Transcription
[Music] Hi everyone, welcome to the Further Faster podcast. My name is Bead Moore. I'm a partner and chief commercial officer here at Antler. Today we're coming to you from Sydney where I'm speaking with our guest Ash Fontana. Ash launched Angel List's fundraising platform, which currently manages over $171 billion. He was one of the founding members of Zeta Venture Partners, the world's first AI-focused fund. He was the first or largest investor in category-defining companies such as Canva, Kaggle, and Tractable. And he's also the author of "The AI First Company," about how to develop a competitive advantage with AI.
As a reminder, if you like what you hear in this episode, please subscribe. Ash, so good to have you join us.
>> Thanks for having me. I wanted to start by taking a lightning-fast tour of the last 5,000 years in the history of technology, as quick as quick as you can. And I want to talk about why it is relevant to the current age of artificial intelligence.
>> Yeah. What I do a lot in the book and in general is just try to give a way to think about large swaths of history or particularly big ideas like machine learning. And I think one way or one framework in which to view the history of technology is just simply with the concept of leverage in mind. You know, a lever being the Archimedean sort of sense of a lever. You know, it gives you an ability to move something that you wouldn't otherwise be able to move just with your own might. And really, what technology is, it's the invention of tools to give us leverage.
And if you think about it that way, if you use that lens, you think about the first generation of tools that gave us physical leverage. And so they were things that quite literally were what we think of as a lever. Um, but also physical leverage in the sense that, you know, a weapon gives us the leverage physically to kill an animal that's bigger than we could wrestle with our bare hands. And you go through to the industrial era. I mean, that takes us through pretty much 4,800 years of history of technology through to the industrial era where what is a factory? It's just a collection of lots of little mechanical levers that allow us to build things that we couldn't build quite simply with our own hands, like tables, steam engines, whatever, chairs, everything.
The second wave of leverage, I think, and there that was granted by technology of tool-building, was intellectual leverage that we got through computers and calculators more broadly. You know, we could do more calculations than we could do by just writing them out using a machine. And so obviously, a lot of these early computers were things like what Ada Lovelace and Babbage imagined and then invented, but then moved through, moves us right through to the start of the 20th century with IBM making actual computers and then the real computer as we think of it today, the von Neumann computer, where you have a notion of something going in and out of memory and then being processed. And they gave us intellectual leverage in that they enabled us to think through problems in ways that involved many more steps than we could hold in our head, but also gave us the ability to then make some basic predictions.
And this brings us into the third wave, I think, of technology, which has really only been the last sort of 60 or 70 years. And it started really with why von Neumann wanted to invent the computer, which was to predict weather. And that was something that took a while to do really well, but at least it was there at the beginning. And what is this third wave really giving us? It's giving us decision-making leverage. Leverage over the future. And that's what's really powerful about it. You know, the first wave of tools that gave us mechanical leverage really gave us some power over our very immediate existence. Were we able to eat or not? Were we able to hold that sort of stuff? The second wave that gave us this intellectual leverage gave us a little bit more breathing room, so to speak, like an ability to make decisions a little bit further ahead of time. You know, where is all my stuff? How much does this person owe me? Using a calculator. But this third wave that gives us decision-making leverage gives us a lot more breathing room. And that's the idea. You know, if we can predict the weather, the weather, we can predict well ahead of time if we're going to have enough food. Are we going to have enough food next year?
>> Or are we going to have enough steel to make whatever we need to make next year? Enough steel to make enough cans to put the food in, so to speak. So they're the three waves, roughly, I think. And you know, of course, this is just if you look at things through the lens of leverage and you think of technology as the building of tools to give us more leverage.
>> Okay. So, you've got these three waves. The first wave, Stone Age, it is physical leverage.
>> Uh, you have a second wave that's the printing press, that's informational leverage. And you have then this third wave that we're obviously now in, which is computational prediction, which is temporal leverage, right? I think you describe it, right? And I think that that winds us into the kind of current age, which you describe in a terminology that I think will shock nobody, the AI century.
>> Yeah. And you say though, and I, this thing I think is particularly interesting, is that we've now been in this for 75 years in essence. And so...
