From Business to Warfare: How AI Affects the Modern World | Azeem Azhar

I'm very happy to be here in person.

Yeah, Matt, thank you. We're familiar with the idea of exponential, especially post-COVID and those of us who've worked in tech. It is the thing that, in a way, describes what happens to silicon chips through Moore's Law, which is that they get more and more components on them every couple of years. And it's very, very nonlinear. So that idea that you have technologies that could behave exponentially is one that we've been familiar with for 50 or 60 years. But what I'd noticed in my research from more than a decade ago was that we saw those patterns in other areas.

You saw it in terms of the amount of bandwidth that you could put over an optical fiber. You saw it in terms of storage densities. And we saw it in areas like genome sequencing and in areas like lithium-ion battery energy density, and ultimately cost. And so I noticed that there was this process of a whole series of general technologies that were behaving a bit like silicon chips, which is they were getting dramatically cheaper every single year on a compounding basis. And we know that story.

We know that the iPhone is more powerful than all the computers in the world at the time of the Apollo moon landing, whatever metaphor we want to use for it. And then I wanted to figure out, is this really happening, or are we just reading it in newspaper headlines? And if it's happening, why is it happening and what does it mean? And so, as I dug into it, I did identify that it really was happening. And the main reason it happens is because of learning.

We effectively learn how to improve these technologies as we produce more of them. So this is different to scale. We're getting more efficient, we're finding new methods. And this idea of the learning curve, which was first identified in the 1930s by Theodore Wright, was the thing that was in common with all of these technologies. And then what does it mean? Well, then you go to basic economics, which is that as the price of the technology comes down—and ultimately, technology is things getting cheaper.

As the price comes down, businesses find the breakeven point for using that new technology falls as well. So you get much, much wider use in industry. And as something gets used more heavily in industry, it gives rise to complementary opportunities for new entrepreneurs. And you start to build new industrial structures and ecosystems, which then has a knock-on in terms of our behavior as consumers or as citizens or as family members within society as a whole. So that's the full arc of the thesis.

But at the heart of it is this notion that technologies will behave really differently when they have these learning curve effects, and that, for particular reasons, since about 2010, we have crossed a certain threshold with, effectively, the rate at which these technologies can impact us on a practical daily basis.

That's right.

What has changed, if anything, to the exponential age thesis in the last couple of years? Well, I was right, so that's the big thing. I think what we have seen has been that Moore's Law has not died, and we are delivering more FLOPs for fewer dollars—floating-point operations—for fewer watts of power than we were doing when I wrote the book. The price of lithium-ion batteries, which is another technology, has dramatically fallen even though the raw material prices spiked a couple of years ago and demand has gone through the roof.

The price of solar has really come down. The rate at which we've installed solar photovoltaic has absolutely gone through the roof. A mathematician will probably say, hey, that's not exponential, that's hyperbolic or something. But there's been lots and lots more evidence that shows that these patterns are commonplace in the broad industrial transition that we're going through. And the other thing, of course, that we need to talk about is ChatGPT. So I did talk about large language models in my book.

I talked about GPT-3. I talked about scaling a little bit from 2 to 3 and the improvement that you got through scaling. But I think everyone was really surprised at what happens when you get something slightly better than GPT-3 on that curve and you put a user interface in front of it. And so that has triggered an absolutely new set of behaviors, some of which you can use the exponential analysis to analyze, but others of which I think you have to use other frameworks to make sense of.

But you can't predict every factoid, but the overarching arc, I'm more confident now than I was four years ago when I put the book together.

Yes.

Eight months. Yes, it was something like that.

Well, I think that there were a couple of key questions around the relative power of the big technology companies. And that power was part of the point that I was making, was that you couldn't use the traditional lenses of monopoly to look at companies like Amazon or Apple and so on. You'd have to use these new lenses because they behaved rather differently. Their products are too cheap to use, and also their impact on our public space and political space. So what's happened since then is that there has been a new set of technology competitors, in the sense of OpenAI sort of showing up and changing the game a little bit.

