Name: AI Continues to Eat the World
Uploaded: 2025-05-06
Duration: 16 min 44 s
Description: What matters in AI? What’s going on, what might it mean, and what will happen next?

What matters in AI? What’s going on, what might it mean, and what will happen next?

AI Continues to Eat the World

Conferences

So that's a nice English welcome. Um, in America, there were cheers. Um, I was in Helsinki at the end of last year and they had lasers and dry eyes because it takes a lot to get Finnish people excited. um. So, thanks. I was, I was asked to come and explain everything that's happening in AI and take 15 minutes and not talk too fast, so I'll probably manage one or two of those. I think a good place to start in understanding this stuff is to think about what we mean when we say AI. So this is a classic quote from 1970, AI is whatever machines can't do yet, because once it works, it's not AI anymore. In the 1970s, databases were AI 10 years ago. Machine learning with AI. Now I don't think machine learning is AI anymore. Now it's just software. So AI is this sort of linguistic concept that or kind of thought experiment that really just means whatever doesn't work yet, whatever's just started working. Once it's working, it's just another piece of software that you build stuff with. And so that's kind of a useful question to ask as we look at LLMs, as we look at generative AI, um. Is this just more the same? Is this more software? Um, right now it's very exciting in 10 years' time, it'll just be the way the world's always been, or are we on a track to something much more fundamental to AGI or super intelligence, whatever this means, um. And this is kind of the thing that Bill Gates was trying to express of two years ago when he said, yeah, maybe this is something much more fundamental than the typical transition technology transitions that we've gone through. Here's Sergio Brin saying something similar. We're on the race to AGI. And so the kind of the split here is that if you think about the history of technology over the last 50 or 60 years, it tends to move in platform shifts every 10 or 15 years there's some new fundamental thing that's the center of all the innovation and the investment and the company creation. And so that was PCs, mainframes, then the web, then smartphones, and now the kind of the base case would be to say, well, now generative AI is the thing that we build everything on for the next kind of 10 or 15 years, and then there'll be something else. The other view, the Bill Gates' view, would be to say no, this is a much more fundamental change. This is a discontinuity in those shifts in cycles. This is more like computing or maybe more like electricity, but any rate, it's not just like the new iPhone or the web or the PC, it's something more than that. And so the kind of the range of outcomes that that points to is there's one extreme case that says we have a sort of that computer that runs our whole world, and you can just go through it and ask it how to do your taxes or how to move to Brussels or to buy your house and then there's another extreme that says no, this is just going to end up like software, it's going to end up like spreadsheets or databases. And if you were to say that we need to have a standardization process for spreadsheets or we need to ensure that the governments get together to make sure that the databases interoperable. All this would be kind of meaningless gibberish, and the truth is probably somewhere in between those, but it's kind of important to understand what a wide range of possible outcomes there might be. It's also important to understand how uncertain this is. So Larry Tessler said that AI was whatever didn't work yet at roughly the same time, Marvin Minsky said we were going to have general AI in 3 to 5 years, 3 to 8 years. He was right, we will still have general AI in 3 to 8 years. We may always have general AI in 3 to 8 years. The point is we kind of don't know. And so it seems to me as I I kind of been been talking to groups about this stuff for the last couple of years, all questions that seem to come up around AI have one of two answers. Take your question. Whatever your question is about AI, the answer is either no one knows, or it will work exactly like all the other stuff. It will work just like PCs or just like railways or just like electricity or just like graphic design or the web or smartphones, or nobody has any idea how this is going to work yet, and you'll just have to come back in 10 years and find out what the answer was. So if that is a kind of a framing for thinking about what we even mean and what the underlying debate is, let's talk about what's been happening since GPC launched now kind of almost 2.5 years ago. And what's happened since then, I think at a very high level, is we have a huge amount of activity. We've got lots of new stuff, we've got lots of new buzzwords and acronyms, and people talk about agents, or if they want to sound clever, they say agenttic, which I think is a bit like saying. instead of saying social, it just makes you sound clever. We have reasoning models or chain of thought models, huge increase in the amount of investment. These things are very expensive to build. We also have collapse in the price to achieve any given result. But the most important thing for everybody else is there doesn't seem to be any kind of fundamental note. There doesn't seem to be any barriers to entry. There doesn't seem to be any kind of network effect or winner takes all effect. And so what, what the consequence of that is that we've had this huge increase in investment. This is the thing that gets most people's attention. Last year, the big four cloud companies spent about $220 billion building infrastructure. This year going on their guidance, they'll probably spend something over $300 billion. In fact, last week Meta increase slightly increased their estimate of what they were going to spend this year. Um, it's just kind of worth digging in on Microsoft for a moment. I'm old enough to remember when Microsoft's business was putting a $1 CD in a $5 box and selling it to you for $300. It's kind of a good business to be in. Now they said that this year they plan to spend about 30% of revenue on CapEx. The telco typically spends about 15%, so this has become very, very capital intensive, at least at the moment, very, very quickly. Most of that money in one or another is making its way through to. Nvidia, which has gone from being this kind of interesting niche games hardware company to doing $140 billion run rates