In general, when you get a new technology, you start by doing the same thing that you're already doing, but more. And that's generally not where things end up. Generally, you move to answering questions in different ways and to different kinds of possibilities.
Well, thank you so much for joining us today. I think I've been super thrilled. I've been following what you've been doing for a very long time, and you spent the last 25 years analyzing mobile media technology, worked in equity research, strategy, consulting, venture capital. You're an independent analyst. I think we just had you doing a session at AI at Text Inbox in Cannes, and it was super insightful to see and hear from you. Maybe my first question is: if AI becomes the interface between consumers and brands, what happens to traditional concepts like loyalty, differentiation, brand equity? You talked about the Louis Vuitton example in New York with their building and how people are thinking differently. Like, what's your perspective on that?
It's an excellent question. I don't think anybody knows, and I mean, I don't think it's the case that AI becomes the interface. AI becomes another channel and another way that people can find out about things and another way that things might be suggested to them. We already have many, many channels and many things that do that, and AI will be another one. And so that raises questions around what kinds of choices the AIs would be making and why, what information they have about the consumer, what information they have about brands, how those get shaped and mapped against each other. But I think trying to predict that now is like sort of saying, you know, it's like being in 1996 and saying, well, the web will be the interface between brands and consumers. What will happen? The answer is no one knows.
Nobody knows. Yeah. And one of the things I actually loved about, and you've spoken a lot about this, is that we've kind of been through this evolution of change, and this has kind of happened historically and previously and everything else. And there's so many predictions that people have as we kind of look at it, but not all of them kind of come through across the board. And I think one of the things I loved enjoying hearing you talk about it is also how the way people shop today and how advertising is going to shape. So tell me a little bit more from your perspective. Do you think the shopping experiences will vastly be different? We're seeing places like ChatGPT and others really starting to grow, but do you think AI is really going to impact the shopping behavior from your perspective?
So certainly it will change shopping behavior. How much it changes what recommendation can look like and discovery can look like and search can look like is an open question. Like, how much do we go from finding a link to finding the answer? How do we find out about things we didn't know existed? How do those get surfaced to us in new ways? And we saw multiple waves of that with the internet as we went through, you know, sort of successive waves of different kinds of channel, different discovery paths. I don't think we can know exactly how this is going to work.
Yeah. You know, I spend a lot of time talking to CMOs and CEOs, and one of the questions they always ask me is, like, from marketers specifically, what is one point of differentiation that you would give them as advice on how can they differentiate themselves from the competitors if everybody is using Gemini or Claude or ChatGPT? Where do you think they should focus?
I suppose I think the kind of primary question is how far a proactive product search gets directed in new ways, and on the other hand, how much you're recommended things that you wouldn't have known about, and how much a search might take you into different places. But it's very difficult to kind of know how a completely new medium that doesn't exist yet is going to evolve five or ten years out.
Yeah. So if somebody were to ask you the questions, you would just basically say some things are going to work out, some things are not.
I don't think, I mean, I don't think you can answer these questions. Yeah. You know, none of this exists yet. None of it has been built yet, and nobody is using it yet. So anybody who says that they can tell you how this is going to work is an idiot.
Yeah, I agree with you as well. Yet people are constantly making predictions and asking for the answers. What is one of the biggest things people are missing, you think, when they think about AI's impact on brand, on customer acquisition or shopping behaviors? And obviously the promise, and we talked about it, similar to the internet, and I was lucky enough to kind of go through the revolution, all the promise that it did. We're kind of seeing many things that AI could change, and we just talked about it. But what do you think, what is the one thing that people are maybe missing from your perspective where you said...
So I mean, I wouldn't generally think about it in those terms, but I think in general, when you get a new technology, you start by doing the same thing that you're already doing, but more. And that's generally not where things end up. Generally, you move to answering questions in different ways and to different kinds of possibilities. And you can see that today with recommendation systems, and that wasn't where the internet started.