But also, especially through the EU, the willingness for governments to start to be much more proactive and ex ante about the way in which they look at these companies. And of course, there is Lina Khan, who I did actually write about in the book, who is now the head of the FTC, and she's taking a different stance. So that's one dynamic. I think the other thing that has perhaps been—perhaps we've got a bit more evidence—is what really was the impact of social media on the fragmentation of the public space.

Four or five years since I finished working on the book—it takes time for them to be published—there's more evidence from the digital humanities and the digital social scientists to say that we don't have adequate causal evidence that social media has actually corroded and soiled our public sphere to quite the degree that we thought it did four or five years ago. So I think that ultimately, obviously, Matt, is the beauty of science rather than blind belief, which is that you can continually test and you can refine your thesis.

But again, I think those are aspects of the detail relative to the overarching view of the curve.

I think the evidence is that finding the causal chain is quite difficult and often conflicting, that it's really dependent on other attributes like the institutional depth in a particular country. So, is there good human capital investment through education? Are there good courts? Has there been a history of good-quality media? And those are often key factors as to whether social media has an overall deleterious effect through which you can find a chain, or whether life just seems really fractious and broken.

It's easy to blame technology rather than to blame other factors.

Yeah, absolutely. So Claude is the chatbot and large language model built by Anthropic. Anthropic was created by some escapees from OpenAI, and they came from the safety side of OpenAI. And Claude has kind of got a friendly personality. The way I used to characterize using Claude versus ChatGPT was that ChatGPT was a little bit like the really, really nerdy honors physics student at college who knew everything, but was a little bit pedantic and unpleasant. And Claude was like the really scholarly history humanities student who knew an awful lot, was just really charming.

And so depending on what you were after, you'd go to a different character. And so GPT-4 was definitely, in a sense, better than Claude. But what I would find in the 10 to 20 hours a week that I work with these systems is that I would prefer to use Claude. I would get less tired. I would want to interact with it more. And it came back from—it was based on the personality that the software has. And personality is a kind of interesting phrase to use.

I think about it as the user-friendliness. I think about the way in which, when the Macintosh first came out in 1984, it popularized what we call WIMP: windows, icons, menus, and pointers, over command line. And of course, command line, in a sense, was more powerful.

I think it's about user-friendliness and about helping us achieve what we need to achieve. So I have this discussion with people quite a lot because I anthropomorphize everything, and always have since I was a little child. I'm just one of these people who puts personality into a teapot, right? Or into a car. But I also know that it's only a car and it's a teapot and it's not a person, even if I've given it a name.

And so I think that anthropomorphizing these products has to be done with a degree of care. But when I put my user interface hat on, having something that has a little bit of a personality absolutely helps you get your work done. And so when we think back to these computers in sci-fi, the ones that respond in very curt ways push a lot of cognitive load back onto the person. And if it just says, "Affirmative, negative, affirmative, negative," you get 200 light-years.

All of the work is being done by the person. Whereas once you start to make the interface a little bit more intuitive, the way that we start to think, then the software is taking on some of that load. So I think it is about building better software. Now, if all you're doing is using the large language model to extract entities, you use it for named entity recognition, for example, for sentiment scoring, you probably just need a curt, brief response that says, "Negative, but a bit sarcastic."

And so in that case, maybe the personality training doesn't matter. But for these more general applications, where you don't know the moment you sit at the sort of query box where you're going to take it, I think it really does help.

Can I answer an easier question first, which is where are we?

Where are we? Well, let's talk about where we are in the adoption.

So, we are moving so quickly. The awareness of these tools is just off the charts. And even the use levels in the UK, where I'm based, normally on a survey basis, 9% of people were saying they were using a chatbot weekly or more regularly. And normally when you look at an adoption curve, once you hit about 6%, you're in the sort of exponential part of the S. The other question to ask is how quickly do products actually diffuse, and how quickly are there technology transitions where you replace one behavior type with another one?

