[0:00:09] So we are so pleased to have this next conversation. I don't think you need an introduction, Sam Altman, but welcome to the Kennedy Center Reach, and welcome to an audience of really motivated, interested people from the financial services sector, from the technology sector, and importantly, from the public policy community. We're so pleased to have you. So I'm going to jump right in, unless...
[0:00:38] Sam, you tell us. We have like an hour with you. I think we have a limited amount of time. So thank you for joining us, first of all. Thank you for having me. So we have had conversations this morning with Senator Rounds, regulators, policymakers, and advisors to incoming administration, the Trump administration. We are talking about the use and implications of using artificial intelligence of many different types of models in the financial system.
[0:01:09] The banking world has been using, as you know well, machine learning models, neural nets, and lots of applications. But you obviously have been building models of the future. And you also just recently released the model 01, which is bringing advantages that are going to be really important for applications in the financial sector where we really need to have some sense of how the models work.
[0:01:37] So I thought it would be helpful if you would indulge us and tell us a little bit about the reasoning model 01, and maybe we can start our conversation together there. Yeah. For a long time, one of the major focuses of the field has been can we get to models that do what is often referred to as system two type thinking in humans, where we can think longer about a hard problem and make some hypotheses, test them out in our head, maybe use external tools, scratch paper, whatever.
[0:02:08] Or just think hard for a long time and come up with a better answer. And that has been something that, despite the, in some cases, enormously impressive capabilities of models like GPT-4, has just mostly been missing. And if you want to do really hard math problems or hard coding problems, you know, be useful to a scientist trying to discover a new understanding of the world, you really want that enhanced reasoning ability.
[0:02:35] And so 01 is our first step towards that. There will, of course, be more coming soon that give you that reasoning ability to think very hard about a more difficult problem and try to come up with a better answer. Fantastic. So as I understand it, 01 is obviously a step in the journey. Is it fair to say that potentially the use of agentive AI is what might come next?
[0:03:06] We think so. I should say we hope so. You really want a model that can do a good job at reasoning before you trust it to go off and do complicated agent-like long horizon tasks for you. But now that we have this unlock, I believe we should be able, or hopefully we'll be able, to make good agentic experiences. That's awesome.
[0:03:32] We are all excited about it. We think about things in the financial services world where this kind of technology has such powerful possibilities, especially around the intimate, personal, dynamic needs around our finances, right? We're trying to figure out which bill to stop paying, which insurance product to go after.
[0:03:55] What are our long-term aspirations? And so the possibility of some type of agent financial management tool has a lot of potential for us. We'd love to hear your perspective on how you see what you have been building and where this is going, how this might come to really be utilized in applications like whether it's financial management services, personal financial services.
[0:04:21] Or other applications. Would love to just hear some of your thinking there. It's a good question. Generally speaking, we love to let the experts tell us what to do. And we view ourselves as we know how to make good AI tools. And other people surprise us on the upside with really quite amazing innovation. But I could share some general patterns that we're seeing. Certainly a big one is the ability to automate repeated and mildly complex world.
[0:04:55] Workflows. You know, if before you could only kind of do good business process automation for things that were very static, worked very much the same. Now there can be like a little bit more adaptability and you see people do quite remarkable things there. Then there's cases where entire parts of a business can now be automated. So an example that I have been amazed by is what's happening with customer support.
[0:05:25] I think as a business, you can now offer better customer support with AI than was possible to do at all, no matter what, before. And the world of calling and getting a phone tree and being on hold and transferred and then the thing not working and the call getting disconnected and having to start over, that's just done. You call an AI customer support agent. There's an instant. Something answers the phone instantly. There's no phone tree. You get exactly what you want. It works really well. It's a great experience.
[0:05:52] So there will be areas like that that can be totally done with today's technology already. And then you mentioned this earlier, as the agentic stuff comes online, you'll be able to build a lot of products and services, which everyone's talking about. The additional thing I'll flag there, because I think it gets talked about less, is what that will do to internal productivity of every company. If you imagine like a software engineer at a bank and what they could do, and if they can now do 10 times.
[0:06:21] We've got ourselves a dream way into the future 100 times more. That will have a very profound effect, and it often gets left out of the conversation. So you know I see a lot of heads of bankers in the audience nodding on that point. I think there's a fair bit of enthusiasm for the value of that. And I will say, as the customer on the other side, it is advantageous to be more efficiently cared for by the bank that I use.