Yeah, definitely. If you were giving advice to somebody like me, and I said, I'm starting a new professional services organization today or a new agency today, what advice would you give somebody like myself? Like, what should I focus on?
I think you would need to have a thesis on what it was that AI was changing about how that business worked. And that thesis would probably need to be more than we're going to automate away all the associates. But I don't know what that thesis would be because that depends what industry you're in. Is the thesis that you can automate a whole bunch of stuff that required a cost base and so now you can do the same service with many fewer people? Is the thesis that the existing industry is locked into an old way of doing things and you're going to do a new thing? You have to have some plan for, well, what is it that we're doing here that the incumbents aren't going to do, or why do we have a wedge or a lever that's going to break into the existing industry and do it with some different cost structure or some different customer proposition?
Yeah, I think everybody is chasing the is going to be better, faster, and cheaper from that perspective as something that's actually going to happen across the board, which totally makes sense. I think one of the things that you talk about, and obviously Sam Altman actually said it, was one of the slides that you had was around how AI will become a utility, right, like electricity and water and everything else. Do you share that point of view as well, similarly to him?
Well, so there's a set of questions about how foundation models will evolve and how far up the stack they go. And so on one extreme, there are two or three giant world models and they run everything, and you can go to them, ask them to do your taxes or furnish your home. And at the other extreme, this stuff ends up looking like databases and it's embedded in everything that we use and enables many things, but it needs lots of stuff on top of it. It needs structure and UI and thought about how it should work. And the more that it goes down that path, the more that this looks like a utility. It looks like mobile data, say, or semiconductors, that it's an essential component of everything that we do, but you need lots of stuff on top of it. And so that's a path in which this looks like a utility. There's another path in which, as I said, the models themselves can become very general purpose and you don't need to build lots of UI and tooling and product and go-to-market and thought about how this should work. People can kind of just go to the model and say, do this thing for me. Now, I'm very kind of skeptical that that works very well for many use cases, but you know, it's difficult to say at this stage. Then you have a sort of question as to how far the different model companies will be able to differentiate from each other. So even if the model can do the whole thing, if there are five or ten or twenty of them, that looks very different to if there were two or three that had pricing discipline. So there's a kind of a wide spread of possibilities, like how many use cases need the best model? How many of the best models will there be? How long will they keep getting better? How much can the best model do by itself? And all of those get you a range of outcomes, as I said, from like three giant models doing everything to this stuff looking like spreadsheets or databases.
Yeah, that's awesome. I mean, you talked about the models, and you know, from my point of view, I've been pretty surprised with how quickly the independent models are catching up. I think the latest GLM 5.2 looks to be like almost within two or three percent of Claude. Do you think there is a hope for a lot of these independent open source type of models against the large frontier models that spend billions in infrastructure and capabilities and processing and everything else?
Well, so part of the question is how much do you need to have the last two percent? Yeah, because you have this very steep curve of capacity and capability and this very steep curve of cost. And so clearly there will be some things that will work just fine on a cheap, free, open source model from five years ago. And there will be some stuff that will benefit from squeezing every last percentage of performance out of the latest best model. If you're doing some big, complex, ongoing analytics project across a big, complicated company, then you might continue to get more returns from spending more money on a better model. And then there'll be stuff in the middle. And I don't think we know the shape of that curve.
We don't know how many people will be ahead of the curve.
We don't know how long the head of the curve will keep moving. Like, how long will the models keep getting better?
And so out of that comes a range of possibilities. I mean, one option is that, you know, if you think about that curve, it may look like semiconductors.
