AI_Explorer - Sam Altman (CEO of OpenAI): "You no longer need to write prompts." ...

You with like an affordable amount of spend on tokens, you can do what a hundred person, incredibly great engineering team would do as a startup. And that was just totally impossible. That was like not in the set of options for a startup. And now it is so, so I think what you can take on, uh, the level of ambition you can have, the speed of what you can move, the amount of stuff you can do at once, uh, it's just totally different. And, um, does that change the shape of the problems you feel like you'd assign at the end of the class for people to attack, you know, at the end of that quarter, if you were teaching it again, I don't think assigning problems to attack ever works because if you like, if I can think of a problem, if I can think of like a really great startup idea, uh, if it's like obvious enough to me, uh, then it's probably obvious to a lot of people when we started open AI, we were, we were like the, uh, you know, one of maybe generously speaking for AGI efforts in the world, right. And you want to find something like that. And I'm sure that there exists something today that just wasn't possible at all. Pre like automated coding era, uh, that is totally not obvious. That will be, you know, a multi-trillion dollar market soon. Uh, and that only four companies are working on right now, but I don't know what that is. It's much more likely you all know what that is. Then I know what that is. I just, you know, my brain is like taken over by open AI. Um, but you know, the kind of idea someone can assign you to work on is probably not what you want. Yep. Um, okay. So that that's fair. Um, but I think it'd be helpful since as a systems class to maybe, uh, reason about a particular problem that you have to reason through so that they can then apply the shape of the techniques used to break down from a systems perspective, that problem into solutions to their own problem. Yeah. Um, and a, and a concept that, uh, you had started to tease in the class, you know, back in 2014, and then, uh, clearly you've talked about publicly over the years is, um, scale, right? Scale is its own beast. It's, it's, it's, you know, quantity is its own quality. Scale as a concept has been something. It seems like you've, um, empirically investigated and all kinds of ways over the last 10 years. So, um, could you help on help us first unpack like what you mean by scale now, 10 years later? How would you deconstruct that as a systems design, uh, attribute to apply whether it's a, as, as a tool? Um, can we start there? Yes. Uh, so I don't know why the following observation is true. I offer no theory that I find satisfying to explain it. And that makes me a little bit nervous to suggest you follow it, but I'm going to anyway, because empirically it does seem to be true, which is all of the most interesting things I have observed in my career in watching other, uh, things happen, all of the most interesting ones, uh, have had something to do with emergent properties that scale or scale continuing to provide returns far beyond what the consensus thinks will work. And this obviously happens with like scaling loss for AI models. Um, but this happens with, uh, you know, getting more smart people together to think about one problem. This ha in a, in a research setting. Um, this happens with, uh, companies and the sort of economy of scale. You can get all the, in all these different ways. I really learned this at Y Combinator when, uh, it became clear to me that everybody was saying, oh, Y Combinator has gotten too big, it should shrink. We should fund less companies per batch. You know, the best times of Y Combinator when it was like 10 companies per batch. And a lot of like very smart people were saying this and, and it was like tempting because it would have been like much less work and the theory was that, you know, the best companies are always kind of obvious and then you fund the rest and it's not as helpful. Um, but a huge part of the magic of what made YC work were, uh, was the sort of the network effects inside of the batch. And that was an emergent property at scale that just hadn't been discovered before, no one had tried to fund startups at scale in the same way. And, and thus no one had ever happened upon this observation of when you do that. Um, there's, there's something important that happens that just didn't exist at all at the one hand, the one 100th of the scale, there's a bunch of other examples like this, uh, I, and I'll skip them in the interest of time, but I would say again, I offer no explanation for why, but empirically speaking, when you find a time that you can push on, you can push something to a scale people have not tried before, and it's already working in some interesting way at the smaller scale more often than not, that seems to be a good idea. And it also seems to be something that most people don't do enough. And I don't have offer an explanation for this either, but like in, you know, when we were like, we're really going to scale AR models, um, all of the like geniuses in the field, most of them were, Oh, this isn't really working. You know, that's, that's barely a scientific result. It's not interesting that it gets better at scale. You've already shown that. Why keep scaling it? I mentioned the YC example. Um, I've seen a lot of startup founders where they're like, well, you know, there might be something interesting