Pod Insights a16z Podcast 10 Year Futures (vs. What's Happening Now) About this episode Andreessen Horowitz' Benedict Evans reflects on mobile, Google / Apple / Facebook / Amazon, innovation, machine learning, computer vision, autonomous cars, mixed reality and crypto-currencies - and more, in this keynote delivered at the most recent annual a16z Summit (November 2017 in Las Vegas).
The personal computer S-curve gave us 1.5 billion PCs and 3 billion people online, and now the smartphone S-curve is giving a computer to everyone on earth.
So today, we have two innovation conversations: First, what can we build on the billion-scale platforms we have, what new models can we create, and do the internet giants leave any room? And second, what are the next S-curves -- the next transformative technology waves?
Transcript 0:00 Speaker 0 There's a few things I wanted to talk about today, and I think a good place to start is thinking about the S curve that's dominated, really created the tech industry in the last couple of decades. That's the growth of the PC, which went from nothing to about one and a half billion units installed around the world in the last couple of decades. And of course, what's happened to that as an S curve is it's been replaced by another one, which is the growth of smartphones. And smartphones in turn are coming out of a third S curve, which is the growth of mobile phones. And mobile phones today are close to connecting pretty much everybody on Earth. So there's three billion smartphones, five billion mobile phones, maybe five and a half billion adults. And what you see with all of these curves is that they start slowly, and then they reach a point that they explode, and then they kind of slow down again and get a bit boring. And to systematize that, then what we have are S curves following S curves.
0:48 Speaker 0 So most recently we had the PC internet, which has been followed by the mobile internet. And as we get to the end of all those kind of the the stable and mature part of each of these curves, we start talking about what we can build on top of them. And so talking about mobile now often feels like talking about mobile in 2004, 5, 6, we talk about what you can build on top of the platform. So on PCs, we talked about search and social. On smartphones, we talk about ride sharing and Instagram, Instagram and Instacart and deliveries. We talk about what you can build on top of the platform rather than the platform itself. And so when you get to that phase, of course, people start saying, well, innovation is dead and this is really boring, and the new iPhone looks just like the last iPhone, and what have you done for me lately? And of course, it always feels like that. So back in 1994, a young student called Mark Andreessen came out here and said he's already missed the whole thing, because everything had already happened.
1:37 Speaker 0 And of course, that wasn't quite right. Rather, what happens is that there's always something else coming. And so what that means today is we're having two sets of innovation conversations. We're talking about what we can build on the billion scale platforms that we have already, and we're talking about what the next S curves might be, what the next platforms might be that come along and reset everything. Now, as we talk about the platforms today, there's a massive amount of experimentation actually still going on, whether that's bots or voice or watches or wearables, whether that's tech companies starting to outdo um traditional content companies at their own game, continuing growth and creation of new social companies or the growth of AR and VR and discussions about what those might mean. There's also a sort of a meta conversation around all of this. Is there any white space left? Is this all going to be boring? And can you even compete with Google, Apple, Facebook, Amazon?
2:25 Speaker 0 Is that even a thing you can do? Now, big scary tech giants are not actually something fundamentally new. This has been around for quite a long time. However, there is perhaps something new about these kinds of companies. So here you see the annual revenue of the last three generations of the companies that dominated tech, so IBM and then Wintel, and now Gaffa, Google, Apple, Facebook, Amazon. And there is clearly kind of a different scale here. I think actually the best place to see that change in scale though is in is in employees. So Microsoft had 45,000 people at the end in 2000. Now you can see Apple, Facebook, and Google all having grossive employees to make them almost bigger companies than Microsoft itself was. Of course, if we zoom out a little bit, um, we have to zoom out a little bit, then we can add Amazon, which is adding people at a different kind of scale, not just engineers and salespeople, also retail and logistics.
