Ex-Google CEO Eric Schmidt: 'The AI bubble is a good thing'

0:02In order to win in this new global order, Europe needs a strong tech sector. So the founders that are here, please don't sell for$100 million. Please wait for$10 billion. Hello, and welcome to The Sifted Podcast, the show where we help you get to know the brightest and boldest people in Europe's startup ecosystem. I'm Amy, Sifted's editor, and today's episode of The Sifted Podcast is a fireside chat with the one and only Eric Schmidt, the former CEO of Google. It was recorded live at the Sifted Summit in London in October, where Eric was speaking with Sifted's founder and editorial director, John Thornhill.

0:40I hope you enjoy listening. Good afternoon, everybody, and welcome back to Sifted Sessions. I'm absolutely delighted to have the chance to interview Eric Schmidt this afternoon. Thank you very much for coming along, Eric. I'm really glad to be here. Eric was the CEO of Google from 2001 to 2011 and then executive chairman for Google and Alphabet until 2017. After leaving Google, Eric founded the philanthropic venture Schmidt Futures in 2017. He has also founded the$100 million Schmidt Science Fellows Program to stimulate interdisciplinary scientific research. And he is an active investor and also runs a rocket company.

1:24And as if that is not enough, Eric is also one of the leading thinkers and a presidential advisor on AI. And he is also the author of several books, including The Age of AI and Our Human Future, and Genesis, AI, Hope and the Human Spirit. And one of his co-writers was certain Henry Kissinger. So we've got a lot to talk about, Eric. So let's get started. Many people and skeptical journalists like me would say that we're in the middle of one heck of an AI bubble at the moment. Bubbles are great. May the bubbles continue forever. I liked the one in 2000. Okay, we'll come back to that. But you argue that AI is underhyped.

2:08So tell us why. You know, I wrote two books with Henry Kissinger about this before he died. and we came to the view that the arrival of an alien intelligence that is not quite us and more or less under our control is a very big deal for humanity because humans are used to being at the top of the chain. And I think so far that thesis is proving out that the level of ability of these systems is going to far exceed what humans can do over time. There's lots of evidence of this. It started with image recognition in 2011. In 2015, DeepMind, which is one of the great jewels of the UK, showed that you could beat the Go game.

2:58And most importantly, and we spent a lot of time talking about this at the time, there was a new move called Move 37 that had never been attempted in 2 ,500 years and made no sense. For non-Go players, it's the equivalent of giving up your queen. but nevertheless it worked against the best Go player in the world at the time. The development of deep learning and now the GPT series, which culminated in a chat GPT moment for all of us, where they had 100 million users in two months, which is extraordinary, gives you a sense of the power of this technology. So I think it's underhyped, not overhyped, and I look forward to being proven correct in five or ten years.

3:39I guess it's all a question of time, isn't it? But so do you think, I mean, some of the valuations appear fairly extraordinary. Is OpenAI really worth$500 billion? Is Palantir trading on a forward earnings multiple of more than$200? Does that make sense? It does if you believe that this intelligence is going to give you competitive advantages at scale that's impossible to imagine. So the thought experiment goes like this, and let's use OpenAI. Let's say OpenAI has no competitors, which is not true, and it just marches along, and it eventually gets to intelligence that is general intelligence, and then artificial intelligence, general intelligence, and then eventually something called super intelligence, which is smarter than the sum of the humans.

4:28And furthermore, that one company, you can give it the hardest questions that the world faces, and it can solve them. Now, that's just a thought experiment. What's the value of that company? It's a very, very large number, much larger than any other company in history forever, probably. Now, of course, OpenAI is not alone. They have a strong competitor in Google, right? Thank you, Alphabet. And in fact, Gemini 2.5, the latest one, is now at the top of the leaderboards, I'm quite happy to say, because of the incredible work between DeepMind and Google Brain and those groups. But there's challengers.

