The Future of AI

Welcome to the keynote from Cornell. I’m your host, Lutz Finger. I’m a faculty member at the Johnson School where I teach several courses on AI, including one actually where I replace myself with an AI virtual bot. I have as well a certificate program which is open to the public.

It’s called Designing and Building AI Solutions. Or as Marc Andreessen would call it, it’s time to build. That’s the main focus of the course. Well, beyond academia, I’m a startup founder and a CEO from a GenAI platform for e-commerce where we are transforming how shoppers engage with websites, search filters and products.

And I went there because AI and GenAI is actually transforming many industries. And it will still transform many industries, but sometimes it’s a little bit hard to figure out what it’s going to be like. And I think that’s the main focus of the course. What is hype?

What is value? And therefore, it is a special honor for me today to welcome Martin Cassado, with whom we will discuss what’s actually working. Martin doesn’t need really an introduction. He’s a well-known name in Silicon Valley and beyond.

He is a pioneer in software defined networking. He’s like me, a physicist. He is now a partner at a general partner at A69. He’s one of the best venture capital companies out there.

And he has backed some of the most exciting startups at the moment in this AI space. So let’s talk about what is working. Let’s cut through the hype and let’s see where we are heading. Martin, welcome.

Martin Cassado Happy to be here. Before we go into the what’s working, can you just give me a little bit of background? To our audience about yourself. You have done, you’re European, but you grew up in Montana and you ended up in Silicon Valley.

I’m a fake European. No, it’s true. I was actually, it’s funny. I’m, you know, because my, you know, my mom calls me Martin.

So my name is Martin. I was born in Spain. However, I actually grew up in a dirt road in Montana. So, you know, like actually all my aesthetic is like Nintendo generation nerd country boy.

But yeah, so they could listen. The quick sketch is, you know, I actually grew up in rural towns and, you know, in the West, in the US. I did my undergrad. I’m actually a failed physicist.

So in physics, you can be really good at math. You can be really good at computers. I tend to be better at computers than math. And so I ended up going into computation.

That actually went to, you know, my first job was at a national lab, computational physics. I was actually at the national lab when 9-11 happened. Wow. And so I was actually at the time I was in a wagon.

I was in a weapons program, you know, doing simulations. And the kind of the tenor of the nation changed. Right. At the time, it was like it was kind of almost this whole go over from the Cold War.

But I had the clearances. So they moved me to intelligence and intelligence. What’s important, of course, is more cybersecurity networking systems. And so I didn’t know anything about it.

So I actually took a course at Stanford to learn networking. And, you know, I was just this random guy that just kind of showed up. But from there, I actually got into the Ph.D. program.

That’s how I got into systems. And funny enough, I was actually, you know, I was actually going to be a professor at Cornell. So when I when I got my Ph.D. Yeah.

So I had a, you know, I had a, you know, an offer at Cornell. But instead, I decided to do this startup at a terrible time, which is the bottom of the recession. You know, A16Z invested. Ben Horowitz is on my board.

And then we were kind of their first series eight, a billion dollar exit. We were kind of their first kind of big exit. I ran that business for quite a while. And then I joined as a general partner.

So I’ve kind of I’ve kind of done all the walks, you know, academic founder, investor. And then failed physicist. Failed physicist, successful founder and pretty successful GP at a venture fund. Now, why would you like maybe just give it a little bit thought process?

Because when you are creating companies, when you’re managing people, when you go from Z to Z, you’re not just going to be a business. Like you’re going to Matter or like when I’m an executive in a big company or like, you know, I’m going to go like be an actor, which I’ll be terrible. I know nothing about acting, but like, you know, you have these moments where you’re so stressed, you want to go do something else. But when the stress goes away and I’m like, you know, just like spending time, whatever, like I, what I’m interested in is, is computers.

I think they’re just amazing. I think that really is like this kind of meta language to solving all problems. And so then I view my career as almost moving up in, in zoom level, right. In my twenties.

So I’m 48 right now. So in my twenties, it was like, you know, my zoom level, the thing I was focused on was like an individual paper and the individual line of code, you know, I was a software programmer, you know, maybe an individual project. And then in my, you know, at 30, I literally walked from my dissertation to start the company. I started to view the world more like the abstraction of life, zoomed out a bit.

And I thought of that as products and markets and people, but for one area for like one market, one organization, one product portfolio. And that brought me to my forties. And then in my forties, I zoomed up one more level. And now the primitive that I’m working with is, is companies, right.

And you get to see it play out in the experiment many fold in many different ways. And you can learn a lot more by doing that. And so, I mean, it’s true passion actually for, I think, for innovation and for computer science and for startups, why that I just continue to do it. And I do think the different perspectives as you kind of evolve in your career, incredibly helpful for you to understand how all of this stuff comes together.

