive thing to enterprises. It, like, the companies are investing the most in that the fastest. There's a bunch of, like, quirky reasons why it's a particularly great environment. It so directly translates to a current limiting reagent for most companies and a revenue opportunity. So I would bet that that will be kind of like the big story for the rest of the year. There will be other things too, but like you will just start to see companies who get good at this or teams that get good at this significantly able to outperform other ones.

Um, clearly software development, I think pretty much everybody recognizes, is major area. Major new area. What about traditional enterprise functions beyond software?

Yeah, I think those are— again, this is not a theoretical thing anymore. Like, we hear from companies that are like, I am doing all of my customer support with AI now. I'm doing all of my outbound sales with AI. I'm doing all of my, you know, pick your other vertical. So that's all happening too. I just don't think it'll have like quite the same magnitude.

Yeah.

Yeah. Let me switch to a different topic. You've gone through a lot of decisions at OpenAI in extreme uncertainty around will this technology work? How will it develop? How— what will competitors do? What are some of the hardest things you've— decisions you've made, some that turned out right and others that turned out wrong? Or do you think about that?

For me, the hardest thing is not any individual decision, although there's a lot of hard ones of those, and we've gotten more than our fair share wrong too. It's the number of decisions we have to make and the number of things we have to do at once and the degree to which they all interrelate. And I haven't ever found really good advice on this. And so we're sort of struggling through it. But I think if we could operate at the scale we wanted to operate at and do the number of things we wanted to do, but only have to make like half the decisions, life would feel really easy. But it does feel like we're just oversubscribed there. And I don't know what to do about this, 'cause we really do have to do a lot of things at once. We have to run a great research program, we have to build out massive-scale infrastructure, we have to figure out how to build all these products, like long list of stuff. But it's, I don't think there is good Silicon Valley wisdom for, how to make many decisions quickly. I think there is good advice about how to make a few decisions or a few big decisions, but I haven't figured out how to transform OpenAI yet into an org that only has to make a few big decisions. And that's kind of the thing that I think would most help if we could do that. We have asked ChatGPT, it didn't help. Uh, it—

the best recent use I heard of ChatGPT is somebody said they're a much better executive coach than any executive coach they've used.

The general case of it being really good at like therapy and coaching has surprised a lot of people. Um, it's— it at this point is a significant part of our usage. Yeah, I Maybe the other category of decisions that I think are hard are just the decision, like doing, figuring out how to do something new or deciding to do something new. Like if a thing that we pride ourselves at OpenAI on is that we try to do new things instead of copying competitors or other research labs. So if you look at other research labs, They mostly try to do the same research OpenAI is doing, and they mostly try to exactly clone our products. They look shockingly the same across the industry. And I used to dunk on people and think that that was just some sort of lack of creativity or something. And it is that too, but it's really hard to make the decisions that are that go into doing something new. And if you know something else works, it's so tempting. I now so understand it to just copy that and not have to make these decisions. I don't think it's how the world gets pushed forward, but I really do empathize now with why people do it.

So there's the research labs, and there's a lot of copying each other and building on each other's ideas. A lot of the entrepreneurs are making decisions in extreme uncertainty on what the models will have and how to make decisions in extreme uncertainty. And it may not be the quantity of decisions, but they're really critical strategic decisions. Any advice for them on how to handle this uncertainty? Because that's a question I get probably more than any other question. Where will AI capability go and what assumptions should I make and where— how do I build my startup on top of that, the current models?

I mean, the thing that's worked historically is just assume that the models are going to get better in every way, every dimension. Better, cheaper, like more better multimodality, better intelligence, better reasoning capability, lower costs. And I don't see any reason to believe that's going to slow down in the next few years. I think there will be some temporary slowdowns where people, you know, rewrite their infrastructure or wait for a new cluster to come up or whatever. But on the whole, I think you'll just be well served to not try to outsmart yourself, not try to say, well, is it going to get better at this little thing or that one or exactly how to time it and when. And just, just like ask yourself the question of, okay, an oracle tells me that this model is going to get roughly 10x better every year on roughly every dimension. What product should I build when? And that'll be pretty good high-level guidance.