>> Can you explain a little bit about how that's evolved and and what kind of takes us to the current day?
>> Yeah, I think what a lot of people don't realize is the AI-first century started in 1950, not 2025. Everyone feels right now like AI is exploding. And it is because we've already been working on it for 75 years. And actually, the AI-first century, in my view, will probably end sometime in the next 5 to 25 years, before 2050 for sure. What does that mean? It means we will have fully exploited AI and we'll be in a whole new realm, which we might cover later in this interview if we have time.
But the '50s were when we really started understanding this notion of a neural network. And a lot of people came at it differently. Some people came at it by trying to understand the nervous system. So literally shocking frogs, applying electricity to the nerves of frogs and seeing how their muscles moved based on that electricity going through the nervous system. Then trying to sort of understand what a neuron was and then recreate it on a basic computer. And so there were two guys, McCulloch and Pitts, they were really the first to do this. Then once they had like a very basic neuron working on a computational substrate, like with an old valve computer, then some people that were studying in a similar area, like physically in Boston, such as Marvin Minsky, started thinking about, all right, what if you have multiple neurons connected together? You have a neural network. And he started calling it the "society of the mind." And what other minds could you create if you had multiple sort of artificial neurons there?
And then there was famously a bit of a dark period because a lot of what they thought would happen didn't happen, and then research funding was taken away and whatnot. In other parts of the world, like in Canada and all over Europe, people continued to develop this notion of a neural network and made some interesting progress right through to when you get to 2012-13 and people started working on what was called, or well, making work, deep neural networks. So multiple layers of these neural networks. And that's when really the modern AI era started, when you had deep learning computer vision systems getting amazing results in image recognition, and then that being transposed over or transferred really over to speech recognition, speech generation.
And then you get to today where we have these transformers that are versions of these deep learning networks that transform a lot, in very simple terms, to be able to understand huge volumes of text, like the meaning, multiple layers of meaning, in unstructured data. But the point is, it all started in the '50s. You know, language translation at IBM in the '50s and '60s was a big area of research, and that was machine translation of text, which is what we're mostly doing today in the AI industry, which is doing it with much more powerful computers, much deeper networks, and a much better understanding of how we can instantiate and run those networks on a computational substrate.
>> So, it's been going for a while, and we're really getting to the pointy end, I think, of the AI-first century now.
>> Yeah. Yeah, I mean, you and I have been talking recently, and one of the points that you made is, you know, Google as a company was set up with this in mind, uh, whatever it was, 27 years ago.
>> Talk about how they have positioned themselves from the very inception to be an AI-first company, and yet really, you think that the latest phase or the fastest acceleration has happened in the last decade. How did they kind of preempt that, and what do you think is the change that happens in that last 10 years?
>> Yeah. So, PageRank, which was the algorithm that Larry Page invented that then formed the very first, was on the basis of the very first Google search engine, is essentially an algorithm that can be machine-learned. Just to refresh everyone's memory, if a lot of people refer to something, it's probably a very relevant thing. And what you can add on top of that is, okay, a feedback layer, which is, we'll suggest that this is probably relevant. And then if a lot of people click on it, then we confirm its relevance. And if a lot of people ignore it, then we say it maybe is wrong, and we'll just lower it in the rankings. So, PageRank, what Google's based on, is an algorithm that can machine-learn pretty quickly.
And what they did was they realized that, and from day one, they said, if you look at the earliest founding documents of Google, they say, "We're an AI company." And Sergey Brin especially was very into AI very early on, and still is today. Today, he works on Gemini and came back to Google to work on Gemini. Google was set up that way. And then they put a lot of effort into building the distributed computing systems to run these algorithms, because they realized quite quickly that they're quite computationally intensive. And so they developed a competitive advantage in running those more cheaply than anyone else. And then, of course, they have been really good about acquiring data sets or building products that acquire huge amounts of data really quickly, like Google Maps, like Android, etc. And also the investments they've made in Waymo, having cars driving around and collecting a whole bunch of data all over the place through the sensors that are on those cars.
