AI and SaaS

Hi, I'm Tony Cameron. And I'm Benedict Evans and this time I'm recording. Great. First step in the voting podcast. AI changing software. Is that what we're doing today? I think so. Yeah. Yeah. So I was saying earlier, it felt like last week everyone who works in enterprise software and there's anything about it kind of put their head in their hands and said, oh my god, do I actually have to explain why this is done? You know, you'll use code instead of SAP. Fine. If you need me to explain this, I can't help you. So let's talk about something else. And it struck me that there's sort of a couple of more interesting building blocks to talk about here. The first of them is like, I started my career in the doc on bubble. And you had a period of erratic optimism where people said, no, you don't understand. Everything's going to change and it's going to change tomorrow. This is just like right now. And they were right, but not for five and ten years. And it wasn't five or ten years because you had to wait for people to build more about either. It just takes time to build all that stuff and get people to do that stuff. And then the only other side after the spring of 2020, 2020, there was a rational pessimism. And in particular, what would happen is you'd have a story that everyone would just knew this is ridiculous. And the stock would go down 15% or 20%. And the best part was a month or two months later, there would be another version of the same story. Again, and people would say, but we just did that. And the answer was stuff can be in the price and then it can be in the price again. And a lot of the sentiment at the moment is look, the big four companies, platform companies spent 400 billion dollars on CapEx last year. This year, they said they'll spend 650 more or less. Microsoft hasn't been given full regard. That metter and Google and Amazon of more or less double that capital say they're more or less double this cap that CapEx this year. And so you're going from apart from Amazon, these were companies that did not have a lot of CapEx. And suddenly, they're talking about 30, 40, 50% of their revenue going on CapEx. Or as the gelco spends 20%, 15 to 20%. So everyone is sitting and going, whoa, whoa, whoa, whoa, whoa, whoa, whoa, whoa, whoa, where is the return from this going to come from? Are these going to be structurally low margin companies? How long is the business going to come from this? So you get this super nervousness about Microsoft results where Microsoft is a mess, but Amazon results is open AI really going to get the investment from Nvidia. And then you get this anything of Claude having a skill for legal and like all the software companies go up 20%. And then they're back up to where they started at the end of the day. So there's a lot of nervousness around just the amount of money involved here. I think there's also nervousness around the fact that you know, with the dot com, but boom, this was about creating new stuff. When it wasn't obvious to everybody, this was going to be hugely destructive to a bunch of other industries. Whereas here, the first way for all of this is to take a bunch of post at. And so that's a revenue that could go, which makes people nervous. There's a big piece in the FT this morning about how much money the type of equity industry put into buying software companies in the last 10 or 15 years and a whole amount of debt, like half a billion or half trillion dollars of debt around that. And everyone see who were, we're going to be able to report these software companies and now what? So you've got all this kind of financial wall street east stuff that makes and a lot of this sort of euphoria on the other side of 25 year olds who've never worked in another industry and have never seen a financial, never seen an industry cycle and just think everything always goes up and they were born with iPhones in their hands and don't really understand that like stuff changes before. Do you just a question there? Do you think if there wasn't that amount of money being poured into this that people would be a little less nervous? It's interesting in San Francisco hearing people who are in there, you know, 50, 60, 70, saying, I feel like this is like the fourth wave of AI in this time. It's actually sticking and it's crazy that I've been working on artificial intelligence in LLM's for 40 plus years. But there's definitely a different moment right now. But I'm curious where your head is out on that. Oh, so I think there's unpicking different things within that. Yes, this is a breakout. I mean, the last way that AI was machine learning 10, 15 years ago now, 10 years ago, really, that was not transformative across everything. It was a new bunch of stuff that everybody could build and there were a bunch of new companies with it. But it wasn't the same scale of change that we have now. It was additive and a nice evolution about it. It was more like, I don't know, it was like going to gurus rather than the invention of the PC or the invention of the web. It was a smaller wave rather than a big wave. And I think, you know, I should prefer this all of this by saying everything I'm about to say it's going to be a bit skeptical. But like, no, this is an enormously big deal. This is hugely important. I still see people who say, no, it's just like a stochastic parody. It doesn't do causation. It doesn't work very well. It's got hallucinations. It's