Thank you very much for having me.
So it might be helpful just to sort of frame where we are with the landscape of AI and innovation more broadly. Would you just kind of set that framework for us?
Sure. Only five years ago, AI was still thought of as something that was in the distant future if it was going to happen at all. And even two and a half years ago, which was right around the time ChatGPT launched, it still hadn't moved past the sort of nerds in Silicon Valley. ChatGPT launched on November 30th of 2022. That was even before GPT-4, and since then the progress has been quite rapid. The adoption, the economic impact has also started to be quite rapid. Just last week we had a model that was able to achieve gold-level performance on the IMO. This is something that I think if you told most people in the field would happen even a few years ago, they would say absolutely not. That's as good as our smartest humans that are true experts in their field. We're now hearing from scientists saying they're two, three times more productive. We're hearing from computer programmers that say they're ten times more productive. It's completely changed what it means to write software. We have systems already that can perform at expert-level intelligence in many, many fields. Now, they cannot operate on tasks that are as long-horizon as humans can, so there's still a big limitation there. But even if progress were frozen right now, which of course it won't be, I think we still have years ahead of us for society and the economy to really digest this technology and figure out what the impact is going to be. You know, there was a saying for a long time that I thought was great and we should try to get to again, which is electricity too cheap to meter.
And we didn't quite deliver on that as a society, although I think we still should. It does in fact look like we're about to deliver on intelligence too cheap to meter. We've been able to drive down the cost of each unit of intelligence by more than a factor of ten each year for the last five years. Looks like we'll do that for the next five years, too. Maybe even more. This weekend I used one of our upcoming models to do a computer programming task that I had wanted to do. Sort of like a home automation nerd—I wanted the lights and music in my house to do this specific thing. I knew that before this technology, it would have taken me days to do that. I was hopeful that with this technology, given our recent progress, I'd be able to do it in hours. I was able to do it in five minutes. The AI did almost all of the work. This is something that, you know, just a year ago you would have paid a very high-end programmer twenty hours, forty hours, something like that to do. And an AI did it for probably less than a dollar's worth of compute tokens. So this is an amazing change, and the speed with which it has happened, the speed with which it will continue for the next few years, I think is still quite unappreciated. We weren't sure even a year ago how much further our current research roadmap was going to continue, if we were going to hit some sort of limit. At this point it looks like we've got many years ahead of us of almost certain progress.
Well, fantastic. That really helps us frame the next conversation, which is you're talking to a room full of people in the financial industry and in banking, many of whom are already thinking about how they're going to use AI or already using AI. But how could you compare the potential for AI and productivity with other tech gains that we've seen in the past? You know, I'm of the age that I was just starting business when the internet became something that we used more broadly. So, how could you kind of frame that for us? And is there an analogy that you might use to describe where we are?
I've never seen a technology revolution quite like this. You know, the historical examples are people talk about the industrial revolution, they talk about the computer revolution. The internet did change a lot of things, but I don't ever remember something that, you know, would have cost $10,000 of knowledge work a year ago costing a dollar or ten cents or whatever it would cost now. Like this is, you know, for my programming example from a minute ago, I think this is just unprecedented. Now, it doesn't work like that for everything. Things in the physical world, robots, whatever, that's going to take a lot longer. Like we will, you know, maybe in 2020 it cost $10 to have an Uber drive you across town. Actually, let me say maybe it cost $100 to have an emergency parcel service delivered across some amount of time and it cost $100,000 to have an application written. In 2030, maybe to have that software application written cost literally ten cents, $100,000 down to ten cents. And having that package delivered has gone from $100 to $1,000. And, you know, it's going to take a while for the robotics, for the full humanoid that can drive the car, take out the package, go upstairs, push the elevator button, the whole thing, that's just going to take a little while. But for tasks that can be done in front of a computer, it's quite—I don't know any precedent for what's happening.