And we often think that they take a really long time, but generally 15 years is a lot of time. So when we moved from sailing ships to steamships in the Europe-to-New York passenger markets at the turn of the 20th century, that entire process took about 15 years. And there was a terrible public relations disaster in the form of the Titanic during that time. And the market expanded and prices came down. And when you look at the iPhone and smartphones, you are looking at a sort of seven- to eight-year period from the arrival of the first smartphone to the point at which 75% of all phones being sold are smartphones.

And the point at which, by 2014, you and I will remember this, we were in the mobile economy. Every startup was a mobile startup. Every government was moving tax onto mobile websites and so on. So with this particular product, where it works on our phone, the interface is really quite easy. I think you have to be making quite a bold and heterodox claim to say that we won't all, by which I mean 75% of us, be using these things daily in less than seven years from when ChatGPT launched, which was November '22.

And that means that it's the next four or five years where it would happen. And that would be, I think, quite conservative because that would say that's an iPhone trajectory. And I think this will be faster than the iPhone trajectory. So we're early. You're not too late. We always have to say that. You're not too late. Still get involved. But I think it will happen very, very quickly.

And then we can talk about the other harder question, which is the AGI one.

What do you mean by AGI?

Yes. No, well, I'm asking you. Yeah.

Yeah.

Yeah.

Mm-hmm.

Mm-hmm.

That was a great definition. And I look at this question in a couple of ways. So one way is that there is a path where we can look at these technologies like all our previous software, and they get more and more capable, and we don't really have a milestone that is AGI, that is before or after. So the first spreadsheet I used was VisiCalc. The one I used most often when I was a teenager, didn't have many friends, was Lotus 1-2-3.

You'll understand why, because of Lotus 1-2-3. But what is Excel today? More of a spreadsheet than 1-2-3? Is 1-2-3 not a spreadsheet because we now know what a spreadsheet looks like as Excel? So there's partly, I think, about this in terms of maturation of software and capabilities that it ends up having. So that is one lens that I use when I look at this. But the other lens, because lots of smart people, Demis Hassabis, Dario Amodei, and Ilya Sutskever, and so on, say we can build AGI and we can do it in a reasonable period.

I mean, there are other smart people like Yann LeCun who say it's going to be another 15, 20 years. So we also have to take into consideration what Demis and Ilya and so on are saying about what they think they can deliver. And as someone who uses these tools really regularly, I would say that, are we past the VisiCalc level? Because VisiCalc was obviously a spreadsheet. It feels that we are. Do I put things into ChatGPT and Claude where it blows me away?

In terms of what it can do? Yes. Do I end up in little dead ends, or sort of embarrassing results, and I just think this is useless? Yes, that happens as well. But that also happens in my career with teams I've had of people working for me, or in work I've submitted when I've been younger, absolute bullcrap, if I can use that term. So, I mean, I think that we are already at a point where you can use these tools in quite general ways and get very, very good results.

And I do that really, really reliably. What I struggle to imagine is what it means when you have the exponential increment from GPT-4 quality to GPT-5 quality. And I know the OpenAI team uses some analogies that are not perfect to try to express to people what they think that's going to look like, because we don't really enumerate any of this. So I think we have to stick with it. The thing that I do question slightly is this idea of ASI, right?

Artificial superintelligence and how quickly that can come and what that's predicated on. So the current model that we use is this scaling system where you just need to throw more and more chips and more and more power at it. And so that model for ASI just requires a lot of power that might be hard to bring on board. But the second thing is that the claims of ASI, artificial superintelligence, are so significant that the path to getting there without there being some kind of magic that bends the laws of physics, and I'm never keen on things that have to break the laws of physics, seems like it's quite a long one.

So, for example, ChatGPT can't really do very much with protein structures. For that, you have to go to AlphaFold or some of these new foundation models that have just been coming out this week and so on. So what would it take to integrate that type of modality into a large language model? And how does that fit a roadmap? And I'm thinking really prosaically. And the magical thinking that I think gets applied is, well, these things will become recursively self-improving, as life has been, by the way, for a couple of billion years through a different mechanism.

