So I was asked to come and explain AI again at SuperAI, and I always think a good place to start, if we want to understand this, is to ask: well, what is it that we mean when we say AI? This is, I think, a really important observation from 1970 from Larry Tesler, who was an AI scientist, that AI is whatever machines can't do yet, because once it works we say, well, that's not AI anymore, that's just software. AI is the stuff that doesn't work yet, or maybe that's only just started working. And if we think about the history of this, in the 1970s databases had superhuman memory, and people made movies about how databases were going to take over the world, and today databases are software, or indeed legacy software. And 10 years ago machine learning was AI. Today I think machine learning is just software. I would guess probably half the people in this room would say it's still AI, and the other half would say, nah, now it's just software. You use your phone, you take a picture of your child or your dog, and you can search for it, and 15 years ago that would have been witchcraft, and 10 years ago it was AI, and now it's just software. And today we have large language models and generative AI, whatever we want to call it, and maybe this is AI, or if you're Sam Altman trying to get out of your Microsoft contract, you will say it's superintelligence. For everybody else, is this AI, or in 10 years' time is there still just going to be more software? Now I think we don't know yet. Bill Gates here, two years ago, three years ago, saying well in his lifetime this was the biggest thing he'd seen since the graphical user interface. So Bill is suggesting no, this isn't just more software, this is something else. Sergey Brin, co-founder of Google, saying something similar, that we're on the track to AGI, that no, this isn't more software, this is something else. And so if we think about the progression of technology over the last 50 or 60 years, we've tended to move in cycles. So we have a platform shift every 10 or 15 years. We had mainframes, and then we had PCs, and then the web, and then smartphones, and you could put other things on here. You could put databases or client-server or open source, and each of these pulled in all of the investment and innovation and company creation. The old thing is still there: airlines still run on mainframes, banks still run on mainframes, but the new thing is where all the innovation, all the investment, all the change, all the company creation happens. And so the view is that now this is the next platform shift: generative AI is the next thing after smartphones, after the web, after PCs. But there's also a view that says that no, this is something fundamentally different. This is breaking the cycle. It's a much bigger change than that. It's something more like electricity or computing or fire, or like the next generation of humanity or something else. And out of that I think you get a range of possible outcomes. It might be that this is just more software, that there will be thousands of models just as there are thousands of databases or thousands of spreadsheets, millions of spreadsheets, and then at the other extreme it might be that we'll have kind of one giant world computer that runs everything, and you go, you can go to it and say like, how do I move to Singapore, or how do I buy a house in Hong Kong, or how do I do my taxes in five different jurisdictions, and it'll just do it for you. The thing is we don't really know what's going to happen there, and we have a range of possibilities and a range of opinions. And it's important to understand we've kind of never quite known how this is going to go. So Larry Tesler on the one hand said AI becomes just software; on the other hand Marvin Minsky at the same time said that in three to eight years we'll have a machine with the general intelligence of a human being, and he was right: in three to eight years we'll have a machine with the general intelligence of a human being, and he may always be right. Check back in three to eight years and see. We'll find out. Either the machines will be treating us as pets, or this will have become software again. And so if we look at the conversations that are happening in the tech industry now, there are these two very different ways of looking at it. On the one hand you have people saying this is going to transform humanity, this is going to change the nature of society, and on the other hand you have people taking these models and making them faster and using them to transform the reconciliation of overdue invoices inside large telecoms companies. So you have this kind of widespread possibility, and I'm mostly going to talk about the first of these. So let's talk about model building. ChatGPT dropped two and a half years ago, three years ago, and what's happened since then? We have much more capital, we have a pricing collapse, we have lots of new exciting stuff, lots of buzzwords, lots of acronyms, lots of new terminology and jargon. We don't appear to have any obvious moats there; there don't seem to be any winner-takes-all effects, any barriers to competition, and we have a huge surge of open source. If we first look at that surge in capital, the big four platform companies last year spent about $220 billion building infrastructure. This year they've said they'll probably spend something over $300 billion building data centers. So we've got more than a doubling of the investment in their, investment in infrastructure in order to build platforms to make this work. It's interesting looking at Microsoft. I'm old enough to remember when Microsoft's business was putting a $1 CD in a $10 box and selling it to you for $250. Now Microsoft is spending over 30% of its revenue on capex, which is double what a typical telco spends. So we've got this enormous surge of investment. A large portion of this, as we know, is going to Nvidia, whose chart is still going up and to the right, as the caption says. This is also a contract manufacturing chart, and so Nvidia is now doing over $45 billion a quarter selling AI chips into the data center industry. But what all of that investment has produced, by and large, is a bunch of models that look kind of like commodities. When ChatGPT launched, they were the first ones out there, and nobody else had