of AI chips and still growing, and this is also as title points out, a chart about TSMC because they're using contract manufacturing in order to scale that volume. So we have these lots of investment. We have all this money going into this. We have all of these chips being sold, and what are we getting out of that? Well, we're getting a lot of models. And so this is there's many ways that you can benchmark this. This is just kind of taking one of the more common ones, and you can see what's happened in the last two years is that Nvidia was way ahead and now Nvidia is kind of just the first amongst equals, and there were lots and lots and lots of models out there and there were more every week. And Deepeek, which got everybody very excited at the end of last year, really just kind of proves the point. They took open source models, they took some leak. Sanctions, they spent about half a billion dollars, maybe a billion dollars building infrastructure, and they got a model that beat the results, matched the results of the best labs out there. And you know half a billion dollars is not nothing, but the tech industry is quite a lot, quite a big industry, and there's a lot of people that have half a billion dollars. And so today kind of anybody who's got half a billion dollars or a billion dollars can have a state of the art model and then can have another one a few months later. If you flip the axis on here, part of what's going on is we've got this divergence between the cutting edge and the almost good enough. So you can have absolutely the best possible model, or you can have something that's kind of 90 to 95% as good as that and will maybe run on your phone for sort of. 5% of the cost. So you have this very, very, very steep cost curve and a kind of proliferation of different models, different kinds of capabilities, different kinds of options, different scores, different benchmarks. And if you go and kind of look at the release papers for these kind of things now, it really reminds me of this like the fees and speeds era back to the PCs in the late PC industry in the late 80s, early 90s, when you had to know how many megahertz and gigahertz and how many bytes and how many ports and sockets and things you had. And you had to kind of know which graphic card you used to buy and what was the good modem to buy, and now you don't care. You just buy the Dell or you buy the Apple, and it works. And this is kind of where these models are today. There's a lot of them. There's a lot of benchmarks, there's a lot of scores, but it's not quite clear what that means to anybody else, in particular, because although the models themselves seem to be commodities and everybody in tech spends a lot of time talking about Gemini and Claw. and perplexity. There's actually only one that has any takeup or any kind of real kind of consumer awareness at the moment. Chat GPT is kind of becoming like Googling as a consumer brand, as a verb. It's also interesting instantly looking at this chart to notice that it goes down in summer holidays and then it goes down in Christmas, which gives you a little bit of a sense of what some of the use case for this might look like. But what does this mean? What does all this mean for the rest of us? Billions of dollars, lots of models, lots of graphics, lots of lots of acronyms. Well, there's massive investment. There's a strong sense of FOMO. This is the next platform fear of missing out. This is the next platform, and you don't want to be the company that didn't participate and then was faced with having other companies dictate how you could do business. That's what happened to Meta in the last platform shift. That's why Google created Android. Meanwhile, the model. cells keep getting much better and they also keep getting much cheaper and they're becoming more and more commodities, but fundamentally the market structure of the whole thing is completely wide open. No one really knows how this is going to evolve and I think it's kind of interesting to look at this diagram which was produced in 1995. You can tell it's 1995 because not only does it say cyberspace, it also says the Matrix, which is a very 1990s terminology. But the point is that the reason I showed. is it wasn't clear that everyone, it was kind of clear everyone would have some kind of computer connected to some kind of a network. It wasn't clear it would be a PC and not like a terminal or a television. It wasn't clear that it was going to be the internet and not all of these other things that were floating around. It wasn't clear that the internet would be the web and not all the things that were other things that were happening on the internet, and it wasn't, of course, clear that the value capture was going to come from such advertising and social because they didn't exist for 5 or 10 years later. So it was obvious that something big was going to happen here, but it really, really wasn't obvious where the money was going to be, what the market structure looked like, was going to look like, what the points of leverage were, and what we were going to do with any of this stuff. And that's kind of where we are, I think, today with generative AI, which gets me to kind of my third question, which is what do we actually do with this? And this is quite a difficult question to answer. Um, what do you do with Technology that can automate things that computers couldn't do before in a much more kind of profound difference to previous platform shifts. What do you do if the technology might get things wrong if what we're really doing here is going from deterministic software to probabilistic software. You're going from software that will follow a series of logical steps the same way every time and give you the same result every time, unless of course you're a postmaster in the UK, which is kind of an important point. To software that is probabilistic and statistical. And produces things that are probably something like the right answer. And so I think a kind of a good way, a good framing to look at that is to think that traditional software does things that are easy to explain to a computer. It's easy to tell a computer how to calculate a mortgage. What machine learning solved 10 years ago were things that are hard to explain to a computer. Why is that credit card transaction weird? Why does that person sound angry? How do you translate this from French into German? And machine learning solved that more or less completely in the last. Decade. How would we think about what an LLM does? Well, maybe this is things that are