But they have to be recursively self-improving. And once you have that, this takeoff can take place. But I think that that, to me, is like a deus ex machina kind of call that you make offstage. So when I look at this question of ASI, I think things can feel, they can feel really intelligent. They can feel that we have these intelligent systems helping us around without me being able to go to my smart light bulb and say, calculate the likely interactions between these two water molecules coming out of my faucet, right?

Which is what an ASI would be able to do.

I tell all of them to go and use ChatGPT or Claude for a couple of weekends solidly on their own, and to not stop until they've done like 10 or 12 hours. And I say, this is not a technology that you can learn through PowerPoints that your juniors have presented to you and you tick off for an ERP or the telephone system or an ATM connection. So you have to go off and use it yourself to make sense of it. And one of the reasons for that is that if you view the current set of AI tools, large language models, as NLP, natural language processing, on steroids, right?

Higher quality, faster. All you'll be doing is looking at point tasks that you can improve. So, can we look at the flow of emails that come in and route them more efficiently than we previously have been? Can we automate this particular task, which is an edge case that we haven't spent the time putting in the nested if-then-else statements in Python, but maybe an LLM can kind of deal with it for us? No one's going to build a great business that way around.

What you need to do is have the experience that I had, which is, wait, I can now think so much better because I've got a cognitive tool, and that helps me. And so one of the reasons I push the top execs to spend time doing that is because my hypothesis is that a lot of the reasons why firms behave the way they behave is because the process of actually exploring their possibility frontier, rather than exploiting their existing operational systems, is too expensive.

And we've been able to bring the price down now through these tools. And what happens when you start to be able to reinvent yourself? And I suspect companies start to look a lot more like, or they'll behave more like Amazon does, or more like Facebook when it was Facebook rather than Meta, which is that there's just lots of experimentation on the edges. Whereas traditional large companies are sort of still driven by waterfall charts. So that's the push that I try to give bosses.

What I would say, having spoken to companies for many, many years, is I can't remember a time where the boards and the C-level were as switched on and had as many projects in the pipeline for new technology as what I'm seeing now. Yeah.

It may not restore imagination. What it may do is it may help you realize you need more imagination, and you have to come in and be imaginative with the questions that you ask. And in a way, I mean, I know that corporate strategy teams have a bit of a sort of bad reputation with frontline workers in general, but it is about being able to apply some of the thinking that is often held in corporate strategy teams at the frontline, by the frontline. And how do you do that?

And then how do you control an organization where you've allowed that to happen? And I think a little bit about what have we actually just done in our organizational design? Because we make these assumptions. And so if I can give you an analogy to this, which is in pharmaceuticals, there's a pipeline and there's an assumption about how many candidates are going to be at each gate. And so everything is designed downstream for an assumption of N, where N is probably less than 20 a year.

Now with AI things, N could be a million. And so that gate is not going to work because that gate that only let through 5%, which was one in 20, is now going to let in 5% of a million, which you'll, VC, know what that is. It's 50,000, is it?

Okay.

Okay. Yeah, fair enough. Okay. Well, it's more than 20. And then that process breaks. So I think the same will be the case when you start to put in people using LLMs in the right kind of ways, as cognitive assistants, as part of a greater collective intelligence across a company's entire corpus of codified knowledge. And also, you bring people to bear with their uncodified knowledge. And what you'll do is you'll create so many more opportunities, which effectively are ideas, and the process by which a company decides how they score those and what they go ahead with.

Will be broken. So those things, those aspects downstream, will need to change.

In the US and in the UK, people are pretty down on it. So Edelman, which is a PR firm, has a thing called the Trust Barometer, and they've been running it for a couple of decades. They interview thousands of people in 28 countries. In Asian countries, people are optimistic about technology and they're optimistic about AI. In the UK and the US, the numbers are really inverted. And so where 70% of people in China and India are optimistic, only 30% are optimistic about these topics in the UK and the US.