that would happen if I scaled this up, but I'm a little worried about it for non-specific reasons. And again, looking back at like a huge data set of people that have scaled their companies in all these different ways, there's almost always interesting stuff there. So I think directionally, that's like an interesting thing to push on and severely underexplored. Um, on the systems design part of that, uh, I think one reason people don't do it as much is stuff breaks, uh, at an accelerating rate and in an unpredictable way as you scale it. And if you are going to really scale something, um, it's always like a little bit broken. There are always like very smart people who say why you shouldn't do this. You know, don't get too ambitious. Don't get too big. Let's try this smaller. And so breaking that down as a systems problem, I'll use the thing of when we were like scaling up AI models, there was technically, can we do this at all? This seems crazy. Like no one had ever thought about trying to do a run across 10,000 or a hundred thousand GPUs, and that was going to require stacks of engineering talent. Um, there was the capital requirements and what it was going to take to do this. And like, how is there ever going to be a business? How can you think about taking this risk? Uh, there was this sort of like cultural stuff of researchers saying, well, if we're going to get all this compute, why do we put it all into this one project? But we're not going to learn something. Why not divide it up among all these, all these projects. And this also happens in kind of every area I've looked at. Almost every area for scale and breaking it down into the sort of each difficult area or each reason not to do it and trying to address them one at a time. Yeah. That's been really important. Um, I'm going to push on that a little bit because there's very few people who've been able to sort of repeatedly scale new products and systems the way. Uh, the OpenAI team has over the years, but it seems like one of the issues is there are all these prior conditioning sort of mental models and expectations humans have. And you said things break. And one of the things it seems often breaks that's the hardest to refactor is, is human, the human side of the system's design, right? Wherever there's human implementers or there's a human participants in that. And so what have you learned about humans at scale, like organizing humans at scale to participate in a system that may not be, uh, like just a redo of some past system that they, they get naively on at a priority on first blush. Um, I think like clear, a clear goal, a clear plan to get there. Uh, and like a clear answer to the way that you're going to get there and kind of how you're going to make decisions along the way that's, that's very important. So, um, you know, if we go back to the example of when we decided to scale up models, there were a lot of people who are like, ah, this isn't really going to work, it's going to have these problems. It's also not, you know, we need a more diversified portfolio, but once we say no, we're going to make a bet on scaling, deep learning, like that's our thing. If we're wrong, we'll fail, but we're going to do that. Here's why we're going to do that. Here's what we believe about what the state of the world can be like if we get there. Uh, that's very powerful. And then for whatever reason, um, we did not evolve to be good at thinking about exponentials. People have a hard time imagining that scaling laws are going to continue exponentially, that revenue will grow exponentially, that an organization can take on exponential complexity. And in my experience, it takes a lot of time to really reason through first principles with people about why that can happen. Can we take two examples to walk through that? The first being chatGPT and the second being Codex, you know, both of these have transformed. Can everyone hear? I'm going to try to project it. Yeah. Okay. Um, so let, let me put in a frame and you can challenge both the assumption and then we can hopefully reason to example of what happened in the case of chatGPT, you know, for a long time in scaling of models, a big mental block that seemed to be prevalent in the space is what are these things going to be useful for? This is, you know, it's a research, uh, sort of solution, uh, solution chasing a problem, research first approach. It's not a product. Um, and then, you know, chatGPT came out and it proved to the world that, you know, that chat experience was a killer app for general models, um, at scale for consumers. And then a couple of years later, it's clear that coding has been a killer enterprise app. So w w how would you compare and contrast the systems you guys used to discover those use cases, ship them, scale them, monetize them, any, any salient learnings from those two systems? Yes. Um, so we had made GPT-3 and we needed to make money cause we wanted to go scale up to, you know, a billion and multi-billion dollar computers. And we had GPT-3 and it was kind of interesting. It was a cool demo, but we couldn't figure out a product to build around it. And we had been thinking, thinking, we just couldn't do it. We had tried a few things. They, they hadn't worked. Um, and so we knew the models were going to get better, but we also wanted to like start a revenue engine sooner. And we said, well, since we can't figure out what product to build, we're just going to put this into an API and we're