Eric Schmidt

Lenny, it was great to be on your show last time. I'm really glad to be back. It's always great to see you.

Well, thank you for that. So let's start with where we are right now. Folks are very familiar now with ChatGPT and its competitors, which includes Claude and my favorite, of course, Gemini from Google and a number of others. And people are amazed that this stuff can write better than certainly I can. They can do songs. They can even write code. So what happens next?

Well, thank you for that. So let's start with where we are right now. Folks are very familiar now with ChatGPT and its competitors, which includes Claude and my favorite, of course, Gemini from Google and a number of others. And people are amazed that this stuff can write better than certainly I can. They can do songs. They can even write code. So what happens next?

The next big change is in the development of what are called agents. And an agent is something which is in a little loop that learns something. So you build an agent that can do the equivalent of a travel agent. Well, it learns how to do travel agents. The key thing about agents is that you can concatenate them. You give it an English command and it gives you an English result.

The next big change is in the development of what are called agents. And an agent is something which is in a little loop that learns something. So you build an agent that can do the equivalent of a travel agent. Well, it learns how to do travel agents. The key thing about agents is that you can concatenate them. You give it an English command and it gives you an English result.

And so then you can take that result and put it into the next agent. And with that, you can design a building, design a ship, design a bomb, whatever. So agents look like The next big step.

And so then you can take that result and put it into the next agent. And with that, you can design a building, design a ship, design a bomb, whatever. So agents look like The next big step.

Once agents are generally available, which will take a few years, I expect that we're going to see systems that are super powerful, where the architect can say, design me a building I'll describe roughly and just make it beautiful. And the system will be capable of understanding that. That's not AGI. That's just really powerful AI.

Once agents are generally available, which will take a few years, I expect that we're going to see systems that are super powerful, where the architect can say, design me a building I'll describe roughly and just make it beautiful. And the system will be capable of understanding that. That's not AGI. That's just really powerful AI.

AGI, which is the general term is general intelligence, is what we have, the ability to essentially have an idea in the morning and pursue it that you didn't have the day before. The consensus in the industry is that that's well more than five years from now. There's something I call the San Francisco school, which says it will be within five years.

AGI, which is the general term is general intelligence, is what we have, the ability to essentially have an idea in the morning and pursue it that you didn't have the day before. The consensus in the industry is that that's well more than five years from now. There's something I call the San Francisco school, which says it will be within five years.

I think it's more like eight to 10, but nobody really knows. And you can see this with the most recent announcement from OpenAI of something called 0.1, where it can begin to show you the work that it does as it solves math problems.

I think it's more like eight to 10, but nobody really knows. And you can see this with the most recent announcement from OpenAI of something called 0.1, where it can begin to show you the work that it does as it solves math problems.

And the latest models are good enough to pass the graduate level exams in physics and chemistry and computer science and material science and art and political science. At some point, these things in the next, say, five years are going to be super brilliant, but they're still going to be under our control.

And the latest models are good enough to pass the graduate level exams in physics and chemistry and computer science and material science and art and political science. At some point, these things in the next, say, five years are going to be super brilliant, but they're still going to be under our control.

At the key point is what we call technically recursive self-improvement, when it can begin to improve itself. And at that point, I think we're in a different ballgame. And it goes something like this. I say to the computer, learn everything, start now, don't do any serious damage. That's the command. Okay. And the system is programmed to be curious, but also to aggregate power and influence.

At the key point is what we call technically recursive self-improvement, when it can begin to improve itself. And at that point, I think we're in a different ballgame. And it goes something like this. I say to the computer, learn everything, start now, don't do any serious damage. That's the command. Okay. And the system is programmed to be curious, but also to aggregate power and influence.

What would it do? We don't know. So that strikes me as a point where we better have a really good way of watching what this thing is doing. And if you think about it for a while, the only way to watch what it's doing is to have another AI system watching it because people won't be able to follow it fast enough.

What would it do? We don't know. So that strikes me as a point where we better have a really good way of watching what this thing is doing. And if you think about it for a while, the only way to watch what it's doing is to have another AI system watching it because people won't be able to follow it fast enough.

Well, the most important thing right now is don't put them anywhere near human life or mission-critical infrastructure. Don't use it to do hard operations. Don't use it to fly an airplane, those sorts of things. The systems today can best be understood as fantastic advisors. In my case, I had a complicated question. I used one of the LLMs and it sorted it out.