They actually explicitly want to have less AI. So what that tells me is that the policy world and politicians and the tech industry have done a really, really terrible job in talking about what the potential for this technology really is. I mean, it's for two reasons. One is that what it is is really contested anyway. And we see these battles between the top AI scientists. So how can an ordinary person who's not put 30 years of their life into this question feel they've got a good answer to that question?

It's like asking people in 1983, are there too many spreadsheets in the world? Well, probably. And I think the second thing is that, and this is an issue, is that if you do think that AI is a valuable or useful technology that most people don't think it is, it's only going to end up being an issue. Now, I suspect that people won't really have a choice because their interface with the world is effectively largely mediated by private companies. And so if you have an iPhone, you'll have AI in it.

Salesforce.com, for your CRM, it'll have AI functionality in it. Your word processor will as well. And opting out of that will be quite hard, but it also won't make sense to opt out of it any more than people opted out of the QWERTY keyboard, which I know a few people did with the Dvorak keyboard. But in general, you don't because it's the market standard.

No, I think it's that Asia has been getting richer rapidly, and connected to getting richer is more technology, more IT. And whether it's causal or correlative, it doesn't really matter. You just notice that technology came as you got richer, and technology created new opportunities to build businesses. And so you're able to be entrepreneurial, and there's a higher degree of optimism. Whereas I think, speaking of Europe, Europe hasn't had a growth story, certainly in Western Europe. And I did specifically mention Western Europe because I can't remember the data for Eastern Europe, so the former Warsaw Pact countries.

Western Europe, there hasn't been that growth story. And I think culturally, we've got tied up on non-growth arguments that are quite far away from the reality of science and technology. So one of the things I think politicians need to take on, and it'll be easier in the UK, where we have an election on July the 4th, than in the US, is a way of putting science and science discussion at the center of whatever public debate there happens to be and whatever public interactions there end up being.

And you have to do that against this very fragmented media landscape. So I think there are two ways of looking at this. One is to say it doesn't matter that 70% of people are negative because the technology is going to move faster than governments will be able to stop it. They couldn't stop it anyway. And everyone will be buying it through their apps and they'll be loving it. That's one approach. I think the other approach is to say it is an issue, and it's an issue because politics will still matter and it'll come back and bite you at some point unless you deal with it.

And that's the view that I would take, which is that it is an issue, that it's, in a sense, misunderstood because it will get reflected in poor politics.

No, I think we have to have regulatory frameworks in place for industries. In general, you can have bad regulations that don't make sense. But the pharmaceutical industry is much bigger for being well regulated than for being completely unregulated. And I think we are starting to see, in the airline and aviation industry, what happens when you change regulatory quality. So what should happen with AI, I think, is really, really complex because you have three different things going on.

You've got: in what way does AI behave like things we already regulate, and therefore we should be thinking that these applications need the same treatment? The second is, are there particular applications that emerge or ways in which AI applications work that create harms or create risks that are not captured by things that we have already recognized as being problematic? And the converse could also be that we thought of something as being problematic, but we don't now because the technology has changed. And then the final bit is, is there something that is unique about AI that means that we need to have some sort of state framework.

And the obvious example is about the safety of frontier models, which people are very concerned with, which is they're going to become so powerful that they should be, I think in the case of a couple of people who said this recently, protected by America's nuclear deterrents. And so if that's the case, then one clearly needs to have some kind of regulatory arrangement.

I'm originally from France.

That's super important, yes.

Well, I think that the U.S. has got a lead, and it's quite hard to read how far behind the Chinese are or aren't. And so, in recent things I've read, some people believe that the gap has narrowed. I would find it surprising if the gap has narrowed for a couple of reasons. The first is that in moments of science like this that are driven by talent, and then you have to build instruments and tools to support the frontier, those ideas and capabilities tend to agglomerate.

