Keynote Discussion | Sam Altman and Vinod Khosla

Vinod Khosla.

I finished with OpenAI. I'd like to start with your goals for OpenAI.

Great.

First, first, tell people what you're doing. Some people may not be totally familiar with it.

Okay. So unlike most companies trying to build artificial intelligence where people take basically deep learning and apply it to narrow ideas and it works sometimes astonishingly well, usually at least pretty well. We are trying to build one intelligence that is smarter and more capable than humans in every way. We're trying to use this to solve the biggest problems facing the world. We're trying to use this— I think we will enable hundreds of billions, trillions of dollars of new businesses because we'll have such powerful artificial intelligence. And someday I think we will sort of build the descendants of humanity and launch them off to colonize the universe. We're trying to do this in a way where we make it go really well for humanity, where the value we create, a large part of it gets shared with the world. This is just such a fundamental technology that we have to design a new corporate structure so that our investors can make a great return and our employees too, but also that we figure out a way that the whole world wins. I think this is different. If we're successful in this quest, and it's very hard, but if we're successful, I think it will be the most significant technological transformation in human history. I think it will eclipse the Agricultural Revolution, the Industrial Revolution, the Internet Revolution all put together. And so we're trying to think about how we want that world to go. The fundamental driver of all of this has been this incredible increase in computing power and a few big breakthroughs in algorithms. Every year for the last the biggest neural network the industry can train has grown by a factor of 10. That'll continue for the next 6 at least, at which point we'll be near one human brain.

Great. Best guess, where do you think we'll be in 5 years or 10 years?

I think in, let's say, 10 years, we will have systems that sound today to be impossible. I think they'll be smarter than humans in every way. I think we will— machines will actually feel to most people like they can think subjectively. I think that we will be able to create— I think it'll be a number of businesses started, hopefully some spun out of OpenAI, that are— look like they're on a path to be bigger than any corporation that currently exists in the world. I think we will have systems that we can talk to in natural language that do complicated tasks. An example I like that really sort of strikes me is when I see a very small child like a 1.5 or a 2-year-old or even less sometimes pick up a magazine and try to treat it like an iPad with touch gestures because that's how they think the world is supposed to work. And I think that shows how quickly sort of we adapt to a new world. And I think that 10 years from now, or you sort of, you know, see these kids try to like talk to everything like it's an Alexa. I think 10 years from now, children that are born then will assume that all systems are actually intelligent. They will treat computers and humans very similarly.

Well, that's exciting. Let's say you being optimistic.

Yeah.

Um, and I say let's narrow the definition of AGI.

Yep.

So we don't need an AGI philosopher or AGI Steven Pinker. We talk— if we talk only about the economically valuable functions humans do, work on assembly lines, be an oncologist, be a doctor, how much easier does the problem get? How much more certain do you get we'll get there in 10 years?

Yeah. So, I think more than 50% of work today is repetitive, non-creative work that does not require a deep emotional connection like you want with in some scenarios. And in those cases, that my confidence interval, let's say out 20 years, is extremely high that we'll be able to do all of that. That gets much simpler. We don't need any more technological breakthroughs for that. We just need system scaling. The work that we've already done at OpenAI, just as that sort of propagates out into the world and we figure out how we're going to build business models around that and other people do too, that alone I think is enough. To get there. I think the easiest layup right now in all of startup investing is to take narrow AI and apply it to every vertical and just do what humans do and generally don't even like doing. It's not the exciting part of work. It's not the creative fire better than humans can. So I think this will be the biggest trend in startups in the next decade, let's say.

Yeah. So Sundar Pichai said the last 10 years was about building everything mobile first. The next 10 years will be about building everything AI first. Now, very few people are actually doing that today. Everybody who talks about AI in their startup is almost always bullshit.

It is— it's become the buzzword. There's always this like buzzword that I think sort of mediocre founders use to— they think it's going to like get them funding. So it was like, you know, Facebook apps and then it was blockchain, that was VR and big data and a whole bunch of other things. And I'm very skeptical when people say, you know, we're AI for X, because it usually— if they say it, if they really believe it, they don't want to tell anyone. And if they say it, it usually means they're not. But I always pay a lot of attention to like what the smartest like college freshmen are going to go spend their time learning. And they're all very good at AI. They're all very good at sort of like applying machine learning to existing problems. And so I do think that that this is— you can see this over the horizon.