In what happened in chips is that with each generation it got more expensive and more difficult. And so like 40 years ago there were dozens of companies at the frontier, and now there's only one at the frontier and a couple that are a generation behind the frontier. And so Apple relies on TSMC and Nvidia relies on TSMC. However, your car is full of chips, but they're 10-year-old chips. And the world runs on chips, but most of them aren't the frontier chips running the most expensive, most current processes. And so even though you've got this continued escalation in cost, actually not everybody needs to be there in capability. Now we'll see the chips themselves get cheaper, so it may look like that. On the other hand, it may not. You know, it may look like something else. It may look like cloud, or it may look like mobile, or it may look like operating systems, although it's not clear at the moment how it would look like operating systems because it's not clear what the network effect would be. But there are kind of many different paths, and I think the guiding point can be that something could be very sophisticated and very expensive and change all of our lives. I mean, cost hundreds of billions of dollars every year and change all of our lives and have different outcomes. There are some outcomes where there's basically no money in it. And like telcos are a trillion-dollar industry and don't make very... nobody makes money, and you know, the curves have gone nowhere in 20 years, and all the cool stuff that we knew was going to happen on telcos happened, but not done by telcos. The other extreme is it may look like operating systems, although as I said, it's not clear how because we don't have network effects yet. And then there are other kind of possibilities in the middle that point to different balances of the degree of value capture, the degree of concentration, the degree of commoditization that you get. I think the only thing you can say with all of this is that, you know, it's a famous quote that history teaches us nothing except that something will happen. And there are many possibilities at this stage. It's not kind of baked in that, you know, Anthropic and OpenAI and Google will win the whole thing and everyone else will be paying them rent.
Yeah. Yeah. I mean, that totally, and I think you talked a lot about cost. We're seeing cost translate in terms of like usage and everything else, especially a lot of organizations are actually investing in it as a part of it. From that perspective, I mean, early on it seemed like a lot of the models were subsidizing our usage, and now it seems it's really trying to catch up with organizations. What's your perspective on it? Do you think cost will be one of the primary drivers of adoption and maybe growth of open source models because it will be a barrier for some people to do it, or do you think it will be in between? If it's driving tremendous value and impact on an organization or particular sector, then people would pay the billions of dollars it costs.
Well, so look, an open source model isn't free because you've got to do inference.
And so that's just kind of one dynamic in what the pricing environment...
It's cheaper technically, maybe not.
You're right. You know, an open source model might be less efficient than a proprietary model that charges you more money. So it depends. I think I'd come back to my point about the cost curve here. Do you need, you know, you've got this kind of moving frontier in that you don't really know what the frontier models are going to cost. You don't know how much capacity there is. You don't know how efficient they're going to be. You don't know how much more compute they'll need. You don't know how many people will need the frontier models. And then you've got this curve of, you know, all the way back, as I said, to, you know, the five-year free model. And you know, at the moment we have this crunch because we had this enormous surge in demand around agentic coding. But that's kind of transitory, like we don't know where it will end up, but like everything that drives that is in play because you've got a trillion dollars of capex coming down the pipeline. The models keep getting more efficient. There are always new models. There are new use cases. There's more capacity. So there's price elasticity. So the only thing we can say for certain about the current situation is it's very unstable. But looking forward, the models have marginal cost, and so you map that against ROI and you make a decision over time against does this ROI justify this cost? And maybe the cheaper model has lower performance but higher ROI, or vice versa. And there always clearly will be some use cases where you can justify spending a lot of money on the latest, most expensive model, and there will be many others where, like, no, the model from two years ago is fine.
Yeah. You mentioned in one of your reports that people still use AI weekly rather than daily. What do you think is the missing piece from turning AI from interesting to essential for most humans?