The analogy that I like the most—people say, you know, is this like the industrial revolution or is this like something else? The one I like is this is like the transistor. The transistor was a discovery of really deep science, physics, that turned out to be very hard to discover, easy to understand once you understood it, completely economically transformative. And the value sort of diffused throughout all of society as a massive productivity win. But there was a short period of time where there were a lot of transistors—they're called semiconductor companies—and it was this huge boom and most of those are now kind of like in the background. You're all using devices with lots of transistors in them. They're all over this room. We don't think about this as a transistor device or that as a transistor device, you know, we think of this as like a microphone and a computer and a screen and a camera and whatever else. But this technology was an incredible discovery. It changed what we were able to build. It became part of kind of everything pretty quickly. And you don't talk about companies much as transistor companies except maybe companies like TSMC or ASML or something like that. And in the same way, I don't think you'll be talking about AI companies for very long. You will just expect products and services to use this technology. You won't think about it that much. You will expect products and services you use to be smarter than you and that will be a normal part of the world. Like the transistor, this is a technology that scales beautifully. So we had Moore's law with transistors. We have—I don't think we have a name for it yet—but we have these scaling laws with AI where it just gets better and better and we learn how to really industrialize this and put this everywhere. So I think that's the best historical analog and we did see phenomenal productivity gains because of the transistor. But again, what we're seeing now with AI—to say nothing of the weird cases like the fact that at some point AI can invent its own next versions and really accelerate progress even further—what we're seeing right now, this just goes quite far.
Well, that's a great analogy, and I think when you said industrial revolution, one of the things that we think about here at the board is the importance of understanding how labor dynamics might change as a result of that. So, could you give us your perspective on how you think the labor market and labor productivity—or productivity more generally—might be affected for some of the banks in the room or other industries that might be tuning in by our live stream?
A thing that I like to remind people in our company of is no one knows what happens next. You know, there's a lot of these really smart-sounding predictions. People make all kinds of, 'Oh, this is going to happen on this and the economy over here.' And we have no idea. No one has, in my opinion. No one really—this is too complex of a system. This is too new and impactful of a technology. It's very hard to predict. There are cases where entire classes of jobs will go away. There are entirely new classes of jobs that will come. And largely I think this will look somewhat like most of history in that the tools people have to use in their jobs will let them do more, achieve things in new ways. What it means to be a doctor or a lawyer or a computer programmer clearly will change, but people will still need medical care and they want to talk to a human. People need legal advice and they'll want someone that they can trust to be on their side. People will still want computers to do things for them. But the amount that one person can do—and also the amount that we expect out of one person—will be incredible. Like all throughout history we have these new technologies come along and people say this is the end of work, this is the end of jobs. It turns out people seem to want unlimited stuff and have a huge desire to express their creativity and be useful to people. And I'm still waiting for that promise from the industrial revolution that we only had to work four hours a week and got to play on the beach and hang out with our kids, and here we all are. But one thing I believe is you never fight biology. Evolution has had too long. We are too finely shaped. The sort of biological drives that we have and what it means to be human—technology doesn't change that. You can't beat that. And so I don't think the fundamental things that make us work and kind of animate society are going to go anywhere. And we'll probably all still be complaining that we have to work too hard even though we're unimaginably rich. Probably if we could look from today at the people 100 years in the future, we would look at those jobs and say, 'Those are not real jobs. You are not actually busy. You have unimaginable luxury. You have everything you could possibly need. You have nothing to do. So you're making up a job to play a silly status game and to fill your time and to feel useful to other people.' And those are exactly the things that someone from 100 or 500 years ago would say about us today. And that's just the way of things. I think that's great.
That's a unique vision of the future. But if we dial it back to today, as regulators, we're always very risk-averse and we're also very cautious about the way that we're safeguarding our data and ensuring that we can use innovation as we're doing our work. And I know banks are very interested in ensuring that they can leverage these kinds of technologies, but the data that we are protecting in our organizations is really critical that we're able to maintain its security. So how should we think about risks, and given that we're in government—one section of government—how do you think about government uses?