So I'm going to flash up a slide. Hopefully it pops up. Somebody in the back there? I have a slide of the top— you can look in front of you— the top 10 employer categories in the United States in 2017.

Yeah.

It's usualists, retail, cashiers, office clerks, food prep.

Yeah, no surprises there.

No surprises there. With relatively high certainty, which of these categories will not be replaceable in the next 10 years? I should have given you a heads up, but—

No problem. So a general statement first that The rate of technological job change is actually higher than most people intuit. It's a little bit spiky, but it averages out over the last few centuries to every 75 years, roughly 50% of the jobs turn over. And we as a society, although we're always anxious about it, always find a way around it. Now, there's like two cases with AI, right? One is this is technology. It will eliminate existing jobs and we'll find new ones. And the other is this is a new life form and it will do to us what the Industrial Revolution did to horses or whatever. And I'm sympathetic to both arguments and I can't say anything with more certainty than I don't know, but I believe that human desire for status and feeling superior to each other seems to be endless. So I assume we will find new things to do, but they will look very different than work today, more different than work on one side or the other of the Industrial Revolution, I think. We can handle job changeover. The question is if it all comes in like 10 or 20 years, can we handle that? That has not been tested yet. Each major technological revolution has compressed in the timeframe that it happens, but it's never compressed inside of one generation. And that feels like something new. If people like, then people actually in one lifetime, in one career, have to change what they do, not just society change. I would guess that some version of teachers are pretty safe because there, there is something about the human connection. And until we get to real AGI, that one feels pretty good. One that I would say feels like pretty bad is cashiers. Yes.

Um, and, and for all of you thinking about if most of the jobs in most of the top 20 employment categories are replaceable, or at least replaceable if not already replaced within the next 10 years, the economic implications for our society is very large. So talk about your personal motivation for why you're working on OpenAI.

Sure.

Given up All the other things?

Look, when I was 18, I made this list of the 5 problems I wanted to sort of help contribute to in my life, and AI was at the top then, and it's been at the top for a while. But until more recently, I didn't believe that I actually was going to be able to meaningfully solve it. And I had this sort of very great job at YC. I always try to— I always want to work on the most highly leveraged thing I can do where I feel like I contribute most to sort of innovation and sort of improving the world. And for a long time, I thought that was YC, which makes sense, right? Because like we get to fund hundreds of companies a year. Many of them go on to do great things and we have a significant impact on the startup movement in general. But when I realized that I actually truly believe AGI is going to happen and then not just present future, it was very difficult for me to be motivated to spend the majority of my time on anything else, because I think this is the platform through which we will solve all the problems that I care about. If you can truly invent AGI, it's the last thing you need to invent. And so—

just elaborate on that, the last thing we need to invent.

Yeah, because then it can do anything. Like, I think we will—

everything else—

we will use AGI, I hope, to solve healthcare, solve climate, educate every kid on Earth.

Um, there's a, there's a fallacy in my view, and I don't know whether you agree, in AI. Uh, people say, well, we'll automate the low-skilled jobs. Uh, I would argue a judge, an AI judge, an AI oncologist, are far easier to do than a warehouse worker.

Yeah. Um, A thing that I always tell people that they're always surprised by is something like 80% of the brain goes to processing sensory input and controlling the body. Only 20% is for thinking as we think of it. The hardware has had much longer an evolution. It's incredible, incredible.

Our fingers can sense a lot of things.

Yeah.

We'll come back to that. So, it's hard for people to imagine that you could have AI judges, AI personal lawyers for every person on the planet, and AI Primary care physician for every person on the planet, 24/7, no appointments needed. The place where people get surprised is we have a startup doing AI artists. And last year, a year ago in the lobby, we bought— we had 5 pieces that were done by artists that were purchased or being sold in galleries for more than $10,000. Yeah. And 5 pieces done by the AI. Nobody could guess which was AI artist and which was a human artist.