3:10 Speaker 0 And if I take this chart forward from 2016 to 2017, we can see there's kind of a something quite dramatic going on here. Amazon will probably end this year with about 600,000 people. You can also look at this in another field in advertising, which is kind of a topical content. Internet took about half of the print industry's advertising revenue and growing, but Google and Facebook took half of that. So creating an ad funded content business today is a pretty challenging proposition. And so there's a sort of a thesis that emerges here in the valley that you have sort of super evolved organisms in Google, Apple, Facebook, Amazon. So you have the massively greater scale. You also have the fact that everyone read the book. Um they all saw Microsoft um fail, they saw Yahoo and MySpace fail, they read the disruption book, and if everyone's read the disruption book, then disruption maybe doesn't work. And so they've all sort of seen what seen what happens, and so they're much more aware of how they might be overtaken.
3:59 Speaker 0 Then they're all except for Apple controlled by founders, and that's really crucial because it's being controlled by founders that means that um Mark Zuckerberg could decide to spend 10% of the company buying WhatsApp and nobody could tell him not to. And then, of course, there's four of them, not one. And that's really perhaps the biggest change that we have four giants competing on equal terms and trying to elbow each other out of the way. On the other hand, the companies that win always look invulnerable until they don't. So here you see Microsoft's total dominance of the computing industry, right up until the iPhone came along and made it pretty much irrelevant to the computing platform side of the industry. And then companies that we talk about in the tech industry as having won or being dominant haven't quite won yet. So Amazon might look like a big company, but in the scale of overall retail, it's got still got quite a long way to go.
4:45 Speaker 0 And so I think the thing to talk about here is that the winners always look invulnerable until they don't. So IBM looked invulnerable, then Microsoft, Intel, Nokia did, then the internet giants of the um the last bubble, now Google, Apple, Facebook, Amazon look invulnerable, but all the previous ones fell, and it was not predictable at the time how or why they would fall, but they all did. And so something will always come along afterwards, which is really what we're trying to work out now. And so the question becomes well, what are the new S curves that will come along and change all of this? Um as all previous generations got changed. And the thing to think about S-curves is that they're never actually smooth. So here you have iPhone unit sales, and you know the first three data points here are proving Steve Ballmer right when he said it wouldn't sell anything at $700. Then they change a business model, they find product market fit, and then they work out the right product.
5:33 Speaker 0 You see the same thing for the iPod, which took almost five years to work, an overnight success that took five years. And the thing that's happening here is you sort of have three phases at the beginning of the S-curve. You have the period where the tech doesn't quite work at all, but you think it will. Then you have the period where you've got it sort of working, but you don't have product market fit and you're still iterating on the product, and then there's the point that it explodes and you're pouring on rocket fuel. And so we have, I think, four S-curves that we're thinking about today that are the various stages along that process. First, we have autonomy, which doesn't work yet. Um, anyone in the field will tell you that. We have mixed reality, which is just reaching the point that the technology actually is starting to work and it's about to become a real product. We have cryptocurrencies, which are in a slightly different situation in that the the technology works, but we're working out what the market is, and then there's AI, where we are definitely in the rocket fuel for rocket fuel phase.
6:26 Speaker 0 And so I think AI is probably the best thing to talk about first. Now, we probably all heard of AI. Everybody has had the conversation with their boss when they get off the plane and said, I heard about this AI thing in a piece in Business Week, and it seems a big deal. What's our AI strategy? But it's kind of there's a bunch of unhelpful ways to talk about this. Um you say artificial intelligence, it feels like someone's put the black model list from the beginning of 2001 into the room, and then we've all turned into monkeys and we're dancing around and screaming at it, and we don't actually know what it is. And so there's unhelpful ways to talk about AI. Saying it's oil, saying Google has the data, saying China will have all the AI. In fact, I think calling it AI itself is unhelpful. Talking about artificial intelligence is sort of unhelpful. It's more useful first of all to say machine learning, which is the primary technology we're we're interested in, to talk about automation and I think to talk about enabling layers.