going to hope that somebody else can figure out what product to build. And so we launched in like, I don't know, something in the summer of 2020, the GPT-3 API. And initially it kind of got no traction at all. And then about a month later, randomly, as far as we can tell, it went viral on Twitter on the same day, a few different developers kind of found, got it to do something cool, posted it. Other people started trying. And, and then like a lot of people started trying the API. Um, but it was shockingly bad. If you go back and use GPT-3 or 3.5, um, you will be astonished at how bad the models were then, uh, relative to the amount of excitement they generated at the time. Uh, so people tried all of these things and really the only business that people got to work in a significant way with GPT-3 was copywriting. Um, and that was like not that great and not that exciting. And we were kind of like, you know, ah, it's just gonna have to wait for a better model. But although that was the only business that was working, developers had figured out how to like put in a prompt and get and be able to chat with it. And we saw this a lot, like more people were using, they couldn't get the API to work for their business, but they were using their API key to just chat. And we said, well, we can build a good chatbot. People clearly want that. And we had a new model. We actually had GPT-4 done, but we had a new model. We were ready to release in between called 3.5 and we had figured out a new kind of post-training where we can get the models to do like a good job with instruction following. So it can make it easier to chat with. And we said, well, you know, the API is not working great. Maybe it was like a 10 or a $20 million run rate kind of business. But there is this thing that people love, uh, and under the YC principle of see what your users love and do that. We said, well, we'll build a chatbot around it. And we put that out and we still didn't think it was going to do that. Well, uh, there was, it was really meant as like a research demo, uh, to convince other people that they should build chat like products and pay us for the API. But that went like crazy viral. And another thing I had learned from YC is when something really starts growing and it's not very good, you have like a guaranteed hit on your hands. And so we had like five days where the traffic would shoot up, fall off and everybody be like, well, that was just a hype cycle. But then the next day it would get to a higher peak fall off again. Later in the day, people would say that's a hype cycle. By the fourth or fifth day, I was like, I know how this works. I know what's going to happen. Like we have the potential here at a killer product. Um, and we knew we could make it much better. We knew we could, we knew we had GPT-4. We knew we could keep scaling. Um, but by that fifth day, we got everybody together and said, this is an emergency. This is a good kind of emergency, but we have to build a company. And the product all at once. Uh, we then had like two months of crazy scaling. Uh, and then we said, you know, we have to figure out a business model later. For now, we're just going to charge people so that we don't like run out our compute bills. But that's obviously not the long-term answer. That also turned out just to work. Um, and that was the story of ChatGT. And then there was so much utility that people just had not gotten over the activation energy to find that that has worked really well. Um, and then Codex, actually the plan before ChatGT was that we were going to go all in on code. Um, we knew these models could write code. Uh, we knew that they could be really, and we knew that that would be like a valuable area, but then we had this incredibly exciting thing happen. Um, but our kind of internal belief at the time was that coding was how these models would control things on computers. And robots were how these models would control things in the physical world. And if you made a smart enough model that had sort of the actuators of writing code and robot and driving a robot, you could then kind of actually get this intelligence to do stuff for you in the world. So, uh, then it took us a while to get there. And then I think Codex got really good by early this year. But with 5.5 is when we saw this real inflection point where people are now like doing just incredible things with it. And, um, you know, that early in the class, we've talked about how the capabilities pipeline, uh, is starting to look, it's starting to become somewhat more legibly standard across different research groups. You've got, you know, pre-training, mid-training, post-training, then you've got the RL and supervised feedback loop. Is that, do you think that's roughly like the shape of the pipeline that allowed Codex to, you know, go to a capability jump and that will basically stay stable now and consistent, or are we going to go through a major rewrite of that pipeline? I think that is definitely the current pipeline. I expect we will go through a major rewrite. I don't know when it'll happen or exactly how. Um, but it is a little odd to me that it's so happens as a pipeline and it doesn't quite feel like the optimal solution, um, what would be an optimal solution in your head? I think that's a research problem for the AIs to figure out. Um, I think we're at a point where we've set this goal that by September of this year, we will use 500,000 A100 equivalent GPUs, like a lot of