all useless. These people are more on. Right. They are just idiots. But they're idiots in the sense of somebody who looked at the internet in 1999 and said nobody is going to use this. Not in the sense of people who looked at it in 1999 and said, this is going to change absolutely everything, but their company is too expensive, which in the end was kind of the right for you to take. So that's kind of a, that's kind of the bubbly conversation. I think then you actually sit and look, well, what does AI mean for software? And there's kind of two or three or four different, like fairly obvious building box work and put on the table. The third thing that I mean is, well clearly this is an order of magnitude. Many, maybe several orders of magnitude cheaper to make any given piece of software. Piece of software that you might have thought of 10 years ago that has nothing to do with AI. You can use AI to make that thing much more quickly, much more cheaply. And so that means, if it weren't for any other conversation, that would mean that I mean a way more software, way more problems being automated that weren't automated before, way more stuff getting done inside enterprises, way more jobs that maybe needed people now being turned into software, just software that runs on a database because you couldn't have made a database for that before because it was too expensive. And of course, that means competition for the existing companies and, you know, churn within the industry and so secondly, this software can do something radically different. That's the ball disgusted, great link. So there's another whole class and stuff that will get automated and changed, including stuff that today is a big giant company. Thirdly, then you get to the kind of the delusional idea, which is, you know, Dwight and Fat Keith and Gareth from the office will come in one morning and say, hey, you know, I say he doesn't work very well. I'm going to make my own instead where I saw a post on social media from someone posting the picture of Samuel L. Jackson saying, saying, which you get, say, say, say what again? Don't say system of record again. I dare you, I double day, that like you've got 25,000 people in your company and you all need to be on the same database and you need to have the same processes and the same flow and the same buttons on the screen. No, you can't just have one man and people making their own tools. What's much more interesting and, you know, we could talk about that for hours and we kind of have talked about that for hours. What's more interesting to me and then I'll stop monologuing is was to kind of scratch my head and think, well, today you might, you've got the stuff that systematized and institutionalized and turned into a rigid process in SAP, like a general purpose enterprise software. And then you've also got hundreds of individual single purpose, special purpose, vertical pieces of enterprise software, like, you know, your accounts, payable system or, you know, the thing that tracks graduate recruiting and so on. And then you've got the middle case of Excel and Google sheets and email where like we don't haven't got a dedicated tool to do that thing. Maybe we don't do it often enough. Maybe it's just come up today. You're kind of improvising solutions and sometimes there's improvised solutions will turn into a south company, particularly because of the two previous things I've just said, but often they won't. It's like I need to do it. Answer this question. I'm going to export a CSV and find out the answer because I can't do it in Salesforce or I can't do an SAP or whatever it is. So you've got this kind of improvised middle space. And you know, there's a joke that everything that you see in the file, the screen and Excel terms into a company. And now the answer might be, well, maybe I'll use AI instead. Maybe you'll say to AI, hey, look at this thing and do this analysis. Hey, look at this thing. I'll answer that question for me. And that's a different way of saying this thing's making tools for you than saying you're going to tell the code for you and then get you an AWS account so you can run it, which is purely delusional. One of the sort of, sorry, I was going to stop something else. I just wanted to tell you to say is there's a recurring delusion throughout the history of software that firstly people always think there's going to be a general purpose of abstraction layer that will just do everything. And secondly, people always think that everyone who doesn't write code will like code. We've made a little bit easier than everyone will write code and those are just completely wrong. I mean, there's also an element that those people don't, there was an interesting video the other day of someone just reminding everyone that AI is about writing code. Like, that is what it is. It writes code and I had that realization also. I'm just like, there's still a lot of people out there who know that they don't write code. Well, generalize this to put a slightly different about the daily lives of them having to think about how they could optimize their job. And even if they did, there's a huge difference between saying it's kind of a pain in the ass to do this thing and working out the right way of fixing it, which is why you see very often, as this, I think, Paul Graham had this phrase, the tar pit, where are there are the things that people have tried and tried and