So, if you think about Google's founding principles, the algorithms, what they put a lot of capital into, all their initial capital into, if you think about where they put their capital in terms of acquisitions, really, it's a data learning effect. They've acquired a lot of companies to get a critical mass of data. They've built information processing systems to turn that data into information. And they have invented a whole bunch of very interesting algorithms that can learn over that information. And those three things are a data learning effect. And Google has very deliberately built that from day one, which, you know, has allowed them to be the leader in so many ways. All the breakthroughs really over the last couple of years have come out of more of them have come out of DeepMind than anywhere else. You know, they've got a lot of firsts there. The first protein folding, the first winner of an international math Olympiad, etc., etc., first winner of a Go competition. And they have really the only chips yielding very interesting and cost-effective results in running models, their TPUs. So, at every layer, they've got something going on that's quite far ahead, or at least doing something good and different to a lot of other AI labs.
Let's just take a moment to kind of go back there because you just talked about data learning effects, and I think that this is an incredibly useful concept for anybody who is building an AI-first company to kind of grasp and understand. Perhaps explain that for us a little bit more deeply, and then maybe we can even do a couple of practical examples of how it might work in practice.
>> Sure. The data learning effect is a critical mass of data, a process to turn that data into information, and a way to use that data such that you learn over it and it improves itself over time. So, you might recognize some of these components. You know, critical mass of data sounds like economies of scale. Those people who study competitive strategy. And turning data into information sounds like a learning effect. And you're adding that data to or feeding that information that you get from the data into an algorithm such that it gets better and better over time with incremental data sounds like a network effect. So, really, a data learning effect sort of combines three existing sources or traditional sources of competitive advantage and actually shows how it's a completely different source of competitive advantage.
But there's nuance there, like it's not really about economies of scale to data. It's about critical mass of data for the given algorithm. So, there's a lot of interplay between these three things. The point is, in the AI-first century, there's a new type of competitive advantage. It's called a data learning effect. They're the words to use to describe it. It involves all those three things. And if any one of those three things is missing, you don't have it. And I think it's really important to understand this source of competitive advantage because it's way more powerful than any of these individual sources of competitive advantage. Just having a whole bunch of data doesn't necessarily mean you've got a defensible business on your hands. You have to have all three. But if you do have all three, it's very powerful because it reinforces itself over time. And again, the clear evidence of this is the best business on the internet, which is Google. It's a complete cash cow, even 27 years later.
>> I want to talk about the practical implications of that for founders who are building today, because I think that that is, uh, really the most interesting. But I suspect that there are also a bunch of people listening who are like, what are the implications for my company if my company did not get started with the sort of foresight that Google got started, and a lot of other companies got started? So, if you are a company that doesn't have a DLE functioning today, what does that mean, and how do you course-correct?
>> Look, what it means, I think, depends on how close your company is to a physical product or the real world, because DLEs are much harder to get going where there's a physical instantiation of a product, as in, it's much harder to get data about a process that happens in the real world, because you have to put sensors up everywhere, cameras, microphones, gas sensors, etc. Like, if you think about a manufacturing facility, if you want to use machine learning to automate or improve the efficiency of a manufacturing facility, you've got to know what's going on in that facility. So you've got to put sensors up everywhere.
So I think in industries where there's very much a physical basis or instantiation of production, DLEs are very important. They can yield great benefit over time, but they're not being built as quickly as they are, for example, in industries where the product is not physically instantiated, software. So, you know, if you have a CRM product like Salesforce, or you have a piece of software that helps people order supplies in a hospital more efficiently, or just helps them order supplies, then a company with a DLE will come along if you don't build it yourself. And what that company will probably do is take an existing source of data about the inventory in that hospital or a whole bunch of sales leads, and very quickly build an opinionated algorithm that makes some sort of prediction which is very useful to the users of those products.
So, for example, if you sell CRM software and you don't have a DLE today, the vulnerability to your business is someone could come along, get a bunch of existing leads out of your CRM, out of the current system of record, so to speak, and add an algorithm to it, which helps you prioritize those leads more efficiently and more effectively. And then, as you target those leads and give it feedback that, you know, that was a good suggestion, that lead actually closed really quickly or not, then they're going to get a better prediction next time. And the point is, you, as the existing CRM, get subjugated to being a system of record, and the new entrant that has the predictive system on top of that becomes the one that gets most of the share of wallet there, because customers want to pay for effective leads. They don't actually want to pay for a database. They just have to pay for a database. That's just step one.