And so there's a lot of advantages that the U.S. has from all of the transformer authors working at Google, the venture capital community, OpenAI and Anthropic and other companies, and then the supply chain, or the Scale AIs and Improbables and so on, that do the training and that are all quite mature. They are an ecosystem that has evolved within the dynamics of the U.S. system. And so how do you race against that, especially as the tools themselves help you build better tools?

So we're using GPT-4 to look at GPT-3 to figure out better explainability, which we can then hopefully apply to bigger models later on. So I think that the underlying network dynamics that exist within that ecosystem would naturally favor the most mature and developed ecosystem, provided it maintains dynamism. And so I find it quite surprising, this idea that someone could catch up, unless someone has made the wrong scientific decision and it's a dead end. So heavy water, the German nuclear program, where the Germans were ahead and then they decided to go down the heavy water route, and that was a dead end for this.

So maybe there's an analogy there, and maybe there's something that's been discovered that is better, more efficient, gets you up the curve quicker that we don't know. So I think the U.S. will, I could be persuaded otherwise, but will maintain a lead, and the lead ought to accelerate. But, and here's the but, the but is when we move from models that are text- and video-based into models that have, say, physical interactions. So from learning from humanoid robots or learning from driving in the streets, where you might say there's going to be much more of that data coming out of China for regulatory market structure reasons, which would mean that the U.S. would run out of the key input that they need to maintain a lead, and that would allow the Chinese to catch up and then ultimately overtake.

I haven't said much about Europe at this point, and I don't have a great framework for looking at Europe other than to say: small, fragmented market, far less talent, far less capital, but enough talent. I mean, lots of it. And then one gets surprised because, A, I now consider DeepMind part of the U.S. rather than part of Europe. And, of course, the other thing is that there's been this breakthrough around safety through the UK AI Safety Institute, which is doing really great bits of safety work, which again become important for making products in mass market.

But it's hard for me to find any framing where I don't see the U.S. compounding its advantage for that reason. But that's not to say we can't have enormous winners coming out of Europe. Mistral, which is this Parisian company, has been doing some remarkable things. It has done a couple of things that I thought were really interesting. One was that they were able to get close to GPT-4 quality much more computationally efficiently than the American firms were, which again sort of plays to, like, a Ferrari has a 3-liter engine, is much faster than a 9-liter V12 Dodge, right?

In the '70s. But the second is that Mistral just has done a deal with AXA, which is a big insurance company, to build a really, really deep subject-specific LLM and offer applications in that domain. So you can imagine that there is a path where you can build actually really, really meaningful value. And I've seen a fantastic biological foundation model come out of Paris as well. So you can see those also emerging. The reason why those might not matter is if you really believe you are two years away from AGI and AGI is only two years away from ASI, sort of none of this matters.

But I personally don't believe that, right? I think we've got kind of plenty of time. So I still think that there is room and space for Europe in that world. But for these sort of big general models, without there being a kind of breakthrough bit of science elsewhere, I can't see why the dynamics stay the same with the U.S.

I think it's—I don't know how to interpret that OpenAI took Paul Nakasone, who was Obama's top cybersecurity and cyber offensive guy, on their board, and that there are so many conversations happening with NatSec on these topics, because clearly it's a really important defensive and offensive tool. You asked how it changes things. It changes things because we've got this sort of ladder of escalation where, still in our head, the worst possible thing a nation can do to another nation is to physically lob a kinetic projectile at it.

And so all of the work that happens in that sort of gray space of information operations or cyberattacks has not really been seen as conflict. And in my book, I write that what AI does is it reduces the cost of aggression because it makes information operations much cheaper. You can flood the zone with realistic-sounding tweets. By the way, you have a wonderful Twitter feed. I think you're one of the funniest venture capitalists.

Yes. But you'll be even better now with GPT-4 helping you, right? So you'll be able to flood that zone. Then the second is the ability to use it for cyberattacks, cyber offense. And the third, which I didn't write about but I think is clear, is that as these large language models become part of foundational infrastructure, they would become vulnerable to attacks of particular types. And then, of course, the tooling will make drones, drone warfare, more applicable. So what happens there is the price comes down.