So for the people who are building on top of your models or Anthropic or Google's models, what do you think is most misunderstood? What is it you wish people understood? The entrepreneurs who are building on top, not the people who are trying to compete with you.

I'm super impressed by the people building AI applications at this point. I think they understand so much about the models and exactly what they can do and what they can't and how to, you know, get the most performance out of them. I don't I don't have like a wishlist of what else I'd like them to understand. It's, there are, um, I had a dinner at my house recently where I tried to invite like, you know, the 20, like 20 of the most impressive like new generation AI founders I had heard about or met. And like, I was just going around the table and I was like, man, these are like such better founders than the world had 5 or 10 years ago. Like, I don't really have a wish list here. I'm super impressed.

I want to talk about some of the global implications of AI, but before I go there, I want to ask you my favorite question, and I want to leave enough time for audience questions too. When do you think we get a 10-person company that has $1 billion in revenue? Has that started? Will start soon? Not going to be possible? Obviously, I have my bias.

I would bet the company has either already started or will start in the next couple of years, few years.

Yeah, I completely— I've been speculating it probably already has started. It's a pretty amazing thought to think about how you create value and what it takes to create value.

This is why I think the AI for science stuff is like so interesting. You could imagine a single drug that generates more than a billion dollars of revenue that could be generated, that could be discovered and taken through trials by one person plus, you know, 50,000 GPUs or something.

Yeah, I just saw an article, somebody sent me an article this morning in my email, which I haven't had time to read, said that using AI only discovered a cure for macular degeneration. So your point is right, but the same could happen in entertainment and a number of other areas. Uh, let's go to sort of the global rollout and how do we ensure benefits of AI spread more globally, more equally? So there's, there's sort of this worry of the rich get richer because then the inside circle— the situational awareness paper talks about there's only a few hundred people who understand how large this change is and how fast it's happening. But talk about the global implications and how we ensure equity and broader usage and spread of the benefits? First, within even this country, because there's an asymmetry to it, and then globally.

So I don't want to be like all— I resent the question, but I don't want to be all negative on it because I think there is something important to it. But You know, ChatGPT is, I think, the 5th biggest website in the world now. It is, if it stays on its current trajectory, which is hard to do, it will be the biggest website in the world. There will be billions of people that are using free AGI at some point. Everyone will have access to great medical advice. Everyone will have access to great education. Everybody will have the ability to like ask for any software to be made and for free, we will just do it for you. And that is like, that is how technology works. And I think capitalism is like really great. It does not mean we don't have to correct around the edges. I think we do. I think there's important stuff that we will need to address differently. But on the whole, the way this benefits the world is that technology benefits the world. You put tools in the hands of people, you make them free or low cost, and people do amazing things. And I think like there are a lot of people who I think there are a lot of people who think that, oh, you know, the world is not ready for this, they can't handle it, there's only a few people that can understand it, there's that paper, there's many others. But I think people really kind of do know what they need and they're pretty good about learning how to use new technology. And it's— and again, this is not like theoretical, this is like at mass scale happening today. And I think it's quite wonderful. And it's happening because like the existing systems and incentives in the world are phenomenal in many ways. Again, there will be some things we have to do differently. If AI starts making massive scientific discoveries, you know, some company will use AI to discover cures for every cancer. And I hope those people get really rich, but also everybody else in the world, I hope they get a cheap cure for cancer. Someone uses AI to make fusion commercializable. I hope those people get really rich, but the whole world will benefit from plummeting electricity prices. So, I kind of just think like technology on the whole does a lot to spread massive global benefit. And the story of the last couple of hundred years has been that. And we should not be ashamed of that. We should not like try to explain away why what we do is not so evil. Like it's really been great. There will be some things that I think are different about AI. You can imagine a world where compute becomes a crazy scarce resource and deciding, somehow democratically deciding what problems we go after in what order with massive amounts of compute needs a little bit of a different approach than we've had in the past. You can imagine access to compute for, you know, you could imagine like most capital in the world decides it wants to like have access to compute and that becomes like a crazy expensive, crazy limited thing. That would be bad, but my answer about how to solve that, at least the first try on how to solve it, would be just like make way, way more compute. I think this— gone back and forth, so I'm not confident, but I think this should be a force towards more equality in the world, not further divergence.