easy to explain to a person or easy to explain to an intern, easy to explain to someone who's got some generalized generalized intelligence but doesn't have context, doesn't have tacit knowledge, doesn't have weeks of sitting next to you seeing how you want the work done. If you could ask a 10 or 15 year old to do this in 30 seconds, an LLM can probably do it for you. If they need to sit next to you for a month, it probably isn't going to be able to do that for you. And one way to think about what that might mean tangibly is to think about wait times this has happened in the past. This is always my favorite example, talking about the history of computing. This is a spreadsheet. These get Dan Bricklin creates VisiCalc in the late 70s. He shows it to accountants and they say, holy fucking shit, this thing just did my week of work in 10 minutes, more or less direct quote from the kind of reactions that people had to this. You can change the interest rate here and all the other numbers change. And that was transformative if you're an accountant. But if you are a lawyer, you would look at this and think, well, that's very clever, but I don't do that. What am I supposed to do with this? And I think that's a good way of capturing reactions that people have when they look at generative AI today. And there are some people who really, really have that use case. The places where this really works right now is in software development, in marketing, in customer support, and of course kind of a broad kind of swathe of early adopters, the people who are always kind of looking for interesting new tools and new ways of doing their work. It's worth in particular. Just pausing and looking at software development where this seems like a kind of a step change in the cost of building software, the cost of getting software companies to market, kind of in the way that AWS was kind of 10 or 15 years ago and again kind of an order of magnitude change in how easy and quick and cheap it was to build software and to take it to market. But for almost everybody else, people are kind of puzzled. There's a small number of people and basically everyone in tech who says this changes the world, this changes I use it all the time every day, but that's sort of 5 or 10% of people. Then there's another sort of 20% of people who use it kind of every week or two, and then there's like 50% of people who've looked at it a couple of times and said, I don't know what to do with this. I don't know how I would use this. And there are some sort of proximate questions here like you've got a blank screen problem. It's not actually clear how you tell the LLM to do it. It's not really clear how you manhandle error rates, where you should put the human in the loop, and in a sense that's not really the user's problem. This is a canonical quote from Steve Jobs. It's actually not the user's job to work out what you would do with a new technology. It's the software industry's job and the company's job to work out what to do with it. So if we think about the progression. That new technologies are going through now. This is the progression we've always gone through. You get the new thing and you make it a feature. You use it to automate the stuff you already do. This is what every big company's been doing in the last two years. Then over time you start to create new things, you start to innovate, you create new possibilities, and every now and then you kind of disrupt and you change the market. So what's happening now is every big company has gone off and hired Accenture, which did $1.4 billion. Of new bookings last quarter, every new startup is out there trying to create things that will change some very specific problem deep inside some very specific industry. Every CIO is sort of looking at this and thinking, yeah, maybe I'll do something about that next year or the year after, something like 20 30% of CEOs have got something deployed now, but 40 to 50% don't plan to do anything until next year. There's a bunch of reasons why you would hesitate around that. People are worried about where the data goes, what the use cases are. They're kind of worried about what layer they put it in the stack. Do you use the LLM to control all the other stuff and get correct answers from your existing systems, or do you use your existing systems to talk to LLMs and so use that to kind of control it, and you always have these kinds of questions. It always takes time. To deploy new software, it takes time to deploy the new thing. If you're in the tech industry, cloud is kind of old and boring and dumb, but it's actually only sort of 2020 to 25% of enterprise workflows. But as you go through that process, you just have way, way more software, way more tools, way more possibilities, way more applications. When we went from the mainframe to the PC to the web, you have an order of magnitude change in how much software and how many things you have that you can automate. And we'll probably have the same thing with LLMs. And so if we go back to looking at big companies, everybody has pilots now. Everybody's got a couple of things in deployment, maybe, and everyone's thinking about what they can do with this and what they could build next. Every new startup is thinking about building stuff around a generative AI. And so if we ask kind of questions like, well, what would you Do with AI or what would your AI strategy or how would you use this, I think it's kind of interesting to ask, well, what kind of a question is that? Is this a CIO question or a CEO question? Well, probably both. Is this a question for Accenture or for Bain BCG McKinsey? Is this top line or bottom line innovation? And the answer is kind of all of those, um. And I think kind of a final point to leave you with as we think about this is that that what generally happens in all of those platform shifts I've talked about in the past is that the tech industry changes everything in some industry. It changes all the dynamics of that industry, and then it leaves, and it's for people. That industry to work out what this means. So what does AI enabled robotics? What does humanoid robotics mean for the logistics industry? That's not really a tech industry question anymore. It's a logistics question. All of the things that AI now changes for publishing or government or for education, those are all questions for those kinds of industries. And those will be the same kinds of questions that we now pose as we think about what generative AI will mean for these industries as well. And so with that, I'll say thank you.