7:16 Speaker 0 And so when I think about enabling layers, I think a great analogy here is relational databases. Um, some people in the room may actually remember what these were first time around. And so relational databases emerge in the late 70s and early 80s, and they allow you to say, show me all customers in the zip code for the first time. Before that, if you wanted to ask a question like that, that was an engineering project, and somebody would have to go and spend a couple of weeks building it. Databases were not built with that kind of structure in mind. And the fact that relational databases meant you could ask those questions, meant that databases went from being record keeping systems that just replaced paper filing cabinets to being business intelligence systems. And that enabled all sorts of things. Most of all, of course, it enabled billion-dollar new companies. So we got Oracle. Then we got SAP, and SAP is not just a database, SAP is an ERP system, and SAP means that you can move your supply chain to China.
8:05 Speaker 0 And so SAP changes what companies are on how you can run a company. And then by the 1990s, pretty much every enterprise software company is a relational database, but nobody looks at COM or Salesforce or Success Factors and says, well, that won't work because Oracle has all the database. Rather, what happened is this is an enabling layer that enables new things but enables them in all sorts of different companies, and you get new billion dollar companies, but you also get lots of companies that no one actually looks at and says, Well, that's a database. Nobody looks at Starbucks and says this is a database company, but they're all enabled by databases. That I think is the best way to think about machine learning. As we ask, well, well, what is it the innate that it enables? The kind of the best, if perhaps the crudest way to think about this is to think about patterns. So you can take a data set and machine learning will let you find patterns.
8:51 Speaker 0 And those patterns could be all sorts of things. They could be cats and dogs, they could be um different delivery routes, they could be fraudulent behavior on the network. In this case, this is the graphic from Instacart, so these are different grocery types. And so what patterns can you find? Well, you can find cat pictures, but the thing that that that really changes is that finding patterns becomes a generalizable solution. So is there a cat in this picture becomes the same kind of question as which customers are going to churn, or is that car going to let me merge, or is there something odd happening on our network, or is this customer um behaving fraudulently? That's to say you find lots of different questions, and you also find things in the data that you didn't know were there. You find the unknown unknowns, the thing you would know things you would never have looked for. And so as you think of all of those different kinds of questions, it's kind of useful to go back and think about automation.
9:39 Speaker 0 So with every wave of automation, we always imagine we're going to get the thing on the left. We imagine we're going to get a humanoid robot. And so in the 1950s, people imagined that we would have domestic robots that would walk around the kitchen with two arms and two legs, and they would pick up your laundry and your dishes and clean things. And of course, we didn't get a domestic robot. What we got was a washing machine, and a washing machine is a Robot, but it can't do anything other than wash clothes. I mean, you could put dishes in them and they would get clean, but it wouldn't be really the result you're looking for. And so the point in each of these cases is that you get a single, you don't get a general purpose solution. You don't get something that can ask every question. You get something that can do one thing. And that's kind of a point about every kind of automation. That people look at automation and imagine it's intelligent.
10:25 Speaker 0 But of course, Charlie Chaplin falls into the machine, it doesn't know it's Charlie Chaplin, just carries on doing the thing that the machine has been built to do. And so in the same way, you don't just get washing machines, you also get dishwashers. So data is super extremely specific to each vertical, extremely specific to each problem. And then in fact, the solution you build with that data gets very specific depending on the problem and the solution that you're looking for. So we have washing machines and dishwashers and microwaves. We don't get humanoid robots, we don't get Hal 9000. So what are the washing machines of machine learning? What might we build? I think there's sort of three buckets to go and look for stuff. The first is current analysis. So questions we're asking now of data that we have now, machine learning seems to produce better results. Then secondly, um machine learning actually asks you lets you ask different kinds of questions of that data.
11:12 Speaker 0 And then, of course, most interestingly, machine learning lets you address entirely new kinds of data. So it lets you look at stuff that computers just couldn't read before at all. So voice, images, and video, most obviously. And as we think about those things, it's worth going back to that mobile S-curve chart. Because here are the things we're talking about now, but when smartphones and mobile phones and mobile internet was launching 10, 15, 20 years ago, we had a whole other bunch of ideas as to what we were going to do with this. Mostly it was going to be email and stocks and weather. Although actually I've written this wrong, it was going to be M email and M stocks and M weather and M banking, because some for some reason everything had to have an M at the beginning of it. Before that, of course, it was everything had an I at the beginning of it. And the point here is that when you start looking at a new enabling technology, the first demos that you see and the first things that are built with it are just expressing the capability, but they're not everything that you're going to do with it.