tried to make a piece of software to solve that problem and failed over and over again, because it's just really hard to get all the incentive aligned. The underlying point is I think the hard part of making software is almost never writing the code. There are some things where you're solving some hard technical problem. But mostly what you'll do, the problem is working out is realizing that the problem in the company even existed and then working out what would be the right way of solving that and then working out how you would get everybody across the industry to use it. And that's not something that, you know, random middle manager or person at one part of the ecosystem, and so on without going off and making that all they're going to do for the next five years of their lives. Yeah, there's a complexity element there that's really interesting that I think we talked about briefly at some point of just like when we moved from the complexity of being on prem versus in the cloud and how people were thinking about that as a step change as well, which is interesting. Well, the on-term thing is to cloud thing is kind of interesting because this is something that starts in, you know, you could date it to Salesforce if you like, you could argue about when exactly it starts. Yeah. So it's going to be a long time to be on the platform. But then this stuff that comes with cloud. So you do from when you don't have to do an installation in the clients data center. So then it becomes you can have way more software because it's much easier to deploy it. It's continuous deployment. So you don't do a version and then do another version 18 months or two or three years later. You don't have different client. So you don't have to do a lot of work. It's a lot more than you do. One of the things that I sort of think about as I talk to companies about this is as I talk to companies about this is OK, so step one is I need to find OK, so step one is I need to find that data. Step two, step 1.5 is maybe step 1.5 is maybe what data should I be looking for. Steps and that will do is that will steps and that will do is that will steps and that will do is that will find you ask the other them and it will find you the thing in Salesforce or SAP or whatever the data system that you're using is. maybe it will find it across 10 or 20 different systems and it will put them all together for you in a dashboard. And yes, you could have done that with Snap Logic or all these other orchestration project project project tools 10 years ago. And maybe Snap Logic will and maybe Snap Logic will and maybe Snap Logic will use those kind of software will use an LM in order to help you build this. So the step two is build me the dashboard. I'll find me all this information so I don't have to go and build it, hold it all together. But what you're seeing there is a deterministic screen. You're seeing a screen of data feeds that come from your deterministic systems. So you've got this kind of nicks of probabilistic and deterministic systems. But then step three is, hey, I manage a warm up in Jersey and there's a once every 30 years storm coming through. What questions should I be asking? What should I be worried about? Should I build me a dashboard around that? Now that's a different way of solving those kinds of problems. And you're stepping up to a different level of abstraction what you could be asking for. I mean, there's a sort of an e-commerce equivalent here, which is, you know, it's the difference between, I don't know, you take a picture of a coat and you say suggest coats like this. And you say, have a look at my Instagram and tell me what I would like. There's a very different types of questions. You know, you remember the old XKCD cartoon about image recognition. So in those KCD cartoon where boss is in, you know, this is like 2010, the boss has to the engineer. I want to flag every picture if it's taken a national park and the engineer says, OK, latitude longer to block up, encoded in the metadata. I'll map it against the GOS database. I can do that softening. And then the boss says, also, it should say, but if the picture contains a book, it's a book. And the engineer says, I'll need a research group in five years. It can feel like the same question to fundamentally different. Yeah, ask. And it looks like the same question. It's not the same question. And of course, the arms five years later, today, that's an API call and it's 10 minutes. It does this contain a bird. Do an API call to an image recognition model on AWS. And there's a set again. There's a set of sense of what kind of question should I be asking my data? And I think that's the question. And the kind of question should I be asking my data? Yeah, and you get these kind of dumb enterprise software ads where they say, imagine having a conversation with your data. But and there's a sort of, you know, there's a middle point here of know what should actually happen is that the LLM, that the enterprise software should be saying, hmm, this guy is a manager in Jersey and that looks like there's a big storm coming up and you shouldn't have to have coded the is there a big storm, which you just be looking and everything. And seeing, well, there's a storm and it's a big storm and it's a big storm. Hmm, that sounds like it might be good big deal for retailer. Let me look that up. It should be doing that level of predictive analysis. And so you shift