a model that was anything like as good. Today, there are lots of people who are making models that are pretty much the same level of quality. Sometimes OpenAI are ahead, sometimes they aren't, but there's a lot of models. Indeed, I think what DeepSeek proved six months ago, DeepSeek just kind of proved the point that anybody with half a billion dollars can have a frontier model, and half a billion dollars isn't nothing. Only a few people in this room have half a billion dollars in the bank, but the tech company, the tech industry is a big industry. There's a lot of big companies, there's a lot of people who can afford to invest that kind of money if they really want to. On the other hand, what's happened in parallel to this is that the models themselves have become much cheaper. So we have an order of magnitude, maybe two orders of magnitude, decline in the cost to produce a given result every year. This is an observation someone has made that LLMs are both the most expensive and the fastest appreciating asset in tech history, possibly any history. But if you kind of step back from all of this, we've got all of these acronyms, all of these charts, all of these measurements, all of these benchmarks. This all looks very like the PC industry 20 or 30 years ago. These are a couple of examples I found online: 191 of the fastest 486 PCs, which would be sort of sometime in the early '90s, 19 of the fastest 56k modems, and there was a time when you had to know what all the acronyms were, you had to know how many MHz and how many gigabytes and what chips you were buying, and then it all got abstracted away and you didn't need to know. You had this wave of furious competition, and then it got abstracted away and the value moved somewhere else, and that's probably what's going to happen with this stuff as well. It all becomes commoditized and the value moves somewhere else in the stack. It's also worth pointing out, though, that although having the models themselves seem to be commodities, consumer awareness doesn't look like a commodity. So here we have a model score versus app store rank and Google Trends. Perplexity, which gets a huge amount of buzz in the industry, it doesn't even rank on either of these charts. It does seem like ChatGPT is on the way to becoming a verb in the way that Google became a verb. On the other hand, it may be becoming a verb in the way that MySpace became a verb. The differentiation appears to be in what you build on top, not on the engineering and the technology, at least at this stage. So if we pull those threads together, like what does this mean if you're not an Nvidia shareholder, if you're not a TSMC shareholder? Well, there's an enormous amount of investment, an enormous amount of fear of missing out. The models are getting better and cheaper, but the models seem to be commodities, and for anybody else, the models are just there and available. You can probably download them, you can run them on a Mac, and the value moves somewhere else. The value moves up the stack, and that gets, I think, to the kind of the more interesting set of conversations, which is around deployment. Like, what do we do with this? How do we use this? How does this get used? What do we spend on it? And whenever you have a new technology, there's a, there's a set of puzzles around what is it that this is, how does this work, how do we understand this? And I think there's a kind of a useful way to think about generative AI is that traditional software is deterministic. It follows logical steps. It does things that are easy to explain to a computer. It's easy to tell a computer how you calculate a mortgage. It's hard to do it in your head, but it's easy to tell a computer how you would do that. What machine learning did was move this to a statistics question, to move this to a probabilistic model, and machine learning solved things that were hard to explain to a computer, like why is that credit card transaction weird? What does this Chinese word mean in Malay? What does this Japanese sentence mean in English? Why is that a picture of a cat and not a dog? Why does our robot keep falling over? And so machine learning solved that. Large language models, again, are probabilistic, and it seems that what they're good at right now is things that would be easy to explain to a person, things that would be easy to explain maybe to an intern. Anything where you could ask a human to do this for you in 10 seconds and they would come back with it the next day, you can probably ask an LLM to do, and it will give you an answer. Anything where it would take you a half-hour meeting in order to set up the project and you need to give a bunch of context and have a bunch of questions, LLMs probably fail at those, not because they can't do the task, but because we can't explain the task. It's hard to explain. All of that kind of gets us to these, these questions that every big company is asking now: is this all sounds very cool, but what is it that we do with this? This is something that computers could never do before, except that they might get it wrong because it's a statistical probabilistic system. What do we do with this? How do we use it? What happens when you open ChatGPT and you have this blank screen and you don't know what to ask? If you ask it and what you get back isn't quite what you wanted, how do you fix that? How do you deal with the fact that these models have a probabilistic answer, and so the answer will not always be exactly correct? Where do you put the human in the loop? What's the right way to use this to build software? At a very kind of conceptual level, I think it's interesting to ask: do you use the LLM to control all the other software? So do you let the LLM talk to Oracle and Salesforce and SAP and get answers back and tell them what to do? Does the LLM sit at the top of the stack, or is the LLM just another API, and Oracle or Salesforce can use a Claude API to do some sentiment analysis or to draft you an email, but all of your source of truth and all of your fundamental control should sit in your traditional software? The answer to that is probably yes, depending on what you're trying to build. But if we step back from this and step back from kind of the AI bubble of everybody who's using this all the time and talking to everybody else