computing power as an AI research intern. And by March of 2028, that we will have a full end to end, very talented researcher, like figuring out complete new architectures. Um, so I think we are going to get like with the current pipeline, the current architectures, I think we're going to get over the line of when AIs can do incredible, incredible work. Um, you know, one of the things that you just described there is you, we've been talking a lot in the class about systems, frameworks, and analogies to make concepts from one domain legible to other people who may not have all the context in another, and that sometimes because of the translation problem, you know, reasoning by analogy is not helpful because then errors compound. Um, right there, you said, you know, our goal is to try to use it as an AI intern, which obviously is a very useful metaphor within the context of, you know, Silicon Valley, a class that understands how these pipelines work and so on, and then as, as you scale, actually that metaphor globally, people who might not have all that context go start analogizing these models in ways that they shouldn't be. Like, how should we think about the limits of, of that, of, of what are the limits to scale of, um, what are the product analogies, the research analogies you find most useful within the Valley and which one of the, what have you found about, found about the limits of those analogies scaling? And now how do you navigate between those two problems? I've been very interested in studying how, but I, I think what is happening is we are, we are in the process of creating a new utility. This doesn't happen very often. You know, electricity is utility. Internet's a utility. There's water, I guess. There's not a lot of these. Uh, and so there are not a lot of examples that we can study for good metaphors or learnings about how to explain this to the world. Um, but I was recently looking at what happened when electricity became a utility. And it's a good analogy for many reasons. It's imperfect of course, too, but the electricity companies, at least the ones I could find information about, they didn't talk about selling electricity because no one knew what that was or why they wanted, it sounds like very scary. It's this thing that's like going to come into your house and it can kill you. And it's like gruesome way. And you know, it feels sort of like very different than the world before. Uh, and maybe they tried to sell electricity or market electricity first. I don't know, but in any case that didn't work. And then what they started marketing, selling to people was light at night. You know, we are gonna, what you are getting from us is not electricity. It's light at night. By the way, you can use the same thing that lets you get light for all these other things. But people are like, well, why would I want that? And I'm like, well, you know, it'll wash your clothes for you someday. And no, no, it won't. I can't, that's too far of a jump for me. Um, so I don't know what our analogy for this should be. Um, but I suspect that even if, even if we're totally right and intelligence is going to become this new utility that every company, every customer, every government just needs access to, and is going to use in all sorts of incredible ways, and you will have like a open AI token subscription that you will plug into everything and use to access everything and you have running for you all the time and doing this amazing stuff. I kind of don't think at least right now, the right way for us to analogize that is we're selling intelligence because people are just like somehow not resonating. I don't know what our equivalent of we're selling you light at night is going to be, but I think if we're going to become a new utility, we need to find a way to explain to the world what it means to have this like intelligence pike that you can just do whatever you'd like with it. So, um, one question that has emerged, an emergent property of this class is of, of having a diversity of different speakers is that the utility analogy has come up several times, but in reference to different things. So Jensen likened you to like compute to a utility. Um, and why there should be access and so on and talk about how Stanford should pull budget and so on and, and, and procure that as a utility for everybody on campus, whereas you just likened the intelligence part to both of these things. True. Is one of them true? One is one more likely to be true. How should people reason about compute as a utility versus tokens as a utility? And by compute, I mean here chips versus tokens. Does that, does that make sense? I think as a consumer, as like a business or an individual, um, you will think in something closer to tokens or probably even one level up from tokens. I don't think you'll care very much about, you know, where the hardware is, what particular chip it is, what's powering it. I think that stuff will be abstracted out. And what you will care about is when you're interacting with the system. Um, can you use it a lot? Is it cheap? Is it doing a good job? Um, so right now it's like tokens. It may get, as we move into a world where we all just have like this constant agent running for us, being useful to us all the time. Um, you may think about it as even one level up, but yeah, my, my guess is, is you, when you like pay for your cell phone bill, you're like, all right, I'm buying access to airtime and some number of gigabytes