5:07You have Claude, I see a booth here, which is essentially another startup, which looks like they're going to be one of the great players. You have Microsoft and Mustafa, the former co-founder of DeepMind, now competing with his former co-founder at Microsoft. These are companies that have enormous amount of capital. They are committing in general. I forgot to mention Meta, which managed to say to the president,$600 billion of capital over five years. The other companies decided to issue statements of a little bit less than that because they didn't want to hype it too much. So if you add it all up, it's like$1.5 trillion in capital.

5:49So these people are not stupid. They know what they're doing. You were mentioning bubbles and how much you love them. I think last week Jeff Bezos was talking about making a distinction between a financial bubble and an industrial bubble. So the financial bubble we had at the time of the global financial crisis didn't do anyone any good because the banking system blew up. That was a bad thing. Industrial bubbles, whether you're building railways or telecoms infrastructures, you can have a bubble that bursts, companies go bust, but you have the underlying infrastructure. You've got the railway network, you've got the telecoms.

6:25Is that going to be similar with AI or not? I don't think so. My favorite bubble story is Global Crossing, which in the late 90s was an incredible company. And they built an enormous amount of dark fiber around the world. They went bankrupt shortly thereafter when the bubble burst. But all of that fiber optic was still in place. and in my early years at Google, I did IRU after IRU after IRU to get the necessary wavelengths to build the data network and Google's data network is probably one third of the global internet right now. So let's go back to that bubble, which I said I love. I just don't like the downside of it.

7:09Another example for British people is the channel tunnel, right? Went bankrupt twice and it works really well. I don't think that's going to happen here, but I'm not a professional investor. What I do know is that the people who are investing hard-earned dollars believe the economic return over a long period of time is enormous. Why else would they take the risk? These are not just random consumers who are reading things in the press. What does it mean for startups? Do they have a chance to compete against these giant companies? Or as an investor, how are you thinking about what startups So let me describe the conventional wisdom.

7:51The conventional wisdom is that there are four or five companies that are now so large you can't catch up with them. Let's go through the list. Microsoft, Google, Meta, Anthropic, maybe Grok. We'll see what happens. Doesn't have a lot of users right now, but they're working on that. And in China, you have investment at a much lower level trying to compete. So for purposes of argument, there are five or six of these things. It's highly unlikely that there'll be a seventh that will emerge at that level. I'm familiar with one startup, because I'm funding them, which claims that they can re-architect the transformer architecture so they can do what is called test time compute using inference machines in a way that's different from traditional transformers.

8:38There are people working really hard to prove the thesis wrong. But those are startups. Maybe they'll succeed. Maybe they'll fail. My current view, Google, of course, invented transformers in 2017. I remember when the paper went out, I didn't even think very much about it. And yet the transformer architecture is the basis of the scaling of knowledge and learning right now. And everything you see is based on that. If you go back to my own history, when I was there, when CMOS was originally invented, that's how old I am. There have been many candidates to replace CMOS architectures for 50 years.

9:13but the installed base, the investment, the specialized training. Look, in a positive example, look at the lock-in that NVIDIA has with the CUDA microarchitecture. For the non-technical people here, why would NVIDIA be worth$4 trillion or whatever the number is? Well, one possibility is that customers who want to be at the leading edge have to use their underlying architecture due to the specialized software, which they will explain to you they've been building since 2007. So again, there are reasons why the value is being created and why the big companies are winning. You were talking about scaling there.

9:49I mean, there's quite a fierce debate at the moment about whether scaling by itself will take us to AGI or superintelligence. What's your view on that? So I have meetings once a week, private meetings in the industry where we talk about these things. And there is absolute agreement that we are not done with scaling. there is not agreement as to when we might be done with scaling right so there is complete agreement that the scaling laws are still running dario who's the founder of open ai and sorry was it open ai now founder of anthropic and the close friend says there's actually three scaling laws playing out one is the initial essentially deep learning training which is what you know about but there's another one involving reinforcement learning and another one involving test time training, which we're just beginning to see.