the layers of abstraction in how you're using your your systems of record. There's there's that word again, the system of record. The databases fill down the databases, the databases, the database and it's drawing the data in the same way. But what happening around it? How are you getting stuff in now? What are you asking? What can you ask? How can you use that? How can you extend what kind of insights you can get out of that? That to me is a lot more interesting than explaining why no people aren't going to use the tool to make the NSAP. It's a new way of thinking and approaching and problem solving that I think most people are just not used to. Well, it's not used to, but also there's a fuzzy gap of is at your job. We haven't talked about that. Yeah. Should you be thinking about optimizing your work? Is it your job to optimize your actual job? At a certain point, if you're working on the data, you're working on the data and you're working on the data and you're working on the data and you're working on the data. At a certain point, if you're being, I mean, if you're being paid half a million dollars and running a Walmart with, you know, that hundred employees, then it kind of is your job. But there's a big fuzzy question of how much do you as a normal person working for a company after they're thinking about what the software could be doing for you, which I think gets you into this fallacy of like we're going to make everybody use AI, which to me is very like, you know, being in 1998 and saying we're going to make everybody use the internet. Or indeed, it's actually what happened with, with Facebook in the late 2000s, where they said we're going to make everybody use mobile. We're going to force all of our app engineers to use Android's because they've all got their thousand dollar or seven hundred dollar or six hundred dollar iPhone. And that was great. The most of your users have a hundred dollar Android and your Android app is terrible, but you know, you use the Android app and certainly not on a hundred dollar Android. So, you know, you're going to have a hundred dollar Android. And so, there is a layer of like, you can just brute force people over the hump to get them into the new thing. But that only works. That works if your Facebook. I'm not sure that works if you are, you know, you're done eventually. Does it mean just some point though, like, is it your job to optimize your job? But there's also a flip side to that, which is, shouldn't we all want to have a look at what we're doing today and optimize it and be excited by the fact that we could be doing more with less and in less time and it requires less skills and it's less complicated. I don't know. I think that product management is hard and working out what the tools should be and how you should be doing this is hard. And there is a whole profession of thinking, well, what should the product be? I mean, this is, you know, it's this sort of thing that we are going to have to do. And then, you know, you know, we're going to have a lot of work to do. I mean, this is, you know, it's sort of, you know, it's another sort of idiots on social media. Nobody asked for this ad. Nobody asked for anything. Nobody asked for cars or airplanes or electricity. Nobody asked for. After time, we don't know what we need to have. It's not your job to invent everything that you that you haven't find useful in your life. You shouldn't have to. You know, there is a whole field, but we should be in a fine to be curious about the whole field in a career that you're working with. And then, you know, you know, you know, I'm thinking, you know, hey, this would be an interesting thing to do. And maybe people might like this. And maybe this is useful. And that's a different set of skills. I mean, you know, it's sort of thinking about to a company. I remember looking at when I was at a 16z, which is a frame.io. And the, the, the model of a frame.io is you're working. Say you're making a TV commercial or music video or something. How many people and how many different organizations need to see that? Five, 10 organizations, 20, 30 people, of which some subset of people who are actually doing stuff. And there's a different people doing different things. You know, there's people doing the color and the editor and the sound and all sorts of different things going on in that. So how do you manage that? Well, you're FedExing hard disks around the country. Or maybe you've got a Google Drive. If the files aren't too big and you've got enough bandwidth. And then there's a Google sheet with time codes and comments. And a whole bunch of email. And somebody's whole job is to go into the email and put stuff into the Google time sheets and then take stuff out of the Google time sheets and email people. And frame.io says, now we're going to make Google Oxford video. It's not literally Google Oxford and you can't actually edit the video, but it's everything around that. It takes all of it and it seems like I can start saying this. You kind of imagine all the stuff you want, like you be here. And then I say, but and also you can draw around that thing on that frame. And as I say that, you think, oh, of course, would you sort of do that? Would you sort of that? Would I have thought of that? If I had the product didn't have that. And then it's six months or a year or two years as people actually systematically thinking, okay, watch it every screen