So the desirability for those tools to be used will increase. And so I think that that creates this very, very febrile environment. And we have seen, we are seeing, an increase in information operations. We're seeing an increase, of course, in cyberattacks. I don't think that's necessarily a bad thing. I think that the reason that there are billions of cyberattacks every month is because there's useful stuff on the cyber. And I would rather have a useful cyber that gets attacked than one that just doesn't work and is useless and no one wants to do anything on it.

But it will mean that the CISO, which was not even a job a few years ago, is going to become more and more important.

Well, exponential is just branding, really. So at this point in the game, the asymmetry is that what we've started to see around the Horn of Africa, around the Sea of Aden and so on, is the Houthis using incredibly low-cost missiles and drones, and they're having to be defended against by state-of-the-art weaponry. And we saw this happen in Ukraine, right? A $20,000 Iranian drone having to be shot down by a million-dollar missile interceptor. And the point is, what are you protecting?

And maybe you're protecting something that's worth more. But the trouble is, if we go and look at what's happening with shipping around the Gulf, a U.S. Ticonderoga-class air defense cruiser has 90 missiles on board. So, shoot down 90 drones, and then it has to go to a major U.S. port to be rearmed, which will take it three weeks while it's there. And yes, it'll of course cost half a billion dollars to do that. So the asymmetry is that we've built these big and chunky systems that are incredibly powerful and make for wonderful Hollywood movies, but the threats will be $500 drones.

But it's not even that. I mean, we've seen that 80% of the ships that go through that region are now going around the Horn of Africa. So the U.S. Navy, for the first time since before Top Gun was released in '87, does not have an effective presence in that part of the world. And it's not been a nation state that's pushed it out. But it's this, and you're a software guy. Software does work in powers of 10. When we look at things like RPA, robotic process automation, the first year you did an RPA project, you did one small project.

The next year you did 10, and they were all bigger. The year after that you did 100, and they were all bigger individually. So these things scale by powers of 10. So we start to see drone swarms of 20. Well, getting to 200 is not going to be that hard because it's now a software problem. Getting from 200 to 2,000 is not going to be that hard. Getting from 2,000 to 20,000, and so on. So we have to now start to say, we cannot think linearly and say, well, we dealt with 40 Shaheds, how do we deal with 50?

We've got to say, we're dealing with 40, how do we deal with 40,000? And what are the tools that we need to do that, be it electronic warfare or close-in weapon systems or whatever it happens to be, to protect against those? And right now, there is clear evidence that it's advantage attacker, because all the ships are going around the Cape of Good Hope.

Yeah.

Well, the energy transition is happening really, really quickly, and more money is going into renewable investment than fossil investment. The vast bulk, more than 90%, of all the new electricity generation that was brought online in the US was renewables. Texas obviously has got this amazing renewable story, and the curves are absurd because they just—they do that whenever there's more vertical than horizontal. That's quite NVIDIA, right? Until three days ago. Until three days ago, right. And what's happening, of course, alongside that is the energy system is rapidly starting to change.

So scale is going up, the amount of storage that's coming on is increasing. People are starting to notice that in order to have a renewable-powered grid, you design it completely differently. One thing to think about is the design of the internet versus the design of a phone system. The traditional phone network was physically robust, guaranteed circuits from place to place. And the internet was, let's just build loads of circuits and let's kind of chunk things into packets.

And if there's enough of them, it'll be resilient and robust. And it is. And that's what the renewable grid will end up looking like. And what I find absolutely fascinating is how it's going to require different mental models. So, in order, for example, for a grid that is largely renewable, largely solar with a bit of wind and a bit of nuclear, say in the UK, for that to guarantee us electricity 24 hours a day in December, we'll need 10 times more solar panels than we might need for July because we don't get much sun in the UK in December.