10:39No one knows where the top is. There is a top. We just haven't found it yet. Now, people claimed that data was the new oil, and then we ran out of data, and we kept getting better, right? So there's something else going on here. And my own view, Jeff Hinton, who invented a lot of this, explains that you understand that these underlying algorithms use next word prediction. So all it's doing when it writes those beautiful paragraphs is predicting the next word. He argues, and he has a right to make the argument as the inventor way back when, is that that's how our brains work too. That there's a parallelism between the reasoning that we think we're so good at and simple next word prediction.

11:25This has an important implication. If you assume that there's a scaling law and you assume that the scaling law is a function of the size of the number of tokens, that is the number of words, then limited language sets, my two favorite being math and software, should have huge scaling advantages because you have a relatively small amount of language you have to predict in that sense and you can concentrated and we're seeing this extraordinary gains in software now my profession is software programming and i've been doing this for 55 years maybe longer and to watch a computer write out the methods and the codes and get it right it's just for me it's like i don't know how to describe the feeling it's it's both horrifying and incredibly impressive so if i were a mathematician how would I feel about the mathematician agent doing the same thing and coming up with a conjecture and then falsifying it or proving it.

12:29It must be the same feeling. The reason those two are going to happen fastest is they're technically called scale-free. In other words, there's no limit because you can just do more programming and you just do more math. Whereas in physics and biology, you actually need real data of one kind or another. A quick one from me. If your company would like to get a message across to Sifted's audience of startup and scale-up leaders, VC frontrunners and tech advisors, why not consider sponsoring the Sifted podcast? You'll help us interview even more movers and shakers of Europe's venture ecosystem and analyse even more of the most impactful trends confronting them.

13:10And you'll reach a hyper-engaged listenership too. For more information, see the podcast description.

13:44My own view is that the robotics revolution is going to happen much slower because it's very much not scale-free. You need real robots, and they're hard to do. I also believe that the Chinese, in terms of volume, will dominate the robotics industry. So that's not a good scenario, given the current geopolitical. There are a whole bunch of startups that are doing world models of one kind or another, which should solve that problem. And remember that, and what I haven't talked about is that the first phase of this was language prediction, but now you have essentially agents and stepwise planning.

14:18The term in the industry is called planning. So an example, I live in California. I want to build a new house. I like houses. I ask an agent to go find a plot. I ask an agent to go review the rules. I ask an agent to go hire an architect. I ask an agent to then hire a contractor to actually build a house, which the computer cannot do. And then in California, I then have an agent sue the contractor for failure to build a house. Right. Welcome to America. Now, why did I give you that stupid example? I just defined workflow. And the agent revolution, technically called the agentic revolution, is occurring right now.

14:58And it's language to language. So what happens is the agent talks in English to the next agent that talks in English to the next agent. agent. One of the jokes we have, and it's not a joke, is what's the point when you should turn off the computer? Well, one point is when the computer discovers it's easier to have the agents talk in their own language and not human language, right? Because then we don't know what they're doing. And how scared are you by that prospect? Well, all of us believe that there's a set of red lines. I'm worried about that one. There's another one which is called recursive self improvement where the basic idea is that the system begins to learn on its own.

15:38So the example is you tell the computer, learn French history, and then it learns French history, and then you say, learn something else, and it chooses to learn emotion, and then it learns about military, then it learns about weapons, and then it learns about desire, and it decides to desire to get in charge of weapons. So the recursive self-improvement could get out of control. DeepMind actually produced last year a very nice paper of a list of these things. Access to weapons is a most obvious one. There are others. That takes us on in a way to geopolitics and the book you wrote with Henry Kissinger.

16:10I remember interviewing you and Henry on video. I asked him about China, and it's quite something when he turns around to you and says, well, I remember when I was talking to Mao Zedong in 1972. But Henry argued that AI is really in a category of its own as a technology, and it's more dangerous than a lot of technologies that have gone before, like nuclear weapons, because it is dual use, it's easily deployable, and in the wrong hands has enormous destructive power. So that combination of three things doesn't exist with nuclear weapons or any other weapons. How worried should we be about the destructive potential of AI?