who's using this all the time, it's kind of important to remember that most people are kind of puzzled. So this is a bunch of survey data that I pulled from the end of last year on how many people are actually using a generative AI chatbot, and I very carefully use that definition so that you're not capturing people who've taken a picture in Snap and put a cool filter on it. And the answer is that sort of something between sort of 7, 8, 9, 10% of people are using this every day, and another kind of 20% of people are using it maybe once a week or every two weeks, and then probably double that used it once and thought, I have no idea what to do with this. And this is kind of a really interesting kind of statistic because it says there's a lot of people who use this, understand it, get it, they've got the account, they go to the website, and it's useful maybe once a week. And that isn't what you hear from people who think this is going to transform humanity. If it's going to transform humanity, why is it something that I use every week rather than something that I have open all of the time? If we look at how enterprises are thinking about this, you get a slightly different picture, which is the kind of the series over on the right here. What do I do with this? That's the lowest. That's not actually a big reason to hesitate. The real reasons to hesitate are, well, we get it, but it's hard. We have to worry about security. We don't have the in-house expertise. We're worried about the data rates. We haven't got our data in order. It's actually more difficult to deploy this, even if you can just understand what the use cases might be. Now, however, I think it's kind of important to remember that the future always takes time. This is a chart I love of enterprise adoption of cloud, and if you're in the tech industry, cloud is kind of old and boring and done, and it's a thing from the 2000s, but it's still only sort of 20 to 30% of enterprise workflows, and the CIOs are always desperately hoping that they'll get the CFO to give them budget to replatform in three years, but they never actually manage to get the budget. It does actually take time to deploy new technology, and we should remember, like, LLMs didn't exist three years ago. Like, it takes a while to build software, and then it takes a while to get people to deploy it, and we should expect this to take a bit longer. And we should think that more generally about any new technology that we had these kind of questions with every platform shift. And I was a baby analyst when 3G was the big thing, and analysts and investors would keep saying, well, what's the killer app for 3G? What should we do with 3G? Why would consumers buy 3G? And the answer was that having the internet in your pocket was kind of a big deal, and smartphones replaced PCs. The same kind of questions came up with the cloud, like only an idiot would trust Amazon or Google with their data, and the same kind of questions that come up now with LLMs, that this is kind of weird and strange and different, and we don't want to do with it. But then it's always weird and strange and different, and we don't know what to do with it. So maybe actually we should step back and think, well, what's the standard pattern for how you deploy new technologies? And to begin with, with every new technology, you absorb it, you make it a feature, the incumbents add it to all of their existing stuff, you use it to automate the obvious easy things where you can see that that would be useful immediately, you do the early obvious stuff. A little bit later on, then you start innovating, you start thinking of new things that you could do, new things that weren't possible, you move from bottom-line innovation to top-line innovation, startups get founded that unbundle the incumbents or unbundle existing business models, and every now and then there's a third step where somebody actually disrupts the industry, changes the question, asks different questions, and creates Airbnb as opposed to putting hotel bookings online. But that takes longer again. I think a kind of a useful way to understand this is to go back and look again at a platform shift, which is to look at spreadsheets, which some of you may maybe heard me talk about this before, but I always think it's a really useful thing to look at. This is VisiCalc, which was the first successful software spreadsheet in the late 1970s, and before this, spreadsheets were paper. You can buy paper spreadsheet, and so with VisiCalc, you can change the interest rate here, and all the other numbers change. And if you were an accountant in 1978, you would see that, and that would transform your life because it could do a week of work in an afternoon or in five minutes. But if you were a lawyer and you looked at this, you would think, well, that's great, but I don't have that use case. I don't do spreadsheets. Mike should see it, but I don't do that. And that's kind of the stage we're at today with LLMs. There's a bunch of people who look at this and say, "I have that use case." So we're already seeing huge adoption in software development, we're already seeing huge adoption in marketing, there's a bunch of people very excited about this in customer support, although you kind of have to be careful because it will tell you what the answer looks like rather than what the answer is, and you see a lot of adoption in early adopters, probably many people in this room, which to me looks a lot like people who used Coda or Airtable or any kind of no-code tool. There's a certain kind of person that has a certain kind of job, a certain kind of task, a lot of flexibility in how they do that, they're looking for new ways to do that, and so those people are all using ChatGPT and Claude and Perplexity and Gemini and Copilot. But for everyone else, there's kind of puzzlement. It's not clear what we would do this, do with this. It's worth, incidentally, I think, just kind of double-tapping on software development. It does feel like even if nothing else happens, using AI to generate code is kind of the new AWS in that this is a new step change in what it takes to get software out of the door, how quickly you can deploy software, how quickly you can build software companies. But for everyone else, there's this sort of puzzlement: what do I do with