So, yeah, I think the impact on your business of not having a DLE going really depends on what industry you're in, but most businesses will probably be quite affected by this sometime in the next 25 years.
What about the, and again, it's hard to look at this in theory, but what about the implications for a company that is kind of weighing trying to build a DLE and the much more kind of obvious pressure that's both bubbling up from the bottom and also coming top-down in regards to implementing as much AI tooling as possible? Because I think there's often this kind of conflation between being an AI-first company and a company that just implements a lot of kind of contemporary tooling.
>> Yeah. How do you think about that challenging balance of wanting to improve process efficiency through tooling versus the more substantive change of putting a DLE in place?
>> Yeah, that's a really important distinction to make, because what I certainly see a lot of people do is buy all the tools before they build anything. It's like the typical handyman thing where you get an idea that you're going to make a table, and you spend many weekends in a row at the hardware store buying all the right tools, but you never actually make the table. Um, whereas actually, you could make the table with very simple tools if you really wanted to. A screwdriver, you know, a chisel, that's all you really need. And I think a lot of that happens in this era or these days with AI, as in, people spend a lot of money organizing their data, labeling their data, cleaning up their data, all that sort of stuff. Then they spend a lot of money going and purchasing a model from someone and then trying to push their data into that model, etc.
Whereas actually, what I see as being much more successful in most industries, you know, there's a little bit of an exception here, which is that if you do actually need to deal with a huge amount of unstructured text data, probably better using a GPT of some form to get started. But what I see being much more successful is when people go and actually just look at the data they have, think about what prediction they really need to make to offer a better, faster, cheaper service, and then try to just make that prediction manually using basic statistical methods, or doing something like running a random forest or a clustering algorithm to see if there are any patterns in the data that might be a clue to what might be predictive.
So, for example, we think that all of our customers tend to order with a 40-day gap between orders. And if we can predict that with some degree of accuracy, or a high degree of accuracy, we can sort of pre-ship them on the 39th day their next order, and they can send it back for free if they don't want it. But most of them won't. And so if you're able to make that prediction, you could make a lot more money and have like much higher retention. But the way to do that is not to go and buy some e-commerce tool that has some AI thing on it that they're marketing. It's probably to just look at your own data first and see if there's a cluster of orders in various ways, and then run that prediction manually, actually send those orders, and then see what your feedback is. And then if it's working, sure, then go and implement it as a system that runs by itself. But don't buy all the tooling first. Just do some pretty basic statistical analysis first.
>> Obviously, easier said than in a podcast studio than, uh, than inside the, uh, than inside the boardroom. I'm interested again, just let's go back and think about what you talked about in terms of getting a DLE in place, like a number of practical steps that a founder, if a founder is listening to this and going, wow, that that's something that is a priority for our business, which it should be.
>> What do they need to do in simple terms?
>> First thing people need to do is be honest, which is try to sit down with a very small piece of paper, like post-it note size paper, and draw out what either is or could be or might be the DLE that's existing or could exist in their business. And that is just write out, what do I think is my most unique data that I have? What information do I think that gives me about like customer behavior or something like that? And how do I think I could make a prediction that could be useful to everyone that I can actually get feedback on? It's sort of no good in the context of building a DLE to make a prediction and then not get any feedback on it, because then it won't improve next time. I think most people just need to sit down with a piece of paper and a pen and try to draw that loop. And we can draw that loop for things like Uber with route optimization. We can draw that loop for DoorDash. We can draw that loop for a lot of these companies like quite easily. And I think people get a little bit carried away. At least what I see is they get a little bit carried away with, as you said, buying tools or thinking about something else around AI and not about the core prediction they need to make in order to make more money or deliver a better service and then make more money, or just actually make more money by automating something.
>> What about the, the kind of fallacy of boiling the ocean, right? Just having this view of, well, five, six, seven years ago, you heard it a lot in investment pitches, people being like, oh, we, we're just going to have an absolute massive data. We're going to have a real data advantage. And this kind of fallacy that the competitive advantage comes in having the data as opposed to, as you said, like having a scientific method about understanding what are the specific features that are going to be predictive within that data. How should somebody approach it when they're, again, when they're starting with a large volume of data but trying to get to a predictive outcome? What's that process?