You know, one person— people love to say the example that like art is the thing that will stay human. AI can't do that. We recently released something called MuseNet. It took the same technology we used for our language model at OpenAI and made it for music. And so we have this language model that's pretty impressive and getting more impressive every month that can do unsupervised language modeling. And someone said, well, one person said, what if I do that for music? She put it together, trained it on a bunch of music on the internet, and got incredible results. And we made it available for some period of time. And I heard from a number of people that they would rather listen to that than human music. It got really good. And it was endless. And if you love Rachmaninoff, you could hear as many Rachmaninoff concertos as you ever wanted. You never ran out. It was new every time. And I think we're going to learn something about these things that we consider magically human.

It's one of the things people don't realize is if you have 10 different things, styles of music you like, the AI could figure out the features of music you like and actually custom synthesize music, a personal musician for every person?

Yeah.

Because it's what their brain responds to emotionally? Would you agree with me?

Yeah. I think everyone is going to have customized— I mean, we've seen this already with some online services where everyone gets a customized version. But I think that trend is just going to keep going.

In fact, we are surprising a few of you, but let me go a little further in the surprise. Let's talk about will AGI do far more than what humans can do? So the flip question, what is it that humans cannot do that AI will be, or AGI will be able to do? Give me some examples.

A meta comment first about how to think about intelligence. It's very hard to think about how much smarter we really are than other primates. It feels like a lot, right? Like, it feels like we walked out of the trees or the savanna or whatever you want and we're just like unbelievably more intelligent than our sort of nearest relatives. Because we can discover physics and all this stuff. We built all— we sort of dug stuff out of the ground and figured out what to do with it. And at some point, we got computers and phones and buildings and everything. And we've had this— at this point, we have this sort of intelligence outside of biology, this evolu— where we kind of like have created society and this body of knowledge and this set of tools that every generation gets to build on. And it's this incredibly exponential curve. And we feel so smart. I suspect we will learn that the limits of intelligence, although I expect they exist somewhere because of the speed of light and a computer system, if nothing else, are very far. And there will be— like, we feel like much smarter than a chicken or something like that. But probably relative to systems that we will build, sort of the children of humanity that we will someday build, we are probably not very smart at all. And in the same way that that chicken has a hard time about, like, thinking about what we're capable of, it's probably, like, very difficult to explain to the chicken, like, the concept of, like, leaving Earth and going to the moon. I think it's very hard for us to sit here and talk about what the systems we build will be capable of. But it is my genuine belief that long after we sort of create incredible economic value and improve human lives, that these systems will someday become truly awake. And either we destroy humanity before we get there, or this will be the moment where humans' biological evolution successfully bootloads digital intelligence, digital intelligence leaves Earth on Von Neumann probes and sort of colonizes the universe until the heat death.

So—

And I don't— can't quite articulate why I should care about that so much, But it does make me happy to think— much happier to think that the universe will sort of continue to observe itself rather than the kind of light of consciousness going out.

So the critical question in all this is when. And I'm going to get another slide. Can we leave the slides up, please? OK, so this is a chart, years from 2016 and probability of high-level machine intelligence. And each curve is the experts' prediction. They asked 100 supposed experts. You can see there's no agreement.

Look, I think this stuff is like always crap. I can make a reasonable argument. I can make a lot of cases against it. But I think it's a dumb debate. I think it is a very small-minded, short-term debate. If we can accept that there's a 75% chance of getting to this most important moment in human history in the next 100 years, that should be enough for worldwide effort and focus on this. I believe it's much shorter. But whether you think it's 10 years or 20 years, like there's so much energy that goes into debating that. And I think if it's within a few decades, there's nothing more important in the world to work on.

Well, not only that, I'd add, it depends on how much computing power we put at the problem and when certain breakthroughs happen that are not predictable.

Yeah. So one way that we talk about it, like let's say backpropagation is a 10, like the quality of the importance of that idea in deep learning is a 10. And let's say something like the transformer is like a 7. We think— my guess is that we need like one more 10 and about 10 more 7s. And algorithmically, that might be it. We do need much more compute, which is why OpenAI has to raise so much money. But we know how to do it. Yeah, there's no new— I mean, there is new physics. There's no miracles required there.

The flip side of AGI doing far more than humans can is all the stuff today's AI does in a silly way. And I'm going to put up some examples, right? Telling— this is the old picture that's used often in AI. To tell the difference between a Muppet and a Chihuahua.