this? Where there's, we have this canonical quote from Steve Jobs: it's not actually the customer's job to know what they want, it's not the customer's job to know what this new technology is supposed to be useful for, it's our job, it's the software industry's job, it's entrepreneurs' job to, it's CIO's job to work out what you would do with this. One part of that is that you go to someone like Accenture, you go and you build new software in the enterprise, the CIO and the consultants work out the new tasks that you could do with this, and so Accenture booked 1.4 billion of generative AI bookings, reported 1.4 billion of generative AI bookings last quarter. There's probably a little bit of AI washing in that, as every CIO and CEO wants to say they have an AI strategy, and every Accenture engagement manager wants to report this as generative AI rather than legacy software, but that kind of reflects how enterprise software works. The other extreme is that startups create new use cases. This is kind of some examples of startups that are using LLMs in order to solve very, very specific problems deep inside enterprises, like converting COBOL code into Java, reconfiguring telco billing systems. This isn't the future of humanity, it's just more software, but it's still really useful. And so if we look at the startups coming through Y Combinator in Silicon Valley, basically all of them are AI startups. Every big company in the same way is looking at this, experimenting, building pilots, sometimes moving those pilots into production, sometimes not, mostly concentrated in software and customer service, but again, it's kind of early. And of course, it's always early. So this is what CIOs expect to deploy, already have deployed. Typically about a quarter of big companies have got something deployed now, another 40% don't expect to do anything until next year, maybe longer, and that's kind of always how it works. You go through these platform shifts, and you go from having one app to dozens of apps to hundreds of apps, and you automate more things with every step. And so that gets to me to this question: well, what's our AI strategy? What should we do with AI? And the answer is, well, what kind of question are you asking? Is this a business process optimization question, or is this a sort of fundamental existential threat to your company? And the answer is, yes, all of the above. Existential threats to your company can look different. We always talk about Uber and Airbnb as redefining what a market looked like. It's kind of interesting to compare what Uber and Airbnb actually did. Uber demolished the existing taxi industry. Airbnb was kind of mostly additive to hotels because it was such a different product. So disruption can look different to different industries depending on the technology. We have a lot of those kind of questions now with LLMs and particularly with search. The classic PE question people always ask is, what happens if I go to Google or ChatGPT and just get the answer instead of doing a web search and getting a bunch of links? How does that change the internet? The more interesting question to me, I think, is, well, what happens if you just completely change the question? So what I've done here on the right here is given ChatGPT an image of my fridge and said, what should I cook? That's not a Google query, that's not something that would produce links, that's changing the nature of the question and how we change how we think about what this might be. And of course, that's also redirecting a trillion dollars of annual ad spend, about half of which currently goes to Amazon, Meta, and Amazon, sorry, Google, Meta, and Amazon. If we change how the internet works, if we change what you mean when you ask a question or you search for something, then all the money moves. And there's a whole bunch of places where this is going to change things. We just heard about robotics, but the interesting things I think that I took away from listening to the robotics conversation is there's an AI question here, but these are mostly logistics questions. What do humanoid robots mean for logistics? Ask a logistics person. How does, what does AI mean for Hollywood? Ask a Hollywood person. And the tech industry changes the industry, and then everything else changes around it. So just sort of some concluding thoughts, and then I'll, I'll give up for time. After two or three years of talking to AI at events about AI at events like this, it strikes me that all AI questions have one of two answers. The answer is either it will work like every other platform shift. Should we go with Microsoft or Google or a startup? Should we hire Accenture or Bain? Well, how did you do it the last five times? The other class of question, no one really knows the answer, and that's also the, the way every platform shift worked. This is a chart from a diagram of the early internet from 1995. No one knew how this was going to work. No one knew that it was going to be the internet. No one knew that the internet was going to be the web. No one knew that the value capture in the web was going to be in search advertising and then in social. We're kind of in the same place now with LLMs. We know there's going to be some AI stuff, but we don't know the market structure, we don't know how it's going to work, we don't know where the value capture is going to be, we don't know where the money is going to be. Meanwhile, all the other stuff we were excited about before this is still there. So e-commerce is now 20 or 30% of global retail. Google actually has autonomous cars working, maybe. And so all the stuff that we were excited about before is still there, and all the stuff that we were excited about for the future is still there. So now we talk about crypto and robotics and space and nuclear, but meanwhile most of the rest of the economy is being disrupted by ideas from 15, 20, 30 years ago, ideas like e-commerce and streaming and SaaS and workflows. And so we have this kind of progression across the tech industry that we're always excited about the things that are going to happen in the future and the things we're building now, but all of the money, the trillions of dollars of revenue, is in ideas from 20 years ago, and that's probably how it will always be. And with that, I will say thank you.
