Transcription
So now we have the arrival of a new nonhuman intelligence which is likely to have better reasoning skills than humans can have, generally known as AGI and essentially super intelligence. This is going to happen really, really fast, and I think people are not prepared for it.
So when I talk to governments, what I tell them is: one, this is not ChatGPT. ChatGPT is great, but that was two years ago. Everything's changed again. Second, you're not prepared for it. And two, you better get organized around it. The good and the bad.
>> Okay. So, in the following part, Dr. Schmidt gives a timeline for AGI's arrival.
There's a group of people that I work with. They're all in San Francisco, and they've all basically convinced themselves that in the next two to four years, the average is three years, the entire world will change. And here's roughly how they make their argument. And I call it a consensus because it's true that it's a consensus, but it's not necessarily true that the consensus is true, right? So you have to follow that reasoning.
So the scenario goes like this. First, you have language-to-language LLMs. ChatGPT is the best one. They did a great job. Uh, the other folks get started, and they're catching up in various forms in that department. Meanwhile, you have the development of reasoning, uh, the adoption of memory inside these systems. And so the agentic revolution, which everybody understands here what that means because many of you are in the middle of the agentic revolution, can be understood as language in, memory, and language out.
So, a typical example is, I'll use an example. Uh, I have a house in California. I want to build another one. I have an agent that finds the lot. I have another agent that works on what the rules are. Another agent that works on designing the building. Another selects the contractor. And at least in America, you have an agent that then sues the contractor when the house doesn't work, right? That's sort of how.
Now, why would I give you such a stupid example? I just gave you a workflow example that's true of every business, every government, and every group human activity. So when you understand that the agentic revolution and the reasoning revolution together really change the way we operate as humans, then you understand why the San Francisco consensus is so powerful.
Um, the reasoning revolution goes something like this. Take a look at the, uh, I'll use GPT-4 from ChatGPT. Uh, there are others as well. Watch it go forward and backward, forward and backward, and it's reasoning, and it will blow your mind away. Especially when you ask it a question that you have no idea what the question is about, and you watch it reason.
Uh, in the Google case, Google now has a math model that is at the 90th percentile of math graduate students. If you look at the other ranking systems, it's fair to say that we now have systems now that are 90% of all of the high-order, uh, graduate school skills, you know, math, physics, and so forth and so on. That tells you that when this thing happens, it will happen at a big scale.
So the consensus is that this marches forward, and that there's a moment when what is called recursive self-improvement, the system begins to learn on itself, where it goes forward at a rate that is impossible for us to understand. It becomes combinatorial in a way that we as humans do not understand combinatorics. This is both incredibly exciting and also very worrisome. And by the way, the consensus is three years. My own view is at six, but who knows?
>> Okay, so in the upcoming part, Eric is going to talk about the difference between AGI and ASI and how big ASI is going to be. Watch this.
The short-term answer is there are people who are now looking at how do you control a super intelligence with a dumb intelligence, right? And there's math that I did not understand which says that to some degree you can contain it, but we don't know.
Um, but I think the first point to make is that recursive self-improvement or recursive intelligence is where the system is learning itself, and you're beginning to see agentic solutions that are doing self-learning. Um, I expect by the end of this year, so six months, you'll start using systems that are learning as you're using them. It's the natural progression.
One of the reasons to think that the hardware investments that you're seeing, which are insanely large, make sense is that those reasoning models are infinitely more expensive than a Google search. Many, many thousands of times more electricity queries and so forth and so on. And so, the recursive stuff has already started.
The general consensus, there are different definitions of AGI, is that AGI means it's generally intelligent. Now, generally intelligent means that you get up there in the morning and you have free will. Uh, I think largely invented here in France, and you have free will, and you basically have this notion of, "I get up, I do what I want, I learn what I want, I seek what I want." That kind of strategic intelligence is not yet present in these systems. The consensus is that occurs, my own dates are between, you know, four to six years, relatively soon.
The step after that is when the system, we the general term there is super intelligence, where the system is smarter than the sum of all humans. And the test for super intelligence is pretty simple. It can prove something that we know to be true, but we cannot understand the proof. No human can understand it. Even all of us together cannot understand it, but we know it's true. Henry would say, "What is that? Is that magic?" He also would say, and we said in the book, that when people see that, they may take up arms against it because it'll be so frightening to humans. Now, I think this will occur within a decade. Wo.
So, Eric Schmidt actually thinks that ASI will occur within the next 10 years. That's crazy considering the fact that ASI will be smarter than the sum of all humans. A number of us wrote a paper on super intelligence where we speculated how do you contain super intelligence? I do a lot of national security work, and how do you deal with the fact that your competitor might be ahead of you? Well, in network effect businesses, which is what I've lived in for my whole life, didn't have a name when I started, but understand that all the most valuable companies in the world are essentially network effect businesses because network effects work.
What happens when you've got two countries where one is ahead of the other? And I'll give you a simple example. This is you and me. We have our little tech startup. We have a thousand people, so forth. And then one day, you and I decide that we're going to have AI researchers. That is, computers doing AI research. And how many do we have? Well, we have a million, 'cause we have enough money. And all they need is electricity. You know, they don't eat pizza, they don't sleep, whatever. So now your slope of innovation goes like that. This in a network effect business. This is likely to produce slopes of gains at this level. The opponent may decide that once you get there, they'll never catch up, and then you can do really bad things to them. Then you have an issue, essentially a race condition of preemption.
So these are just the beginnings of these ideas. But if you believe that AI is this powerful, then every aspect of human experience will be touched by it, including national security, politicians, democracies, and so forth. I spend lots of time in conversations with people who say, "Will democracy survive AGI?" Right? These are really the core questions about humanity that are being raised by the arrival of this intelligence.