So I think there are some people where this really, really has product-market fit, most obviously in software development. I think there are some people who have kind of free-form, open, flexible jobs with many different things and many things that could be automated. I think there are also things where it's easy to see that you could do that with this. But if you think about like the process of machine learning in the last 10 years, we've kind of spent 10, 15 years saying, 'Oh, I just realized I could turn that into pattern recognition,' or we worked out that this problem exists and we worked out that we could turn that into pattern recognition and solve it with machine learning. And if you go up another level of abstraction, like what every software company is doing in a sense is saying we realized we found this problem that you... And very often the problem that they point to is one that people didn't know was there. You know, there are some companies where they say, 'We've come to solve this problem that you complain about all the time.' But there are many other software companies where they say, 'You don't know that this problem exists, and you don't know that this piece of software is going to make it better for you, and we're going to explain to you, and we're going to spend years persuading you that doing this thing will be better for your company, and it will be good, and it will save you money, and it will be more efficient and have returns and so on.' And so if you think about those kinds of scenarios, then giving somebody a blank cursor, a blank command line, and saying, 'Now make yourself more efficient,' doesn't get you to those parts. Most people are not tool builders, and the people who are good at building tools and good at thinking of those use cases and those problems and then thinking about how to solve them are often not the same people as the people who have those problems. Very often the people who have the problem don't know that they have it and they don't see it. Or they don't think in terms of... And very often the software company that solves it does it by doing something weird that's different from the last five failed attempts. And so all of this is to say, like, working out what the tool should be is really hard. Knowing that the problem even exists is often really hard. Describing how you do the thing that you do every day is actually kind of hard. Most people couldn't sit down and draw a flowchart of how they run through that complex work process. And so then just giving somebody a command line and saying, 'There you are, go off, be more productive,' works to a certain extent for some kinds of use cases. But it's not a great way to re-engineer a company around a new technology.
No, absolutely. You know, it's interesting. I saw this report, and I'll add it as a part of the citation when we do this, and it said that the number one trend of companies that adopt AI in their organization or not is psychological safety for their employees. Meaning the opportunity and the ability for them to try and fail and try and fail. Just because the promise is so big, people expect that to work automatically, and it does not.
Well, so you could go and, you know, there were books and there were courses and people do MBAs talking about change management. You know, you deployed SAP. What happened? Why did it fail? You rolled out PCs to everyone in your company. What happened? Why did it fail? 1997, you give everybody in the company a web browser and you say, 'There you are. Go off. Be more productive.' And that didn't work. And why not? And trying to create work, you know, working out how these new technologies can make a company work more efficiently is hard. And that's even if you're just talking about doing the stuff you're already doing better. When it comes to actually like re-engineering a process around this whole new thing so that you can do stuff you couldn't do before, like that is not the job of middle managers in accounting. That's not their skill set.
No, it's not. I mean, obviously we talked about technology and coding, and it has tremendous impact on the engineering things. You know, we heard Sundar talk about at Google that 75% of their code now is written by AI. We've heard Mark and Bos, CTO of Meta, talk about that close to 95% of their code is written by AI. So as you said, technology has been an early adopter and is benefiting from it. Now, one of the questions I have for you is, do you believe that we'll be producing more software and it will be one-off software just for one use because it's so much easier and cheaper to produce software for cases, or do you think that the software companies will be impacted negatively and kind of almost go away and we'll work in systems and designs and all be integrated straight into the LLMs?
I think the answer is yes.