12:05 Speaker 0 And so the same thing with machine learning now, the first demos we get a cat pictures and trivia questions, but you know, there will be an awful lot of other things that get built with that, because those are just demos. Those aren't actually what the technology is. So what is it that we can automate next? Well, if you think about the history of automation, we started out by automating maybe human beings as beasts of burden. So we mechanise legs here in the 19th century, and then we mechanize arms with production lines, we mechanise fingers with typewriters a little bit later on. We met we we mechanise arithmetic, we recognize mechanized mass with adding machines and early computers. Everybody you see here is basically in a cell in a spreadsheet, and the whole building is an Excel file. And so now we think, well, what is it that ex conceptually that we can automate with machine learning? Well, a really good way of thinking about this is to think about an intern.
12:52 Speaker 0 So right now the stuff computers can't do, can't do that any 10-year-old can do. And computers are moving along that. So computers will be able to answer the kinds of questions that a 10-year-old could answer. Maybe also the questions a 15-year-old could answer. Probably not any of the questions that people in this room could answer, although I can imagine some exceptions for some people in this room. But what you do is you don't get one 10-year-old or one 15-year-old or one intern. You get 100,000 or you get a million. And so you have a pile of data where you know there's patterns and you know there's that stuff in there, but you could never sort it with a computer, and you don't have a hundred thousand interns to go through that data. Well, now you do. And this is what automation always does. It gives you a massive multiplier effect. A steam engine doesn't replace one horse, it gives you a thousand horses. Machine learning doesn't replace one person, it gives you the capability to do what you could only have done if you had a thousand of those people before.
13:43 Speaker 0 Often very simple, often very mundane tasks, but things that just weren't possible before you had this kind of automation. Obvious sort of set of ways to think about that. And in fact, sometimes I think um if machine learning only did image recognition, that we would still be the biggest thing in the tech industry, because computers are going to be able to see. So for the last 50, 75 years, computers have been able to count and computers have been able to read in some senses, but they couldn't do anything with images or video. Um now they'll be able to do images and video in the same way they've been able to count. What might that be useful for? Well, what happens if you can show your phone a bag or a pair of shoes or a lamp or a meal and say, I like this, what is it? How do I make it? Where do I get it from? What else might I like if I like that? What if you can show your phone your living room and say what lamp should I buy? How do you automate recommendation?
14:30 Speaker 0 How much does that change e-commerce? What happens if you can automate trend analysis? What happens if you can say, show me everybody in Soho and everybody in Berlin and what are they wearing now and what were they wearing six months ago, and what patterns can we find? Not just what are they wearing, but what patterns and trends can you see if you look at millions and millions of photographs taken of what people are wearing? What does that do for trend analysis? What happens if you have a camera over every point of a production line? If you can see what's being done at every all 300 steps used to assemble an iPhone, how can you analyze that um industrial process? What happens if you can show a plant to your phone and it can tell you what's wrong with it? Again, you're automating a process here. And then you sort of get further away from things that are obvious, like hat pictures and fashion, it's almost gets more interesting. What happens if all the cameras in a subway station can now count people automatically?
15:19 Speaker 0 So again, you don't need your thousand interns looking at the video and counting people. But this is computers thinking like computers, not computers doing what people do. What happens to traffic analysis if you can count whoever is in any single place in a subway station or a shopping mall or a store or a whole city? You turn the image sensor into a universal input for a computer. It's not just that computers see like people, it's that computers can see like computers, and that makes all sorts of things possible. Now, one of the things, among other things that that makes possible, is driving, because machine learning, both in general and for image recognition is the reason we're talking about autonomous driving. Now, one of the ways that people talk about this is people often say driverless car. And to me, saying a driverless car is like saying a horseless carriage. So you have this wonderful thing built by Daimler at the beginning of the last century where they've taken the horse off, but they haven't actually changed anything else.