In the interest of time, I'm going to put 3 questions on the table. You can answer any one of them or all of them. Okay. Role of government, what I expect AI will cause is a hugely deflationary economy in the 2030s, and the risk of sentient AI. You pick which ones or all of them you want to address.

Well, how about I start with deflation and you tell me when I've gone on too long and aside from the other ones. I think it should be hugely— I hope it's deflationary. I think it should be. But all of the excess wealth that's gonna get created will need to go somewhere. And I hope that things like water and food and healthcare and education and access to nature and time with family and whatever else, I hope that becomes something that like, everyone can afford, and then that we all find the silliest status games to play and go bid up, you know, Da Vinci paintings to a trillion dollars or whatever, or galaxies to a quadrillion dollars, to make it so that people who like want to play status games and want to create value for other people and want to have something for bragging rights basically get to go do that. But I think it is an interesting question about like, if we want everything to deflate wildly, if we want everything to get super cheap and people are wildly ambitious and creative and like being useful to each other and gonna work hard and we're gonna have some way to keep track of that, where is the excess wealth gonna go? And I think that's like a very interesting design problem.

Yeah, no, it is. And I'm sort of a huge, believer in how deflationary the economy will be. And if it's a good measure of GDP, which I don't think it will be, how we measure GDP growth, which I think will explode. So you'll see these two things happening at the same time. But we'll go to audience questions. Can we get a mic here?

Hey Sam. So most of us here at this summit have been talking about AI as an enabler for our business. And I wanna flip it for a second and think about it as a potential threat, enabling our competitors to build competing models, better models faster, leveling the playing field or even tilting it worse than that, tilting it against us. How would you? How would you think about that?

I'm not sure I understand the difference. Like, I think it's gonna enable someone and disadvantage somebody else, and I think you just have to work to be on the enabled side of that. But I think both of those things are true, but I'm not sure how to act differently. Like, yeah, all of your competitors and new ones are gonna try to like outgun you with AI and all you can do is do a better job.

Another version of that question is, are competitive modes even possible in a world of superintelligence?

I don't know. I mean, yes, the answer is of course they are, but I don't know what they'll look like. I think some of them will look about the same, like network effects, brands. I can imagine those things looking a lot the same. There's a lot of others that I feel certain won't exist. There's new ones that I imagine will come. But this is like, I think this is the fun part of business. Yeah. Like, it's great to go figure out what those new modes are going to look like.

Question over there.

Hi, Sam.

Hey, Vinod Khosla. I think you said something pretty interesting in terms of there will be a scientific problem that can throw 50,000 GPUs. I think fusion is of interest for a number of us here. But just speculating, physical world can't scale up as quickly, but when would you flip a switch and throw 50,000 GPUs at a problem like Fusion?

Probably right now. I think you want to be a little ahead of time because you'll start to learn the contours of this. I think this is the right time to start some big high-compute scientific projects. I don't know if that's the exact one to start with, but Directionally, yes.

Hi.

Similar to Fusion on the energy side, how do you think about energy infrastructure in the world of AI? I was in a conference in Switzerland in January, and I heard Eric Schmidt talk. And he was saying how Europe would never be big in AI because of how high the energy costs were. You need all this compute power. And then DeepSeq came out, and it was much more efficient. Do you see a change in your hypothesis around where you might be going as you develop these data centers to meet the energy needs for your technologies?

I mean, I think both things are true. I think the models will get hugely more efficient, and also there's so much demand in the world, we will need many gigawatts, tens of gigawatts of eventually probably hundreds of gigawatts of energy to run the world's sort of AI demand, even with super efficient models. But I'm like very quite optimistic about new energy sources coming online and I think it'll be abundant and cheap. And I think AI and energy are obviously tremendously related things. Energy is valuable for other reasons too. You want the ability to go make stuff happen in the world, but The degree to which 10 years from now I expect the cost of AI to converge to the cost of electricity is like, it's hard to overstate. Everything else ought to get really cheap. We can go roboticize making chips, but you still need to drive the electrons.