and, you know, it's going to do all these things and I'll use all these apps and whatever else. But like what you think about paying for the kind of internet utility in this case is just like access to the whole system and the particular hardware at the base station, how it connects to the internet, you don't think about that as much. Um, I know I could nerd out about utility infrastructure for a long time, but I want to make sure we switch a little bit to being relevant for the students. Usually we have, uh, questions where we're not hearing those today, unless you're comfortable with it. Oh, okay. Great. How about that? Improved. Okay. Uh, so one final question to start getting the creative juices flowing is, um, the final project for this class, partly in private 183 is the one person frontier lab. So everybody here is working on projects where they're simulating being an individual, uh, as a lab with access to all the right tools. They've got hundreds of thousands of dollars of credits from CloudFlare. I think we've got some open AI tokens, maybe. But there's a bunch of compute at their disposal. Um, what would you, if you were in the class, what would you be working on for your one person frontier lab project? First of all, I think that's an awesome project. Um, I think this is top of mind because, uh, you, we were just like talking about utility frame frameworks. I think there's a lot of very smart people working on, uh, great training ideas, and we're going to have incredible models, no matter what you all do. We're going to have incredible models. I promise you. Uh, like pretty quickly. I think we have not invested enough in being able to deliver at scale, huge amounts of cheap intelligence. So maybe I would go work on like the inference part of the stack and how are we going to get this incredible intelligence to be cheap in a moment? Uh, I think that's under invested in, and I think all of the frontier labs are going to have to become inference companies to a significant degree. Um, okay. Might be too late to pivot your projects, but better late than never. Work on whatever you want to work on. Uh, okay. Let's start taking questions and I'm going to moderate and try to be not, you know, please try to be productive and not spicy, et cetera. Remember it's a CS class, but up to you, Sam. Oh, we've got questions. Oh, perfect. All right. First one. There are questions about your views on beyond the guns view that elements are a dead end. Um, first of all, in terms of achieving human level intelligence, these models have already far surpassed human intelligence in some ways, and then they're wildly worse than others. Like for example, they seem much worse than people are at very long horizon kind of high judgment signal and tasks. Um, on the other hand, yesterday we had one of our models, uh, discover or disprove a conjecture, one of the Erdos problems that had smart people have worked on for a long time. And a lot of people, a lot of smart scientists, I don't know if Lacuna was one of them or not. Had even quite recently said something like that was not going to happen. Uh, and then like the model just did it. And you know, now you have all these mathematicians saying like, is math over, what does this mean for our field? So clearly LLMs are capable of figuring out new knowledge. And clearly they are capable of doing some things that some intelligence tasks that humans just can't do. Um, they are going to scale much further. So how much better and what distribution of the tasks they can do better than humans we'll find out, but I suspect it's a lot and the, you know, in terms of this, like lack of a belief in the exponential we were talking about earlier, um, I think the field was honestly held back by a generation of scientists who just were way too certain on what wouldn't, what, what scaling was not going to produce. And then some people just looked at the graphs and said, well, it looks like it's continuing beautifully. Let's keep going. Um, I think world models are clearly important and to, we'll need that for things like robotics, uh, but betting against LLM scaling at this point, uh, feels quite misguided to me. It doesn't get annoying to be the, I told you so guy. I mean, there are these like Twitter trolls that, you know, for years have just been like, it's not going to work. It's not going to work. This is so dumb. Like, you know, this is a fraud. This company is going to fail. This research approach is going to fail. And I used to get more bothered by them, but I don't even like feel the, I told you so at this point. It's like, you were just, you're still going on about it. Like the data is quite strong on our side. And I don't think it'd be that fun to say, I told you so. And also the fact that you're like still saying we're wrong. It doesn't really bother me. I think there's that kind of move on. There's that saying that like insanity is doing the same thing over and over again when presented with data that is not working and if they keep repeating that in a sense, it's, it's, it's a form of insanity. I think, I think there's something that happens, which is if you make your identity about a particular thing is going to work or not work and you associate yourself with that belief and then the science or the empirical results disprove you and you're like too hung up on your identity, you can't let it go. You can't see the truth. And I think this is like a important reminder in both directions. Yeah.