[0:06:50] How do you differentiate sensitive use cases around the future of Agenda of AI versus maybe less sensitive use cases? I think the example that we were going back and forth with your team is some type of financial agent versus I'm trying to discern where am I going to have dinner tonight and what's the tool that's going to get me there. How do you think about that?
[0:07:15] Those kinds of things. As a general observation with new technology, I think it always makes sense to start with the less sensitive use cases. As a general rule, complex financial services transactions should not be the first thing that one uses any new AI leap forward to go build. But I think this is fine. I think this is just the way of technology. And it's good to start with less sensitive use cases first.
[0:07:45] If you look at how much progress we've made since the launch of GPT-3 until now, there are things that you would have had to be crazy to ever use GPT-3 for that you could now really rely on the latest version of GPT-4 to do. And that took a couple of years, but now we're here, and I think that's what it'll look like going forward. So we'll put out new sorts of capabilities, and it'll take some time to figure out how to make them to be.
[0:08:12] Considered generally safe and robust, and then you can use them for more sensitive cases. The one thing I'll add, though, is the reasoning models are much more reliable than the previous models. Their ability to make a mistake and recover, their ability to say, I don't know, their ability to think more deeply before giving an answer. They're not ready for many sensitive use cases yet, but there's a lot of
[0:08:43] Things to like about this approach. Excellent. Thank you. So in financial services, we give a lot of attention to black box and the need for explainability. How do you contemplate, how do we address the black box challenge for large language models and promote explainability when thinking about using these more complex models?
[0:09:09] Two common ways people talk about explainability. One is I need to know what's happening with every neuron in there. I need to be able to x-ray it and see what happens if I change this weight to that. And that's often called mechanistic interpretability in the field. And then there's another kind, which is you need to explain to me how you got to this answer and tell me what happened.
[0:09:38] At each logical step of the way. And I'll decide if I agree that those were reasonable or not. The former is not something we get to do with humans either. I don't get to look in your brain at the way every neuron is firing and say, OK, well, it's no longer a black box. I understand. But what I do do is exactly that same thing I said a second. I can ask you, OK, this is your answer. Explain to me how you got here. Tell me what data sources you used. Let me go look at those myself if I want. Tell me each.
[0:10:09] Logical step along the way. And I can decide if I think those are reasonable. And, you know, break it down to a level where I can see each jump, but you don't have to tell me like every really silly piece along the way, just the major ones. And then I'll decide if that seems explainable enough. And that's kind of what our reasoning models can help us do. You know, we can ask them to explain how they got to something and decide if we agree with the conclusion. My sense is that will be much more the kind of...
[0:10:38] Interpretability that we find useful. That's really helpful. And these are some of the questions we've been grappling with in the financial sector, going from a logit regression model to even a supervised machine learning model. How much and what do we need to explain and for what purpose? Thank you. I appreciate that. Another component of today's conversation is about how are we engaging the technology sector in the world of financial services? And some of this comes down to culture.
[0:11:06] And different cultures and how we're bringing a very regulated sector to utilize and benefit from the technology that the tech sector is building. It would be great to hear your thoughts around these different cultural differences, but also what do we do together to really understand and engage so that we're in a level of comfort with bringing in the technology to the world of banking, if you will. The number one point.
[0:11:35] I would like to get across is just how quickly the technology is still evolving and how fluid things are. And I think that makes heavily regulated industries and also what makes a lot of things challenging. The things that we believe are possible or not possible today can completely shift in a relatively short period of time. And operating that way is...
[0:12:06] It is difficult to make very long-term plans when the technology is this dynamic. And the best thing I can advise is to just not hold anything too rigidly right now, keep a very close eye on technical progress, and be willing to adapt. Yeah, I appreciate that. That's hard for us, but we want to listen and learn, as you all are figuring that out too.
[0:12:32] So you have yourself invested in a few different fintechs. And it would be interesting for us to hear, you know, what are those fintechs doing? Why did you decide to invest there? And if it gives us some sense, any flavor of what you're thinking about the future of financial and technology together, you know, where the world may be going in light of these models. Would love to hear you on that. I love fintech. As you mentioned, it's been some of my most fun investments.
[0:13:03] I don't know exactly where the world is going to go here. As I mentioned, that point about the rate of dynamism makes it hard. But I think abundant intelligence and really like a level of intelligence that can do very complex tasks has got to mean we can make great new products for this gigantic sector of the world. One of the reasons I got interested in fintech in the first place is
[0:13:35] I met a stranger while traveling who worked on finance in South America. And he said, corruption is the number one problem in the entire world. And you're a tech guy. If you could make tech that could solve financial corruption, that would fix every problem in my country. And for whatever reason, that just stuck in my head. And the more I thought about it, the more it seemed like, wow, this really is a huge thing. And that led me down the path of getting very interested in it.