So if you ask most of the executives in the industry, they will say the following. They'll say that we're in a period of overbuilding. They'll say that there will be overcapacity in two or three years. And when you ask them, they'll say, "But I'll be fine, and the other guys are going to lose all their money." So that's a classic bubble, right?
What would need to be true for that not to be a bubble? If you believe the San Francisco consensus and you believe that reasoning, which is done through inference and through these back-and-forth, um, essentially reinforcement learning chains, if you believe that those are going to be the defining aspects of humanity, then it's underhyped, and we need even more. I know that sounds insane, right? But that's the basis of the reasoning.
I personally don't know. Um, I've looked at this pretty hard because you have these massive data centers, and Nvidia is quite happy to sell them all the chips, you know, and the real estate developers are used to borrowing money to build these things. I've never seen a situation where hardware capacity was not taken up by software. There was an old saying when I was young that was "Grove giveth and Gates taketh away," right? Referring to Intel. So you, the old-timers will remember that the Intel chips would get better, but your computer never got faster. And we used to say, "Why does it not get faster?" Well, because in this case, Microsoft would add all these extra features. So I think it's unlikely, based on my experience, that this is a bubble. It's much more likely that you're seeing a whole new industrial structure.
So one way to think about this is the discovery of LLMs and foundation models, which we're properly going to give credit to the team at OpenAI at the time, is a historically important discovery. It began, right, to give them some credit, although in fairness, quite a few of those people used to work for me. We won't bring that up, but we'll give them credit. That historical event is a very big deal, right? And I say that with great respect for what they, in particular, Sam and Dario and Mira and the team did.
So when you build these systems, you're limited by data, right? And we've sucked up all the data, right? That's what everybody thinks. It's not really true. The way these systems will scale is if they have the, the fastest scaling things will be called scale-free. So what's an example of a scale-free field? Math, right? As you know, mathematicians with the whiteboards or the chalkboards or whatever, they just make stuff up all day, none of which is comprehensible to me. Uh, and they do it over and over again, and they're perfectly happy inventing completely new structures which have no basis in fact because that's the beauty of mathematics, and we know how powerful mathematics is.
So it makes perfect sense that using a protocol called Lean, which is an interchange protocol for defining proofs, you can have one system that generates conjectures and another system that proves it, and that's a scale-free solution, software. And by the way, notice that the number of verbs in math, literally the number of tokens, is much more limited than in language. And that's important for serving and performance architectures, which we can discuss later.
Uh, let's look at software. What do software programmers do? Um, they write code that looks an awful lot like everybody else's. So you have this explosion. You mentioned Cursor. There are competitors as well. Um, Anthropic, of course, on its own is probably the underlying technical leader in that space, with the others competing. The competition is healthy. Um, software can be understood as eventually software people will say what they want, and the computer will write the code. And as a person who spent my undergraduate and graduate and my PhDs on language design and computer OS architectures, to see my entire validation be destroyed in my lifetime is a little concerning. I was talking to one person. I said, "What language do you program in?" And he goes, "Why does it matter? It's a new world."
So software and math, remember, and software is the twin to cyber. So it will also, on the negative side, be possible to generate an enormous number of cyber attacks because you can generate the software, you can generate the attacks against the software because you can, you can do it repetitively until you find the buffer overflows and so on and so on. So the combination of all of that is amazing, right? Those two scale-free will be followed by things which need data. Biology, for example, we don't have enough biology data. Chemistry, we don't have enough chemistry data. Physics, we don't have enough physics data, but it's coming. And that's the basis for this power explosion and infrastructure. That's how we solve climate change. That's how we, you know, we do these extraordinary things in medicine, all because of that.
>> I think you once said, "Silicon is strategy." How does the chip race, um, change global AI power dynamics, and how is China catching up?
>> Well, it's very interesting. Um, I was in a meeting earlier with some folks who are leading open source here. It's called Commit, and they're here at the conference, and they were telling me that on a math basis, there's more open source in Europe than there is in the US. So that was really interesting because I'm historically an open-source person.
The structure that's emerging is very strange in the United States because of the enormous amount of capital that companies are not typically doing open source. They're buying these incredible data centers, offering these incredibly powerful services. And in China, they're doing the exact opposite. With DeepSeek and others, they're doing, uh, essentially open-source, open weights, and presumably the Chinese government is funding it.
Now, what's the consequence of that? It's very interesting. If you have two models, open source and closed source, the open source will get much more adoption, especially in all the countries that we don't go to. There are like a hundred countries that you never visit. Um, they don't have enough economics, they have their own problems, and so forth. So we can end up in a situation where America and the West are ahead, but the majority of the AI use is actually not from US Western models, but rather from open-source Chinese models. I don't think people have quite figured out the geopolitical issues that are part of that.
Maybe the last one before an amazing founder will join us on stage. So you led Google during the mobile shifts. What mistakes were made during the mobile wave that companies should today take into consideration and try to avoid in their new roadmap?
>> I think that I can tell you as a CEO, every mistake I made was fundamentally one of time. When I look back at what we did at Google, we obviously did very well. Android has 90% market share, but we didn't move fast enough. We didn't take it to its logical extreme. Um, and I think other companies, Uber being an obvious example, combined the GPS and so forth to make it happen.
If you look at WhatsApp, um, as a sort of communications mechanism, I did not fully understand that your phone number would become your real unique ID in the world. It's now obvious, but it wasn't obvious to me at the time. But in each case, it was fundamentally an error of time. My message to you is, if you're going to do what I'm, the areas that we're discussing, do it now and move very, very fast. This market has so many players. There's so much money at stake that you will be bypassed if you spend too much time worrying about anything other than building incredible products.
If you enjoyed today's video on Anonymous Official, leave a like and subscribe if you haven't done so already. Thanks for watching and sticking with us.