You know, when I wake up, sometimes when I wake up in the middle of the night, I will like look up at my ceiling in a sort of semi-awake, semi-asleep edge of consciousness state, and I will like see like the lights and the fire sprinklers and stuff on my ceiling, and they look like human faces. So even humans, like something at a very—

I'll give you another example.

Like, we can get tricked if we're not— if we're even a little off our normal state. Like, vision is tough. Optical illusions happen. People that are either really tired or like, you know, on some medicine in some sort of altered state, people make mistakes like this too. And honestly, if I look very quickly at that, I'm not sure I could tell you which is which. I have to take a second. It doesn't come in the first layer of the network or the second. And so it is true. That AI systems, you can trick them, and people love to talk about that. You can trick humans too. And I think a lot of the work that we are doing now, as we make more progress with unsupervised learning, for the first time I think we're actually having systems get to some semblance of conceptual understanding. And it is my hope that in the next few years, we will have a system that never makes this mistake, as with a visual classifier. And that, I think, will be a pretty That'll make true AI feel closer to people.

So back to this question of what today's AI does poorly. I'm more interested in the kinds of the 2 or 3 things, the 2 or 3 technical breakthroughs.

Yeah.

Other than just more computing power. I think that would, that would cause the switch from some level of stupidity in today's AI systems into a more robust world that at least matches human standards?

So one year ago, I would have said the biggest, the most important piece in front of us that was missing was unsupervised learning. And now with our GPT-2 results from earlier this year, I believe we have something pretty important figured out there. We have longer to go, but the fact that we can train these models the same model can generate a story and then be state of the art in almost every text task without being specifically trained for it. It's the first time I felt like the machine is a little bit conscious. You know, these systems that you don't train to do translation or even tell them the concept—

The output of GPT-2, which some of you should look at, it's public, right?

It's public. We haven't shared the latest versions, but there's—

Yeah.

Be happy to show any of you the latest.

Trained on all of the body of text in Reddit?

Not all. No, no. 3 basis points of Reddit. So actually not even that much.

And the answers, when I looked at the answers, they sounded like Fox TV experts, like the same language, the same phrases. Just blew my mind how I couldn't tell experts from—

Yeah. And then the downstream—

The talking heads you see on TV.

The downstream. Performance using that same model to solve all the other language tasks that it wasn't even trained for is surprising. A big thing in front of us now is reasoning. So can— how can we teach a system to have some data and keep thinking, and the more it thinks, the better it does? How can we build a system that can prove unproven mathematical theorems? We're working a lot on that now. We're also interested in how can how can we rerun evolution? So how can we build these very large simulations and have agents with long memories and a lot of autonomy that have to interact with each other and develop sort of social intelligence? It's actually an interesting question why humans— why evolution endowed humans with such big brains. Incredible waste of energy. They make us in our very early months and years sort of like easy prey for other animals. Huge parental investment and this ongoing tax of like 25% of all the food you eat just to run. We don't need those to outrun a lion. We don't need those to run down an antelope. We have them to deal with other humans. And I think this idea that you generate intelligence by interaction with other agents is going to turn out to be quite important. So we're doing a lot of work.

Agents learning from other agents and—

Yeah.

Class of networks called GAN networks. Has the beginnings of that, and BigGAN has a bigger—

Yeah, we have some amazing results there of watching the simulation as agents sort of— because you have this continually escalating curriculum if you have to sort of deal with the other agents in an environment. So that's cool.

China seems to have— so switching topics, and we are running out of time— to be positioning for a global race AI. Comment on that.

I mean, I have just much too much self-confidence, but I think, I think we're gonna do pretty well.

On the flip side, one of the key tenets at OpenAI is safe AI. Talk about safety and the need for regulation and Safety means a lot of things to us.

There's accidental misuse where the system just does something that's not quite what we meant it to, that is awful. There's intentional misuse where a bad guy uses it to sort of conquer the world. There is a policy failure where it's not regulated and we end up doing something that most of the world doesn't want. Most of the world doesn't get input. And so when we say safe AGI, we really just mean beneficial AGI where we sort of maximize human preference and happiness.

We are out of time, but thank you very much.

My pleasure.