16:48So it's much more fun to talk about the bad stuff that AI can do. So let's remind everybody, new drugs, solve climate change, new energy, new efficiencies, better health care, hopefully peace, better communications, better mutual understanding, blah, blah, blah, blah, blah. Let's list all the benefits. Is there a possible of a proliferation problem in AI? Absolutely. There's evidence that you can take models, closed or open, and you can hack them to remove their guardrails. So in the course of their training, they learn a lot of things. A bad example would be they learn how to kill someone. All of the major companies make it impossible for those models to answer that question.

17:34Good decision. Everyone does this. They do it well. And they do it for the right reasons. There's evidence that that can be reverse engineered. And there are many other examples of that nature. We don't have a good nonproliferation regime yet. I'm part of a set of dialogues with China on this. This is a hard problem because the reasons that you described. We're alive today because plutonium, uranium-135, enriched plutonium, etc. are really hard to manufacture and if you screw it up it will kill you as the evil person. The same is not true of proliferation. So we have to have a long conversation about how to contain that.

18:12And what is the Chinese attitude to that? Because clearly you've both got exactly the same mutual interest on that level. The Chinese and the United States have similar interests to maintain a non-proliferation regime around bad actors, at least in the case of cyber and biology attacks. It's not clear if there's agreement on everything. Another danger from AI is misinformation. And depending on what kind of government you are or what kind of politician you are, you may or may not be in favor of misinformation. I think that one's out of the bag. You've just come back from Ukraine. How is AI changing the nature of warfare?

18:47Well, Ukraine is an important place in the world, in addition to the suffering of the people who were invaded by Russia, because they're building a whole new national security architecture based on automation. And Ukraine is today still independent because of their rapid adoption of drones and other kinds of techniques like that. And with a three-to-one mismatch of fighting and at least a three-to-one mismatch in resources, they're holding it up. And what they've done is they've created a whole industry, 600 or 700 startups, which are building very powerful autonomous weapons. Where they are right now is they've gotten to the point where they have integrated command systems.

19:32They can see everything. They have the equivalent of Google Maps for what's going on in the war. It's not Google Maps. So Google didn't do it, but it looks like that. And they have all these things flying around and they can target them and hit them. The next stage will be to use reinforcement learning to control swarms of them. And they're working on that. That's why. And so because militaries are very conservative, it's highly likely that the use of AI in war will emerge in Ukraine. And how alive to that is the Pentagon? I mean, they're obviously investing very heavily in drone systems themselves.

20:08Well, the Pentagon is late to the drone stuff. The Pentagon tends to be dominated by traditional procurement, traditional favoritism for specific states, the large primes and things like that. There's a whole bunch of new startups, a number of which I'm involved with, which are trying to produce sort of the technology world that I do to help in national security, which, by the way, would be done in conjunction with the UK, who is a very close partner. I would say that the U.S. is well behind. the U.S. should spend more time watching what's going on in Ukraine because you're seeing a whole new doctrine of national security be evolved, both offense and defense.

20:47When you say well behind, you mean behind Ukraine or behind China? I can't speak for China. When I go to China, I go a lot. I always think that the building I'm in and next to it is the military building because in China they have a concept called civil-military fusion. There's no notion of a military contractor. So when I go to see these beautiful consumer products, I have to assume that there's, but I don't know. I do know that NATO and the U.S., Britain, whatever, are all well behind Ukraine. I can also tell you, having just been there, when the war is over, hopefully very soon, they will be an enormous supplier of powerful weapons to help the West and NATO and so forth, for which we will be grateful, all of us.

21:33I want to talk more about Europe. How does Europe fit into this AI-defined world? As you know, because we spend a lot of time together, I love Europe and I love the UK. And I've sort of given up trying to tell people the truth because it doesn't seem to change anything. So I'll try again, but I'm going to be skeptical and obnoxious as an American and I apologize ahead of it. In order to win in this new global order, Europe needs a strong tech sector. A strong tech sector means eliminating the barriers between the different countries in Europe. It also means a lot more money in the university graduate programs.