[0:14:01] And then of course realizing how much of our lives it touches. But I do still very deeply believe the premise, which is this just makes the world go round. And if we can make much better products and services, it touches everything and that'll be a huge impact. Yeah, we love that. And some of those challenges are the...
[0:14:27] Some of the barriers and problems, and you're putting your finger on it, I'll say it directly, to what we're really focused on, which is where is the way in which technology and data can actually enable more people to be a part of responsible products, financial products and services, right? How do we reduce the friction, reducing the cost so that there is the opportunity for more people to be a part of a financial system? So I do appreciate that.
[0:14:52] People would be sad if I didn't ask you. We hear a lot about artificial general intelligence. What is it? When is it going to happen? And what will be the implications? This feels a little harder than what's the future of banking in light of AI, I have to say. But you spend more time on this. It used to be a term that people used a lot. And it was this really smart AI that was very far off in the future.
[0:15:20] As we get closer to it, I think it's become a less useful term. People use it to mean very different things. Some people use it to mean something that's not that different than 01. And some people use it to mean true superintelligence, something smarter than all of humanity put together. We try now to use these different levels. We have a five-level framework. We're on level two with agents now. Sorry, with reasoning now.
[0:15:47] Rather than the binary of is it AGI or is it not, I think that became too coarse as we get closer. But I will say, by the end of next year, end of 25, I expect we will have systems that can do truly astonishing cognitive tasks, like where you'll use it and be like, hmm, that thing is smarter than me at a lot of hard problems.
[0:16:19] So we are sitting in DC thinking about public policy. What laws, if you were given the pen and a blank sheet, would you want to see the US do to ensure we build the world's first safe, super intelligent AI? Because there's so much unknown, I think really good testing frameworks are probably the best start. So this has worked for many other industries. It seems like a very reasonable thing to do here. Elaborate on the testing frameworks for us.
[0:16:57] So we're all keeping up with you. Oh, like, you know, before OpenAI gets to release O2 or O3, there should be some sort of federal testing framework that says, here's the dangers that we're most, the harms we're most interested in monitoring, mitigating. Here's a set of tests. And before you can release it, you know, you've got to be able to certify, like you would for a new drug or a new airplane or whatever, that this model is safe in these ways. So that's a good lead to my...
[0:17:25] Potentially last question, which is what would you advise for the Trump administration? It sounds like you're referencing some of the collaborative work that you've been doing with the AI Safety Institute. There's a bunch of things I would advise on safety, but there's one other thing that I'll use this minute to plug, which is unlike other kinds of software, AI needs huge amounts of infrastructure.
[0:17:54] Power, chips, massive data centers, the whole thing. And the U.S. has at other times in our history, under a visionary leader, decided to really get behind a massive infrastructure buildup for national prosperity, national security, all the above. And I would love to see President Trump really go after that with vigor.
[0:18:21] I'm going to keep going because we can. Part of our conversation today is about what are the implications of AI for workers. We heard from Oren Cass earlier today talk about his perspective. Can you give us your thoughts on the impact of AI on the workforce over the long term? There's always this question people get confused on, I get confused on.
[0:18:50] Is this going to only be a tool that makes workers way more productive? Or at some point, does this replace a lot of labor? The answer is clearly both. It makes workers more productive by replacing some amount of tasks and giving people more leverage. And my kind of optimistic default always is that we've never run out of things to do. For a long time, we've been hearing how automation is going to make it so no one has to work and people still have to work. On the other hand, maybe it's different this time and I wouldn't.
[0:19:20] Say I or truly anybody else really knows the answer to that. And then you have talked about universal basic income as a potential policy tool for the possible implications of loss of work, loss of work opportunities. Could you elaborate on that with us a little bit? Yeah, I don't think it's a solution. I think it can be part of a solution. As the society has gotten richer, the expectation of...
[0:19:51] Support has gone up and up, and I think that will continue to happen. But I don't think it's like a binary thing. I also don't think it replaces everything else. Other supports, other training as a for instance. Yep. Great. Is there anything you would like to share with us in the remaining 30 seconds? No, I'll give you the time back, but thank you for having me.
[0:20:21] I might be the only one in the world who takes 20 seconds back. Thank you so much for being with us. I appreciate it, Sam.