16:09 Speaker 0 And of course, that isn't how cars evolved. It's also not how cars will evolve once you get rid of the steering wheel and indeed the internal combustion engine. So you can completely reimagine what the car might be or what it even means to say a car or a bus or truck if you don't have to have an internal combustion engine, if you don't have to have human drivers, and of course, in a fully autonomous world where there are no accidents, you can redesign the vehicle even further. And so you sort of expect totally different vehicle types and totally different ways of thinking about what a vehicle is. But that's just a car itself, you also need to think about the fact that computers are going to be driving these cars automatically, and computers don't drive like people drive. And so you will have a very different way of thinking about what the road is and what lanes are, what signaling is. Um you can have platooning, you can have vehicles driving at 150 miles an hour two feet apart from each other.
16:58 Speaker 0 Obviously, in the Bay Area, people do that already, but you might not have as many accidents. And so you can change not just what cars are, but what roads are, and if you change what roads are, then you kind of change what cities are. So what happens to a city if you don't have any accidents? What happens if you don't have any parking, or if parking doesn't have to be in walking distance of where you need to be, it can be somewhere else. What happens to congestion and traffic signs? What happens to road pricing? You can have dynamic real-time road pricing if you like. And so that changes where you live and where you shop and where you eat and drink. There's an old saying that it was easy to predict mass cars car ownership, but hard to predict Walmart. That's kind of the change of scale uh scale of change that happened from horses. We'll expect to see a sort of a similar scale of change happening from from automating cars. Now, perhaps one of the sort of points to take as an aside from this is that machine learning is actually such an enormous S-curve that change entire cities or let computers see are actually just applications.
17:52 Speaker 0 So those are S-curves of their own that come out of the fact that machine learning gives us this step change in the kind of questions that computers can answer. Now, another thing that is enabled by machine learning, but is much more than that, I think is mixed reality. What happens if you can wear a computer that can see? If you have a little computer sitting on your ear that can see everything that you see. Now, when we look at machine learning, sorry, when we look at mixed reality, it's rather like looking at multi-touch in 2006. So this is in a demo that Jeff Hahn did at the TED conference in 2006, demoing a multi-touch screen, which no one had outside of academia had really seen before. And as you can see, it requires about three packing cases to put together. And he's got a little multi-touch lava lamp. What you can't hear is the soundtrack of this audience of technologists screaming and losing self-control when he says, I can make do this, and the image gets bigger, and everyone in the audience starts screaming because no one has ever seen this before.
18:44 Speaker 0 And so the reason I think this is an interesting comparison is that just as Murray Multitouch was in 2006, and that then became the interaction model for the whole world. Mixed reality is sort of in that kind of stage. It's in the stage of prototypes that work, not in the stage of the shipping commercial product, but pretty close. Give you a sense you can put things into the world. This is a concept shop from magically, one of our portfolio companies. So you can place things into the world. Well, well, what would you place? So there's sort of two ways to think about this, I think. One of them is you can just add stuff to the world. So you can make every wall a screen. You can put Minecraft on the table and move it around with your hands. You can stand in front of the stove and have the recipe hovering in front of you. I think what gets more interesting, as I said, is if the computer is actually looking at stuff, not just trying to find the flat surfaces.
19:30 Speaker 0 So I can stand in front of someone and say, who is this person? And um, when did I last meet them? And am I supposed to be talking to them or not? Or you can look at a product and say, is this cheaper on Amazon? Or you can say, you know, I left that thing that I was holding the other day on a table somewhere, where did I leave it? I saw a poster last week for a concert. Where is it? What's it going to happen? What's going to happen? 11 o'clock last night, I was talking to a bald guy with glasses from Hollywood. What movie studio did his badge say he worked for? And what did we talk about? And those are questions that become possible to answer. Um, and not in a kind of a science fiction way, but with basically all of the building blocks that we have today. Now If mixed reality is at a sort of a 2006 stage, then crypto or cryptocurrencies feel more like a 2000, sorry, a 1994 stage, which is sort of where HTML was HTML was then. And of course, that is, as I mentioned earlier, when Mark came to the valley and thought he'd missed everything.