Exactly. So one of the ways I would sort of think about this is that, you know, any big company, you've got your big iron systems, you've got your SAP and Salesforce and Workday and so on, and then you've got hundreds of vertical tools, hundreds of SaaS apps, probably hundreds of tools you've built internally as well. And then you've got in the middle, you've got kind of Excel and email and shared folders. And there's always a question, like for any given task, is that what SAP or Workday is built for? Can we make them do that? Do we buy a dedicated vertical tool to do that, or do we do it in Excel or in email? And the answer is it depends. So I mean, the example I always think of here is like if you're PwC and you hire 10,000 graduates a year, you have hiring software for that. In fact, you probably have like five different pieces of software for that. If you're a little company that hires like five graduates a year, you do it in Google Sheets. And then when do you switch? And it depends. And when do you build your own tool, or do you build a five-meg Excel spreadsheet that manages this for you every year and tracks everything? How are you managing your cap table? Do you build a five-meg Excel file or do you use Carta? It depends. And so there are these kind of fuzzy thresholds of where you unbundle something from between SAP and Excel and vertical SaaS app and back and forth. And what AI does is, well, now it will be easier to do that in SAP. There will also be ten times more vertical pieces of software being sold to you. And it's also easier for you to do it yourself in email and Google Sheets and spreadsheets and shared folders and ChatGPT and everything else. And so all of those thresholds become fuzzy and everything gets bundled and unbundled, and there'll be way more software, and also you'll do way more stuff yourself. So that's sort of one answer is there's just more software and less software, and the whole industry kind of gets shaken up and turned upside down and held by the ankles and shaken. And there'll be some SaaS software companies that go out of business because that's what happens when you have an industry shift. There'll be ad agencies that go out of business. There will be retailers that go out of business. There will be companies in every industry that go out of business because of this. I think the other answer maybe though is what I've just described is more software, like doing the old thing but more. More interesting to me is to say, well, you know, what if you had a piece of software that could say, go look at all of our app telemetry and all of our customer calls, and because they're all recorded, and then go look at everything in Salesforce, and then go and look at all of our customers' products, and then explain why our churn is getting worse. And so you're looking in five or ten different systems and synthesizing stuff, which is a thing that kind of SAP couldn't do and Excel couldn't do and the SaaS software that you have today couldn't do. So you get different kinds of software answering different kinds of questions, as opposed to just, well, I do that in Excel now instead of doing it in SAP, or I buy more software. It's, can you change the kinds of questions that you're asking? Because that's really what happened with SaaS, and it's what happened with the internet, and it's what happened with...
Excel is — you kind of change what analysis you could do.
Yeah. And you got better at it, and I think you bring an interesting point that I spend a lot of time talking about, which maybe is not the sexiest thing, but about the semantic layer — the importance of having different, disparate data systems and translating them so you can use AI to make a lot of sense of it. I think that's probably the biggest opportunity I've seen across the Fortune 500.
Yeah, it's funny, this is great business for Accenture even though their stock is down by half in the last couple of weeks. And, you know, the funny joke comment about this is that what happens is big companies have spent 20 years not sorting their data out because there was never really enough of an ROI to do it this year, and now they'll sort all their data out so they can do AI, but actually most of what they'll do is SQL stuff they could have done 20 years ago.
Um, and so that's of course what you know. And then, you know, occasionally, like, Snowflake's stock goes through the roof because they say we're doing this, and IBM says we're doing this, and Accenture says we're doing this, and you saw Databricks with our new Lake CDP and everything else.
Yeah, I've always confused Snowflake and Databricks.
We do — I do as well, both of us. But constantly, can they just do a merger? It would make...
They should. That would be a lot easier. You're not aligned on that.
Yeah. Um, there we are. I'll take a commission on that one. Thanks, guys. But you know, there is this sort of sense of how much work it is to change how all this stuff functions. And, you know, do you hire Accenture or Cognizant or IBM or someone like that? And that's one kind of work. Then there's hiring Bain, BCG, McKinsey, which is another kind of work. There's hiring, you know, more dedicated vertical professional services firms — like an ad agency, or like an accountancy firm, or like a law firm — to sit and help you work out how you change this set of processes. Because big companies don't have, like, hundreds of people sitting around not doing anything to do all of that.
Totally. And we talk about it, right — obviously our bet is on our side of that: context and judgment will matter, so be more vertically integrated. You understand that because ultimately you've got to make sense of the data, versus like any data is the same data — it's across verticals and across jobs is different. Speaking of, you talked about splitting jobs and tasks apart, which is great, and again I know this is also hard, but what parts do you think — let's talk about marketing specifically — what parts of marketing do you think are most likely to remain human-led for the future? Because we constantly are hearing marketing automation cloud can just do it, everything else, and I believe we're still far away from that world from my perspective. Would love to hear you...