So we expected that the financial sector and also the government itself were not going to be early adopters of our technology. AI—it's gotten much better, but when we first launched, the thing that the average person thought about AI is that it hallucinated a lot. I remember when we launched GPT-3, there was this whole survey of academics who are supposed to be experts in AI and they were polled on what percent of GPT-3 answers were hallucinations. The real answer was like 0.1% or maybe even less, but the consensus of the academic experts was 50%. They thought half the time people were talking to ChatGPT it was completely hallucinating something. It was obviously not true, but that branding kind of stuck with us for a while. And so we assumed that financial services, to say nothing of the government, were not going to be our early adopters. Some of our biggest early enterprise partners turned out to be financial institutions. Morgan Stanley, Bank of New York—these are major partners of ours that we're doing fantastic work with. And we were kind of like, are y'all sure? And they were like, yeah, we really want to do this. And they have figured out how to use the technology, how to structure it enough that they can rely on it for critical processes. Many other financial institutions too rely on it for critical processes. We are increasingly working with the government to roll out our services to lots of government employees. And a thing someone said that stuck with me is, 'Hey, we realize this is a new technology. We've got to put some new controls on it. But the risk of us not doing this—if we don't adopt this, the risk is that we don't continue to exist as a business. We know we cannot compete—if we're a bank, we know we cannot compete with a new bank that is really going to build an AI-first experience and use AI for the entire stack of things.' So the financial industry turned out to be one of the most innovative in terms of AI adoption, and the rollout has worked better than I thought. There clearly are risks to mitigate. We talked about hallucinations. A new one on the horizon is what's sort of called prompt injections—the idea that once a model really becomes personalized to you and your data, someone else can kind of trick it to tell it something it shouldn't. You know, I may know a lot of private information about you, but I understand when I can share that with him and when I can share that with her and that I should never share it with him. And the models—that's a new thing that's happened—is the models get all of this personal information in them. So there's some work to do there, but we've been able to kind of go through these risks carefully and manage them but really deliver a lot of the upside.
Speaking of personal information and those kinds of things, I might just pull a thread on that. And you may not know much about this, but this is something that the banking industry is really interested in right now—is fraud and impersonating individuals to facilitate fraud. Are there ways that we could mitigate that kind of activity or that we should safeguard if there is the use of AI or identification of that type of impersonation?
Great question. I am very nervous about this. A thing that terrifies me is apparently there are still some financial institutions that will accept a voice print as authentication for you to move a lot of money or do something else. You say a challenge phrase and they just do it. That is a crazy thing to still be doing. AI has fully defeated that. AI has fully defeated most of the ways that people authenticate currently other than passwords, but all of these fancy things—take a selfie and wave or do your voice or whatever. I am very nervous that we have an impending significant fraud crisis because of this. We have tried—I think other people in our industry have tried to sort of warn people like, 'Hey, just because we're not releasing the technology doesn't mean it doesn't exist. Some bad actor is going to release it. This is not a super difficult thing to do. This is coming very, very soon.' And, you know, there's obviously some reports now of these ransom attacks where people have the voice of your kid or your parent and they make this urgent call. That is going to get so compelling. Society has to deal with this problem more generally, but people are going to have to change the way they interact. They're going to have to change the way they verify. Like this person calling me—right now it's a voice call. Soon it's going to be a video FaceTime that will be indistinguishable from reality. And teaching people how to authenticate in a world like that, how to think about the fraud impact—this is a huge deal.
That might be something we could think about. Partnering on the identification of those hallucinations or the intentional impersonations, I think, would be a really beneficial thing to engage in. But as the parent of two teenagers—one just graduated from high school and the other's a sophomore in high school—apparently a lot of these kids these days are using ChatGPT and other AI to facilitate their work and to complete high school. What thoughts do you have about that and how AI can be used in a beneficial way instead of in this way for kids and with teaching?