And that's a bit like in the internet, a Tier 2 ISP having 25 peering relationships rather than one. But what it means in July is that, when we have lots of sun, you're going to be producing lots more electricity than the economy normally needs. And that will be free or negatively priced. So if you're an enterprising person, you'll go and say, what thing do I need to do that uses a ton of electricity and is way too expensive that I can wait to do in July rather than January?

Like train a foundation model or use some kind of electrical process to reduce bauxite into aluminum, or dot, dot, dot, and a whole series of things. So I think we start to see a fundamental change in the way in which we consume energy, which is going to be partly driven by deep science and partly it'll be driven by software and optimizations. Now, I know I didn't talk about AI and energy, but I went on for that point.

Yeah. Which goes back to the theme of shifting geopolitical power, right?

Well, the Chinese solar manufacturers, based on some of the analysis that Bloomberg did, have actually got more—can generate more power than Big Oil prior to any major reinvestment by either industry. So that is a sort of Bloomberg analysis. Now, the thing that you deal with, with being a manufacturer of solar panels, is that it's a commodity business, and you have no guarantee for success. And quite likely, a lot of these businesses will struggle to be super profitable even as they get more and more market share and, sort of, in absolute terms, as the addressable market grows.

But I think the key thing for us to understand is that the price of energy is going to come down, and we're going to replace the way we think of energy as something where the price jumps around based on whatever local dictator is misbehaving to something that follows the price of a transistor, which is, it's consistently declined over time and over long periods of time. And think about what that means for a business. It means that right now, if you want 10 years' worth of stable energy pricing, that's an incredibly expensive forward you have to buy in the market.

Whereas now you can start getting 10-year PPAs where you are pretty much locked in, and you get absolute visibility as to where you're going to go. And the next one you buy will be even cheaper, and the one after that will be even cheaper. And for the oil companies, I think the struggle that they will end up in is that they have actually got quite a brittle system. The brittle system being that they've got the minimum cost to recover oil, which is incredibly low if it's Aramco in Saudi Arabia or it's natural gas in Qatar, but it's much, much more expensive as you get to tar sands and other places.

And when that oil-equivalent price for kilowatt hours generated and stored on solar and batteries crosses over, the economic incentive will absolutely disappear for businesses. It'll just make much more sense to bet on the new technology. And when I say it's brittle, it's that it's actually quite a fragile system. So where I live in London, in my area, a number of people moved to electric cars quite early on, and it pushed the local gas station out of its marginal profit into, like, not a great place to be as a business.

So that's disappeared. And now people have to drive an extra mile and a half, which takes about 20 minutes in London, to fill up their cars. So the cost of owning the petrol car just rose. And that process will, I think, start to take hold, and you'll get these tipping-point effects where it's no longer cost-effective, and therefore the inconvenience for the consumer rises. And that brittleness, I think, will exist in lots of parts of the O&G business over time.

So you will see quite rapid switches. They won't be ChatGPT-level switches, right? Because these are sort of physical infrastructures, but they'll be quicker than people expect.

I think that one of the things that we need to talk about more is how a lot of this enables us. And so my little story of this is I've always wanted to be able to DJ, right? Do house music sets, and never tried to do it because I kind of didn't really understand the technique, like all the things that people put in the turntables, and probably wasn't cool enough to do it. And a few months ago, I came across a piece of software which uses AI tools, or machine learning tools, to help with your mixes.

So you go and you buy the music and you analyze it so that it helps you figure out where transition points are going to be, and it can separate out vocals from the instruments, from the bass, and so on. And it just gave me a superpower. Now, I can't produce mixes as well as Carl Cox or Tiësto or any of these big characters, but I can be a decent hobbyist in it. And so it's sort of enabled me. And I think for all of us, these tools, the AI tools in particular, can help us do that.

And we need to talk a little bit more about how they can help us and what difference that actually means to our lives. I think it's a much more appealing narrative than the alternative narratives of either a life of idleness or a life of unemployment.

Thank you so much, Matt.