Amazing. Now let’s do that zooming out for a second, because I, if we are talking about what’s working and I, I met you a while back, well, I actually met you in a wedding, but like, like I met you for real. I remember that. Yeah.

No, I’m sorry. But I, the, the real moment of my life was when I was like, oh, I’m going to do this. And I was like, oh, I’m going to Matter Here’s the thing. AI has steadily progressed since the 60s.

We’ve solved all sorts of amazing things. In the 90s, we could beat Russians at chess. We’ve been very good at perception for a long time. We’ve had self-driving cars for a long time.

Listen, when I joined Stanford for my PhD, it was in 2003, and Sebastian Thrun had just won the DARPA Grand Challenge. I remember that because everybody was like, hooray, self-driving cars are here. It’s been around for a very long time, but then if you look at the industry, it’s invested $100 billion since then, and finally, we’re actually there. There’s been this very open question.

Actually, we should… Hold on. This is so… Finally, we are there.

In San Francisco, we have self-driving cars. In the rest of the San Francisco… No, 100%. Even then, the technology is finally there, not even a mass deployment.

I think that’s a great microcosm for the story of AI, which is basically, it’s been very, very capable, but we’ve never had AI-native businesses. We’ve never had AI-native startups that have really been successful in a way. Really? The reason is the economics have just not been there.

The economics have been terrible. They required tons of capital investments. There’s perverse economies of scale anywhere. It’s been a great technology, but it’s been a terrible business.

That’s it. That’s literally the story of the last 30 years. When you say AI-native business, and I have this discussion quite often about… Yeah.

If you look at LinkedIn, it’s a social network. Yes, it uses AI to figure out who to connect to, but it’s not an AI-native. Right? If you look at many other…

It has a lot of AI, but if you look at Netflix, it uses AI to suggest the movie, but in reality, it’s a channel for videos. Yeah. Yeah. How do you define AI-native business?

Right. What did you say when you say it hasn’t been a success? Yeah. What do you mean?

Right. So I’m talking about AI starting five years and before that. It’s historically not been a fertile ground for startups. And by AI-native, I mean the primary product offering or value is driven by AI, as opposed to…

I mean, listen, Netflix is set in DVD. You didn’t need AI. LinkedIn was a social network. You didn’t need AI.

So, of course, AI will give you some sort of a lift and some sort of stickiness, but that was not actually at the heart of the thing. heart of what was being built or what was being sold, nor was it at the heart of the people that you were hiring early on, et cetera. This includes Google, right? I mean, Google early on, this was a traditional indexed search.

It was not AI. And so again, just take it from the position of an investor. So an investor, you spend decades hearing about AI and seeing all of these great academic papers. And then startups come in and they show you these great examples, but you never have a wave like you had with mobile or like you had with cloud.

You never had a wave of these kind of AI native companies. And the paradox has always been like, well, why is this the case? And we actually did a lot of studies of this. And recently once we studied, we were like, okay, so why does this never take off?

And the answer is just the economics have never been there, right? What AI has historically done is if you have a business, it can improve it by 20%, which if you’re Google, that’s amazing. Or if you’re LinkedIn to your point or network, that’s amazing. But if you’re a…

Startup, a 20% improvement in something does not a company make. So the big change that happened this time, and probably the most important thing of our entire talk is the following, which is with the generative stuff, it actually has an economic dislocation that’s very specific, which is, well, there’s two of them and the definitions get muddled, but I’m going to simplify it and we can kind of explain it later, which is it brings the marginal cost of creation to zero. And it brings the marginal cost of language reasoning to zero. So if you think about compute, so compute brought the marginal cost of computing two numbers to zero, right?

So before we had compute, you literally… Like I said, the name computer came from the term we gave people that were calculating logarithms by hand. And then we created ENIAC, which is a computer that could do it 10,000 times faster. With the internet, we brought the marginal cost of distribution to zero, right?

And so you could send it in mail, it’d take two weeks, or you can send it immediately over the internet and the price per bit literally dropped to zero. And so with generative AI, the marginal cost of creation has gone to zero. And a lot of the many successful companies you see, this is what they provide. They provide the creation of image, they provide the creation of a story, they provide the creation…

And the economic dislocation is incredibly clear, which is like, let me just give you one example and then I’ll shut up and we’ll go forward. But let’s take me, okay? So if I wanted 10 years ago to get an avatar of me to put on my social profile, right? Let’s say I wanted to be a pixel…

Sorry, character. Martina’s a Pixar character. So to do that, I’d go to a designer and the designer would spend, let’s say, two hours at a hundred bucks an hour, so it costs 200 bucks. That’s how I would get this, or I would do it poorly and look terrible.