beat. What is the work flow? What are the problems here? How is it that we build this? What's the right way of doing this? What are the right trade else? How do we get everybody to use this? And now imagine you are managing the color grading of that video and you're thinking, hey, it's a massive pen in the ask me to share these videos. Are you going to build that? Is that what you, is that your skill? Is that how you think? Is that, do you want that to be the next two years of your life? Is that what you do? That's, that's making working, seeing the problem exists and thinking that I've had to solve it is not about the code. That's not the hard part of any of this. I say this is somebody who hasn't written code except for Excel in 20 years. So like turning Kruger and syndroming in action. But, oh, I don't think that the whole part of that is to code. Listening to you talk about this, there's something that strikes me as well as, like, no, I don't think I haven't heard of anyone be particularly annoyed or angry at the idea that AI is going to change software and change the way we code SaaS products. It's interesting to me that every time people do get very emotionally invested in AI and where people do get mad is in the creative space. In the things where there isn't actually a right or wrong, which I find absolutely fascinating. It's people get mad at, you know, AI generated art or music when the reality is that's the stuff we should actually be trying to enjoy. And we don't need to worry about too much worry and putting this in inverted commas. Not from a job's perspective, but we don't need to worry about if it's right or wrong or what I'm ingesting or what I'm consuming is going to put me down the wrong track of just like don't be an idea that well, you know, it's it's completely hallucinating here and it is in some way, but the fact that it may be wrong is just you don't like the art. And I just find there's something really interesting of what people are getting angry at and where we're spending our emotional capacities. Actually, the stuff that we should just be taking in and appreciate it. Well, yeah, I mean, we were telling about this earlier. I'm my sort of pithy, which we sort of craving of this was something that feels where AI is actually what's best right now. As it feels where people are really angry about it and sure it's useless. And it's not actually solving a problem. It's just offering us. And by first, where as it feels where it doesn't really doesn't quite work or needs an awful lot of other stuff around it. As it feels where people are sure that it works right now. And you're like, and that is completely wrong. You actually had another, when we were chatting a couple of days ago, you had a really interesting layer attached to that as well. Which was this question about authenticity, like canned music versus tape and I think that's a nice way of closing out this conversation because I think that question of authenticity actually does tie into the software piece and what people are getting angry about versus what they aren't. Yeah, well, thanks for reminding me. I mean, this is actually kind of a long way from enterprise software and one of the things that I was sort of thinking about was, I mean, I heard I was interested in a podcast about the history of copyright and image and I was really interested in it. I heard I was interested in a podcast about the history of copyright and it mentioned that John Phillips, he's a hated recorded music. He called it canned music. And it did to has struck me often that the people who are most upset, upset about the image generation appear to have no conception. That there are things called cameras. Like, but it's just a machine. You press the button and it makes the art. Well, yeah. Every photographer enjoys that and and yet somehow I can buy the same camera as Cartier Press and and not get the same pictures and like I go and look at people who are doing interesting stuff with other lambs and I think, but I have no I wouldn't be able to get get get get get get image and to make that. I mean, you and I were saying we wouldn't even know where to start in our prompt to get the result that they are that we are looking at right now. I wouldn't even know what words to use what sentences to string together. What yeah. What vocabulary. Yeah, exactly. So the the but the interesting thing here is is that you can have photography that is purely make that you can have photography that is purely mechanistic purpose like many pictures of my kids making a picture of that product and photography design. And what is it that makes it art? Well, it's something to do it. It's not really about the if that exists. It's to do it the intent and the point that we were making about authenticity is like, why do I want that picture? Do I want that picture? Because I value the artistic integrity, the artists journey, the vision behind it, the message that it conveys. Or do you just want a nice picture or I just want a nice picture of my teacher and I just want a nice picture of my teacher and I just want a nice picture of my teacher and I just want a nice picture of my teacher and I just want a nice picture of my teacher and I just want a nice picture of my teacher and I just want a nice picture of my teacher and I just want a nice picture of my teacher and I just want a nice picture of my teacher and I just want a nice picture of my