Question here, and can we get a mic over there also? So thank you, Sam.

Thank you, Vinod.

I want to go back to Vinod's third question. What is the role of government? Well, on that question of AI infrastructure, I think making sure that we can build enough AI such that it is abundant and not something that just gets all of the availability sucked up by rich people, that seems like an important thing for governments. There will need to be some guardrails here. Overregulation, I think, we can all see how that could be a disaster. Under-regulation for this could be pretty bad in its own way. So setting some global rules, especially as the systems get much more powerful, I think that can only come from governments. I think the companies will propose it and figure out what it should be, but I think it can only be enforced globally by the governments. I think there will come some questions about how distribution and access is shared. There'll be some big questions about how we think about data in this kind of a world that governments will have to weigh in on. I would kind of expect this to be like, by say the US 2028 election, one of the very major topics for politics.

Question over there.

Hi, Sam. So two questions. One on the bandwidth of the interface. So it seems like as more and more agents become more and more powerful, powerful to do different things. I probably wanna get to a point where I can talk to thousands of agents at once. The problem is that if I'm shifting interface over and over and like waiting for them to get back with reports, I'm probably not doing anything else but just dealing with agents, doing one thing or the other. However, the other problem is that I cannot let them go hands-off and they're gonna do things differently than what my exact intention would be that's available. Like that, that understanding becomes clear once you even like make it to write a PRD where it obviously doesn't get all of it what I want and like there's a lot of editing involved. So the bandwidth is what is causing a scaling issue of the number of agents that I can deal with. So how do you think about that? What are the next steps here?

I mean, there will be some agent that's job is like to output to evaluate the outputs of other agents and decide what to surface to you when and when it can make a decision or understand the decision you would make. So that'll be another kind of job for a different kind of agent. And I bet it'll be pretty effective. There is probably like, you know, you'll have a custom-generated UI for you that'll be evolving in real time based off the output of those agents and it'll do a good job of figuring out how to get your attention when, but I think a lot of the problems of today will not be problems of tomorrow because the AI just keeps getting smarter and can work at higher and higher levels of the executive function there.

Was there another question there? Yeah.

Hi, Sam.

Micha.

Sorry. Micha from SOMITE AI. First of all, congrats on bringing Fiji Simo on board. She's an advisor to ours and one of the most amazing leaders I know. At SOMITE, we build foundation models for stem cell differentiation, human stem cells, in order to bring forward novel, novel cures, novel cell replacement therapies with a new modality of data. And my question to you is, when you think through the breakthroughs that are coming to the physical world, to biology, and so on and so forth, where is the bottleneck? Is it gonna— and where's the unlock gonna come from? Is it going to come from generic models? Does it come from domain-specific models? Is going to come from the data itself. Where do you see the breakthroughs or the real first breakthrough in biology for bringing novel cures in?

Obviously, I don't really know. There are— I think there are some areas of science where with no more data, a smarter intelligence can just make new breakthroughs. I think it's possible in physics, for example. That without a single additional data point, without ever building another particle accelerator, AI could just solve physics. I don't think it's super likely, but maybe all the data we need is there and we just have not been, we've not had the ability to solve a hard enough integral or whatever. And then there are areas like biology where I'd say it feels way less likely to me that without ever running a single additional wet lab experiment, we could cure every cancer. Not impossible, but unlikely. And I think we'll have to build like new systems that are— it doesn't have to be an automated wet lab. It's fine if it just sends an email to a scientist saying, can you run this experiment for me and tell me the answer? But somehow like active learning systems that can keep getting new data points. But what problem will fit into what category and exactly where the limiters will be, you know, we'll just We'll take an empirical approach to this.

Well, thank you, Sam. I think we'll stop here.

Thank you for having me.

Yeah, thanks very much for coming.