Let me tell two stories and then I will answer that. I never met my grandfather. He died before I was born. But a story that was told to me by my grandmother was about what happened—he was very good at math—what happened when the calculator came out. And he always kind of liked new technology. When the calculator came out, math teachers at the time apparently said, 'This is a disaster. This is the end of math education. If you don't have to learn how to use a slide rule or look up stuff in the logarithm books or whatever they did at the time, why bother to teach anyone math? These kids are never going to learn.' And it was apparently a real meltdown. And of course what happened is with better tools we use our brain power for something else. We start teaching calculus in high school. We start exposing people to more advanced math. My own version of the story was when I was in junior high and Google came out and the teachers at the high school—I heard this from the older kids—were really panicked because if there was this crazy new thing called Google, you didn't have to memorize facts, you didn't have to memorize the year that a certain war was fought, you could just look it up, and like what was the point of history class? And if you didn't have to fire up your car and drive to the library and learn to use the card catalog and then walk down and try to find this book, but someone had checked it out and then try to find some other book—if I didn't have to suffer that hour, then what was the point of learning? And I had gone to the library and used the card catalog as a little kid a little bit and I was like no, that's a really bad experience. I could do something better with that hour of my life. And again, what happened—so there was like this brief effort at my school to ban Google and get these pledges not to use it and whatever. And then people realized, oh man, we can give our students more tools and we can expect more out of them. And sure, maybe they drive to the library a little bit less but they can use that time to think harder or come up with new ideas or do whatever. And that was really great. We expect more, we get more. The potential is higher, but the expectations are much higher.
And I think it's the same thing now with ChatGPT. It is totally true back to that November 30th date—when ChatGPT first came out, kind of the first mega passionate early adopters were for sure students using it to cheat on their finals or papers or whatever that December. And you had these school districts falling all over themselves to see who could ban ChatGPT the fastest. So, we launched November 30th, by December 7th some people have banned it, by December 14th it's like really careening off the cliff. Schools leave for the break and it looks like ChatGPT is like never going to be allowed inside of any educational institution again. It's a two-week-old product. It's done. By the middle of January you start seeing these things from principals, from superintendents, from serious high-up people in the world of education saying actually we made a huge mistake. This is the best learning tool ever. Our students—at least our self-starting students—are using this to go learn anything. If we ban this from our schools we're just not going to be globally competitive. And we've got to retool our curriculum because this is like the calculator. You know, now we have a calculator for words. So, sure, maybe a take-home essay is no longer the right way to evaluate a student, but that process of learning to write an essay, that process of thinking better—because I still write things that I never show anybody else just to organize my thoughts—that's really important. So, we said, great. Now the good news is students for sure have learned how to use this to learn better, to teach themselves, to think in new ways. We're going to roll out a new way to help students learn better with ChatGPT very soon. The bad news is on the whole, in those two and a half years since, the curriculums have not moved as much as we might have hoped. People seemed to have a lot of energy that they were going to have to start educating and evaluating in new ways and then kind of the molasses of the education system did its thing and we're still giving a lot of take-home essays and I think that's a losing battle. We should be teaching people to use—give them assignments that require them to use tools like ChatGPT, cannot be done without it. That's fine, but expect much more out of them and I still hope we're going to get there.
Sounds like a transformation of our education system.
But it has to be like these, you know, these kids are going to grow up and be adults with unbelievably powerful AI. And if we're not training them for that world, we've really missed the mark.
That's very helpful. Thank you. I'll make sure my daughter understands what our expectations are as she's finishing high school. But to move back to some things that you've talked about in the past and to move a little bit back to the business community, you've said that AI will make small businesses—or it's going to create an environment for entrepreneurs that's going to be second to none. A new era, I think is what you called it. How do you see small businesses? That's something that a number of our board members are really engaged in and we care about making sure that there's new business formation and that there's opportunity there for new entrants. How do you see that being transformed and not just for the Silicon Valley small businesses or new entrepreneurs there but more broadly?