Yeah, I mean, I don't know enough about marketing to have a good answer on that. Um, but I think that's kind of part of the point: is that once the technology starts diffusing, then those are no longer tech questions, those are marketing questions or accounting questions or law questions. So, you know, you could pose a hypothetical, which is like every law firm will get a bunch of AI law tools and they're all using the same ones, and so now they all basically just do 10 times more work in the same amount of time for the same fee and they'll cancel each other out. Is that a good thesis? I have no idea. Ask a lawyer.
I mean, they won't know either, but at least they'll have some idea of whether that's even the right question. Can you go back a generation? So, you know, before computers, if a lawyer wants a precedent, they're going to have to spend all day in the library. And then they get databases and now they can find 10 precedents in less time, or maybe the same amount of time they found 10 precedents. They bill the client the same, but now they've got 10 precedents instead of one. Trouble is, the opponent's lawyer has also got the same database and has also gone and found 10 precedents and billed their client the same amount of money, but now they've got 10 precedents instead of one. And you go to the judge and you get the same result you would have got before.
And this is effectively what economists call consumer surplus, which is you get more stuff for the same money or less money. Is that what happened? That's clearly some of what happened in law. There's obviously also a bunch of other stuff that happened. And I suppose the point I'm getting at is that this is a tool that will cascade out into other industries, and as it hits those other industries, then you get all sorts of puzzles that are specific to that industry that aren't really about AI anymore.
Yeah, that makes sense. I want to kind of start with the beginning at the end. You spend so much time — decades — analyzing technology shifts and everything else, and we're talking a lot about adoption curves and maturity curves. Where do you think we are in the adoption shifts and curve of AI, and what separates this — or if it doesn't — the smartphone, the internet, and the industrial revolution before it, from your perspective?
Um, well, maybe two or three answers. One of them is each of these adoption curves is quicker, and that's because you're compounding and you're standing on the shoulders of giants. So this doesn't need to wait for everyone to get onto the internet. There are five or six billion people who have got a smartphone. ChatGPT is just a website, so you can get 900 million weekly active users. When consumer internet starts working in the mid-'90s, there are less than 100 million PCs on Earth, so you couldn't have 900 million weekly active users. PCs at that time are like two or three thousand dollars, and you don't have broadband, so it takes time. However, the PC industry didn't need to wait for people to invent semiconductors. Mobile didn't need to wait for people to invent mobile. Then of course there's the stuff that you need people to deploy — you need people to deploy broadband, you need people to build all the software on top of it. But the point is you're always kind of compounding: semiconductors don't need to wait for the chemical industry, and chemicals don't need to wait for mining, and so on, go all the way back. So there's always this compounding, accelerating effect.
I think maybe the second answer is there's a kind of — if you think of these new technologies as following a sort of S-curve — there's a period at the beginning where it's very exciting for the small number of people that know about it and it doesn't really work, or to make it work requires an enormous amount of effort, and that small number of people are saying, 'No, it's amazing.' So there are people in the late '70s who are building their own PCs, and there are people like — I have painful memories of using mobile internet in like 2001.
And it was very exciting, but it —
It was like a GPRS or something.
GPRS. It was very exciting and it really didn't work.
Um, and the same thing with trying to use the internet in like 1993, and you've got to go and buy a computer magazine that has a CD-ROM on the cover that has TCP/IP on it so you can install that on your Windows 3.1 computer.
And then there's a period, and at that stage everything is unclear. It wasn't clear that the internet was going to be the web. First of all, it wasn't even clear it was going to be the internet, because there are all these other networks around. And then it wasn't clear that the internet was just going to be the web. And then it wasn't clear how web browsers would work. So Tim Berners-Lee's original web browser had an editor in it, because he kind of thought this was like file sharing. He thought it was like a network drive that you would navigate to a document and you would edit it. He didn't see it as a publishing system; he saw it as a document sharing system.
Um, because of course it was built to help people share documents inside CERN. And at the same time, you know, Netscape goes public in whatever it is, '94, '95. Mark Zuckerberg is 10 years old.
Larry and Sergey are grad students.