20:20 Speaker 0 And as we think about where this technology is, there's sort of two ways to be at the beginning of the S curve. The first of those is that the technology doesn't quite work yet, it's not quite a product yet. And that's where autonomy is now, excuse me, it's also where mixed reality is now. And that's where smartphones were in 2006. The other way you can be at the beginning of the S curve, though, is that the technology works, but we're not quite sure what we should do with it. And that's where HTML and the web were in 1994. It all worked fine. You could install it on your PC and you could go and look at five web pages. But it wasn't quite clear what you were going to do with it and how useful it would be and what it would mean if everybody on earth was using this. And that's pretty much where cryptocurrencies are today. It works, but we're working out what to do with it. And so perhaps a way to get at this is to think again about automation.
21:06 Speaker 0 And so if this is what the internet automated, retrieval of information, or perhaps connecting people to each other and talking to each other. What is it that crypto automates? Well, it automates another kind of technology. It automates, well, money, but it automates trust. And so if you think about the kind of the process of this, money has been in the cloud in some ways for centuries. So we moved gold into banks, and banks in the 18th century were basically the cloud, except they had no backup. And then we moved the gold into paper, and then we moved the records of that paper into punch cards and into mainframes. And so we moved it into databases, but it was never actually software of itself. It was just a record of something. We changed it into a record, but then we didn't do anything with the record. And so the money went to the cloud, and then contracts and bonds and stocks and deeds and every kind of sort of written trust followed it, and it's all stored in databases, but we don't do anything with it.
21:55 Speaker 0 It's all just as inert as the gold was, or just as inert as the paper was. Now, the Spartans famously decided to use iron bars as their money because it was so impractical and so hard to carry around that no one would ever be able to get rich. And the advantage of gold over iron bars is it's kind of easier to carry around, but it's still inert. And then we use paper as a way of saying, well, I've got all this gold somewhere theoretically. And then we sort of forget about the gold part and say this paper is just money. But the paper is still paper, it doesn't do anything, it's not software, it's not um active. And so as we look at cryptocurrency, we start thinking about two parts of this. One of them is the distributed part, that you have a database where you can put things without needing a central authority. But the other is that that database can do stuff, and those records in that database can do things and mean things that were not really possible with any previous way of storing this kind of information.
22:44 Speaker 0 Now, I kind of that takes me to the final thing to think about. You know, we every kind of five years ago we hear that innovation is dead. We oscillate between innovation is dead and we're in a bubble and maybe both at the same time. But it's kind of useful to think about another way of thinking about S curves. And so what tends to happen if you live in a big city is you walk past a construction site every morning for six months, and there's a big hole in the ground, and nobody's actually doing anything. They're just kind of messing about moving cranes back and forth, and there's clearly nothing happening there at all. And then one Monday morning you walk past and they've thrown the whole frame up, and you think, oh, they've actually been doing some work over the weekend. And then they spend another six months messing around on that frame, not obviously doing anything very much, and then they suddenly they say, Oh, it's finished now, um, the skyscraper is done.
23:31 Speaker 0 And so the point here is you've got this kind of period when there's very visible activity that doesn't necessarily correlate with the period when actually stuff is being done or important things are being created. And if I go back then to those 4S curves, you can sort of situate this them on this continuum. Um, that for autonomy, we are still very much down in the muddy hole in the ground digging foundations. Mixed reality is just sort of moving out of that to the point that you can throw the frame up. For cryptocurrency, the frame has gone up. Um, we're just trying to work out what the facade will look like and who the tenants will be. And for AI, the tower is up, and we're filling it up with tenants, and we're working out well, what entirely new businesses can we use, um, can we create with this? Thank you.
Previous The Future of Money: Banking on Fintech Next a16z Podcast: On Data and Data Scientists in the Age of AI Page updated August 25, 2026 · transcript and summary are machine-generated