>> Yeah, it goes back to basics. What's your goal? Like, what do you actually need to predict to provide a better service? It depends what you're doing. Like, if you're growing food, you need to predict, what are the features that you need to sort of think about it? What is predictive of having a good crop? It's basic things: water, fertilizer, and other environmental conditions, how often you prune, all that sort of stuff. So, put all of those things into, run a regression on all of those factors and see what is most predictive of having a good crop. It depends what industry you're in, but just again, taking it back to basics and thinking, what is my goal here? What is the overarching business goal? It's some version of, I want to make more with less. So you think about on the revenue side, I want to make more by better predicting orders, by better segmenting customers, by improving the frequency at which we talk to customers so that they buy more stuff, or whatever it is, or making a product that is more useful to them because it provides, you know, in a second-order way, useful predictions to those customers. Or where can I save a lot of costs? Where can I automate stuff? And a lot of automation is machine learning at its core. So again, I think it just starts with making sure you know what your company's goals are, making sure you know the economics of your own business, and then playing with a bunch of predictions around the economics of your business to figure out what, where you can get some leverage from machine learning.
With the observation that, you know, in many businesses, like in the Uber, DoorDash examples that you gave, that there can be multiple DLEs set up and running at a given time that all equate to building competitive advantage in those different segments of the business, right? It's not just one thing that sits in a business.
>> Exactly. And they can all be complimentary. So, for example, Uber does a little bit of machine learning around pricing and surge pricing and when people will accept a surge price, when they won't, and what not. And if that works out well, you people are encouraged to use Uber more because it's actually really cheap when they want it to be cheap, and it's cheaper than alternatives when other alternatives are expensive. For example, in surge moments. If they get that right, they have more users. What happens when they get more users? They get more data on more rides. When they get more rides going on in more areas, they can figure out the most efficient routes in those areas by learning, okay, if you take this route, it takes this long. If you take this route, it takes that long. Then they can suggest better routes to drivers. Then what happens? People get there quicker. They're happier. And they use Uber again. So all of these things feed into the DLE by generating more data that they can then learn from to improve each of those algorithms later on.
>> If you haven't started your company yet, but you're thinking about starting today, like, what do you think are the implications of not just DLEs, but us coming to the kind of penultimate period of the AI-first century? What are the implications for founders who are building a company today?
>> Don't have to hire as many people. You know, I've always thought it's quite funny when people raise money, the first thing they do is hire people. It should have always really been the last thing you do. Spend it as slowly as possible. But now, certainly, it is one of the last things that you should probably do. The first thing you should do is hire an AI to try and do a job for you. A research task. Fill out a whole bunch of forms that you need to fill out to enter a market. Do a bunch of research on what market you should enter. Run a pricing experiment. Give you ideas about how you can market different products. Write your marketing emails. Run your CRM. You know, AIs can do all of these things now, particularly large language models. They can do all of these things.
So, I think one of the major implications for founders is you can get more done with less. And this has been sort of an inexorable trend over the last, however many, call it 25 years in technology. It's become cheaper and cheaper to start a company if you've been smart about adopting the latest technology. And I think that's certainly an implication for founders today. Another is, because it's more competitive, because it's easier to start a company, therefore more people start companies, it's more competitive. We have to have a stronger source of competitive advantage. And so you have to really think about whether how you're going to generate a DLE. Anyone can grab ChatGPT and then release some sort of wrapper, so to speak, on top of that, some product that is a light layer on top of ChatGPT to, for example, generate marketing emails. Like, that's not a product that's going to have a sustainable source of competitive advantage. However, a product that has its own unique data source and its own process and its own models might be able to produce far better marketing emails than something that's just built on ChatGPT. Something that's, for example, trained on a whole bunch of world-class marketing emails and has been, you know, developed by a really good copywriter and whatnot. It's a silly example, but I think the barriers to entry, the barriers to starting are lower. I think the cost of starting is lower, but competition is higher. So do less of the boring company-building stuff, sales and marketing stuff. Let AIs do that, and more of the machine learning part.