So the first thing we saw with ChatGPT, as I mentioned, was students. One of the coolest things we saw pretty quickly after was everybody who worked at OpenAI started to get these stories of how people were using ChatGPT when it was still very novel in their business life. And everybody has their favorite, but I'll tell you mine. I was in an Uber and the Uber driver was telling me about this amazing new thing called ChatGPT and asking if I'd heard of it and I was like, 'Oh yeah, what are you doing with it?' And he was like, 'It's the craziest thing. I have this small business. It wasn't quite working, but now I have like an employee in every area. It does contracts for me. It answers my customer support emails. It helps me come up with my marketing stuff. It designs my—he goes down this list and he basically had this business running on ChatGPT. This was before it was a common thing. This was before GPT-4. This was an extremely early adopter but he had figured out how to run a business on ChatGPT and it was not like he was taking jobs from other people—his business just would have failed. He couldn't pay for the lawyers. He couldn't pay for the customer support people. He didn't know how to get someone to design the advertisements for him. He didn't know how to have a system that could automatically bid for ads on the internet for him. And ChatGPT just did all of this. Now, obviously, it's much better now. Now, we have a whole industry around us that builds tools with our API to make people really truly one-click automate all those things I just talked about. But the creative spirit of people that were just doing this inside of ChatGPT way back in the stone age really stuck with me and now it feels like it's on easy mode and people are doing that in all sorts of amazing ways.
It's really transformational, I think for me. The first time I used it, I was trying to write a haiku for a company toast at an innovation conference, which actually didn't go very well. But at the urging of my family, of my teenagers, and my husband who's involved in AI too, we tried to make it work. And then went back to the drawing board.
The new version should be much better at haikus.
He's very good at creative writing.
Think about that. But I wanted to make sure we have some time for audience questions if you're up for that. I see a Neil in the front row. If you'll come up at the Chicago booth.
Hi. So I assume one of the things that financial institutions will begin doing is using this to mine their data maybe to do credit scoring, credit evaluation. What do you say to the people that are worried that it's going to be picking up patterns and making decisions on things we wouldn't want to, and how are you guys thinking about protecting against that?
So this is deeply outside my area of expertise, so I'll answer in the general case. One of the things that is cool about these models is they understand us in natural language and they are so steerable. So if you say, 'Hey, you can go look at all this data to make a decision but don't consider X, Y, or Z at all—don't let that have any factor,' it generally speaking will be really good at following that. Doing this in language models is different from the way that we used to do this—sort of just group all this data together into these clusters—so you can really instruct and it really, really follows your intent. The second thing is—I think the question behind the question there is that these models are going to be very biased in ways we don't like. I think humans are quite biased. I think AIs are dispassionate and unemotional and don't have the kind of built-up, you know, whatever that people do. And I think it'll be possible for AI, correctly built, to be a significant debiasing force in many industries. And I think that's not what many people thought, including myself, with the way we used to do AI, but on the current trajectory that seems very possible.
Hi Sam, Scott with a company called Starling. I recently had the opportunity to interview Craig Mundie, who wrote a book with Henry Kissinger and Eric Schmidt, where they describe AI as almost the evolution of a new species. And Craig was saying that we can teach AI human moral codes—they can learn it and reflect it. So there's a lot of talk about the ethics of AI. My question for you is: will AI evolve its own morality?
I think people get very confused about whether they want AI to be a tool or a creature. And I'm very much in the tool camp. I don't think AI will have independent morality. I think that AI can certainly learn—really study humanity and the best of our thinking. It can probably help us point out, 'Here's a problem in your thinking. Here's a thing that should be different. Here is a real moral gap.' And I don't know what the giant moral gaps are in our current worldview, but I guarantee you there are some. And if it can help us find those a little bit faster, if it could point out those inconsistencies, I think that would be helpful. But I think it is easy to sort of impute creature-ness on AI that is not there.