22:14It also means much deeper financial markets. So let's imagine you have a young Elon, or if you don't like him, pick someone else, who's in one of the European countries. And they say, who do I call for$50 billion of capital for a product that doesn't work at all? Where do I get the electricity from and where am I going to get the buildings? Just answer that question. The answer is none of those are available. So if I can be completely honest, I think that the battle, the big battle is now, Europe is not going to play in the big battle. Europe may wish to, you're not going to. What you can do, however, is build companies that use that knowledge because they'll want partners and they'll want long term partners to build extraordinarily powerful systems.

23:04And my specific proposal is take every document in Brussels, feed it into a fine tune and train the model and have it tell you what the laws say. Right. Because I can't figure it out. I spent so many years talking to Brussels and so forth. And furthermore, you could say, what is the fastest path to get through all these regulations to solve my problem? So it's sort of irony that the first use of this powerful technology is to figure out how to actually use it to get through your regulatory structure. And the perspective of a European startup, though, does that matter so much? I mean, if you're a Demis Asabis developing amazing technology, you can sell yourself to Google.

23:44You provide the tens of billions of dollars and the compute power to enable that. Well, I don't think all of the countries that I spend time in in Europe do not want to cede dominance to America. And that makes sense to me. You're not going to end up with$250 billion of training hardware in France, even though they have excess energy. By the way, their energy is more expensive than it should be, as is true in all of Europe. So if you define it as there's going to be this thing that's American-led, those are services that you can build on. And there's huge services. Europe is a huge service economy.

24:23Your banking systems are complicated and need to be reformed. Redo your air traffic control system because of the inefficiencies in it. I mean, I can go on and on and on. There are very large businesses to be built that can provide real growth to Europe and real growth to European countries. and that's ahead of you. And I will tell you emphatically, I hired 10 ,000 people in Europe when I was CEO. They are the best in the world. You do not have a people problem. Lots of countries want your people. The problem is you don't have a way for them to organize in a way that is capitalistic at the scale that we're talking about because of internal regulation problems, complexity of governance.

25:07And the other one comment is, I spent a fair amount of time with UK entrepreneurs, and they tend to make$100 million, which seems like a lot, and then they sell. Well, to show you how crazy the people are in Silicon Valley, people say, oh, I started there. So there's just this huge gap of expectations. So the founders that are here, please don't sell for$100 million. Please wait for$10 billion. In an attempt to make the Europeans feel a bit better about ourselves, there was a really interesting chart in the FT the other day showing that the adoption rate of AI chatbots in Europe generally is a lot higher than it is in America.

25:48So 35 % Norway, France, UK and so on, 25 % in America, which rather surprised me. So is there an opportunity for Europe to play in that application level? Remember, I just said you have the agentic revolution. Build agents. I gave you my stupid California real estate example. Everything you do in Europe is regulated in a series of steps. Welcome to Europe. Just automate that. There's a huge businesses and they're completely under your control. They're under European control, European laws. You can become national champions and there will be successful AI startups. Mistral is an example. I was the first investor in Mistral in France.

26:30We really like that. But the fact of the matter is most of the money will be made on applications that use AI to solve the problems of Europe. Creativity, invention. Look at the bio industry in Europe. It's huge. You could use these things in the fashion industries, in the fine lifestyle industries. There are so many ways where they could be applied. Just do it. Wonderful. I'm afraid it's telling us that time is up. But thank you so much for a really – Thank you all. Thank you all. And I'm really happy to be here.

27:02Thank you for listening to this special episode of the Sifted podcast brought to you from the Sifted Summit. We'll be back with our regular newsroom chats and pod studio interviews soon. In the meantime, if you'd like to win a pair of headphones worth£250, please take our listener survey linked in the episode description. We need just a few dozen more responses to help us close the survey. And that puts you in a really, really good chance of actually winning those headphones. so please do take it.