>> Just to kind of pick at that a little bit, right? If, if I am vastly enhanced by the use of a whole bunch of agents to do a bunch of these manual tasks, surely two of me is going to still go twice as fast because we're using a lot more agents. How, how do you think about that balance?
>> Um, well, it just depends on what stage you're at with your company. Like, you might just not need that many yet because you don't have that many customers to work with, or you haven't really defined your market, or you're still operating in a niche. So, having 500 salespeople or 500 sales agents that are AIs might not be necessary if you're in a market that only has 500 customers, um, to start with. So, I think, yeah, sure, later on, the obvious argument is that you expand in a similar way. You just have a lot more leverage for each incremental person. I think the stage of the company really determines that when you need those extra resources.
>> Do you think that having agents at our disposal, tools like Lovable, Wind Surf, Cursor, etc., changes the profile of top founders in the future? I mean, I know that historically you have really focused on leading researchers, technologists, and obviously they have been the people who built some of the largest companies. Does that change?
>> Yeah, I think for what I tend to do, no. In that I'm still looking to work at the lowest levels at the frontier of intelligence broadly, of systems of intelligence. And so, no, because, you know, the sort of people I work with developing those underlying systems. I think it certainly changes the profile of a SAS company founder, or any sort of business founder, really, is in they have to be very, very fluid and competent in using agents in their business. They have to basically be as good at managing agents as they are at managing people.
>> I think this brings us to probably our final stop, which is where do we go from here to get to the end of the AI-first century? What's the next wave of evolution, do you think?
>> Yeah, I think it is when we've evolved to the point where we have computers built into us. And like the sort of fairly benign sci-fi at this point sci-fi version of this is the cyborg. But obviously, it's going to look very different to how we have been imagining, or just like it always has, it looks very different to how we've been imagining. You know, I don't think we would have imagined AI at this point in 2025 to be something that we chat to like in the chatbot. We probably imagined it to be a little bit different, probably a robot. And turns out robots are harder, and dealing with text is easier. And if you had have been there at the beginning of the AI-first century, you would have realized this because it started with text. So, it's going to finish with text.
But what, what's it going to look like at the very end of the AI-first century? I think it's going to be the start of better integration with our real life, with our real world. And that's probably the next century, which is sort of 2050, but maybe it's going to be a little earlier, like 2030 onwards, which is not quite AI, but something very similar to AI helping us in our daily life and fully integrated within us. So, you know, the areas I'm thinking about as being particularly exciting are areas like brain-computer interfaces, or any form of neuromodulation. That is the addition of some sort of electrical device to the nervous system to help us move through the world a little bit better. Quite literally, being able to move through the world, walk better, etc., even if you've been the victim of a stroke or some sort of paralyzing disease, but even if you're not, just to be able to walk better, faster, because your nervous system has been modulated or is being modulated by a computer. All the way through to how you think, being able to control five computers at once, being able to control a computer without even thinking about controlling a computer because it can read your background thoughts. You know, I think that's going to be probably the next step.
And then, you know, tied in with that, or soon after that, we'll have a lot of little AIs that can actually move and operate in the real world, like go and do something for you. You know, the simplest version of this is like a simple helper robot, but I think there'll be other versions of this. You know, thing, you know, people have talked about nanobots for a while that sit in your food and then do some of your digesting for you, or just things that can go out and do things for you in the real world, go and pick things for you. And everyone can perhaps be a self-sufficient farmer with the assistance of a whole bunch of robots that can actually interact in the real world with some degree of reliability, which we don't have today. So I think that's probably what's next is, um, just more integration and more impact on your, your real life as it's lived in the physical world.
>> We'll take that as an optimistic note on which to finish. Ash Fontana, thanks very much for your time. Thanks for listening to today. As a reminder, if you enjoy this podcast, please subscribe. Reminder that today we talked about the history of technology in humanity and the AI-first century. We talked about data learning effects and how they influence contemporary company building, both for existing companies and for founders who are embarking on their journey, and where we're going in this final 25 years or perhaps fewer of the AI-first century. Thanks for joining us.
[Music]