Today, I just use a model to do that and the inference cost. So the actual cost to generate that image is about one one-hundredth of a penny. Actually, these days it’s less. It’s about one one-thousandth of a penny, right?

So you’re talking about a five order magnitude difference in cost. That’s an economic… I mean, dislocation. These are the things that super cycles are made from.

Yep. Yep. And it’s actually a nice term, super cycle, and we should get back to this. But just as an example, we have…

So I created this online course, right? As I said earlier on. And in this online course, I replace myself with an AI. And the question was always, why would you do that?

And the reason is that I don’t have to fly or to travel to Ithaca to sit in a studio with a computer. I don’t have to fly to Ithaca to sit in a studio with a computer. And I have a nice studio staff here. Shout out to them.

And then I can do this actually from my home, right? So I save cost and I become more effective and can keep the course up to date. Now, that said, why should people… And this is a little bit like the hype discussion.

For me, it makes sense, right? For me, I replace myself. I don’t have to travel. Awesome.

But why should somebody actually create a virtual copy of themselves? For most people, this is like a little bit of weird thing or the uncanny valley. So I don’t know the answer to that because I don’t know that space very well. But what I can talk about is what people are doing.

So in the companies that are doing well, I think there’s a lot of use cases. I don’t want to be distributed. There’s a lot of use cases like the one that you just brought up that are just emerging because the technology is there. And I don’t really know how to reason about them yet just because they’re not in broad use.

But I think that’s a good question. I think that’s a good answer. I think that’s a good question. I think that’s a good answer.

But the companies that are working today, think about 11 Labs or Midjourney or Ideogram. There’s these existing companies today. You can ask a question like, okay, why are people using those companies? What do the products do?

And then in most cases, what they’re doing is they’re taking some content for which there’s market demand for, and they’re basically driving the time and the cost of creating it to zero, whether that’s music, that’s an image, that’s speech. Joe Consorti 08.00 And you need these both in like consumer use cases, like it’s entertainment. I like to listen to music. I liked it, whatever.

Or they are enterprise use cases. Like in the case of speech is a very important one. Like if I’m going to subtitle a movie, I need something to do that. So I can either go through kind of like this laborious process of having actors do it, or I can have a model do it or a model work with an actor, which is very common.

Joe Consorti 08.00 And so again, the reason that this AI wave is so important and successful is you can very clearly articulate like, what is the value? What is it relative to an existing market need? Code is another great one. We clearly need developers.

Developers are a massive market, arguably a trillion dollar market. You know, if you think about like 100k for like, you know, 50 million developers, if you can speed up their productivity by 20%, you’re talking about $200 billion. And that’s just code. And we know there’s so many verticals that are working.

Joe Consorti 08.00 And so you can have a lot of people that are working. There’s all the diffusion models, like I mentioned, which is image and video and music. There’s code, but then all the language stuff, which I haven’t even talked about, which is, you know, think opening eye, chat, TPT, which is like search, retrieval, you know, novel writing, whatever. And so it just, this is whole monopoly of use cases where the economic equation is just totally, you know, disrupted.

Joe Consorti 08.00 This is pretty amazing because I think you talked about the super cycle saying, okay, there is something new. Joe Consorti 08.00 This time, it actually can be not an add on like it was for Netflix or LinkedIn or whatsoever. This time, it can be actually an own business value. And then you talked about coding.

You guys invested into a company called Cursor. If you explain, this is pretty amazing, explain why these guys are so cool and what this means. Joe Consorti 08.00 So Cursor is the most popular development IDE. And what an IDE is, it’s an integrated development environment.

So any, any software developer, the primary tool they use is an IDE, right? And so that’s kind of where they like write the code and it does things. So Cursor is the most popular AI version of that. So what it does is one way to think about it is if I’m a developer and I’m using the IDE, it will use AI as almost like this, you know, companion or sidekick or coworker that you can chat with about your code base.

Joe Consorti 08.00 That will help make edits. Joe Consorti 08.00 To your code base that will make suggestions. And it’s just seen massive, massive adoption. It may be one of the fastest growing companies in the history of the industry.

I mean, so publicly, it went from zero to 100 million in 14 months. I mean, we just, we just never see that in this. And so it’s just a very, very exciting company, but it also shows that in these super cycles, and we also saw this with the internet again, like if you have a disruptive technology, they take off very quickly just because you have an existing market and existing need that’s being filled by, you know, Joe Consorti 08.00 Joe Consorti 08.00 You know, Joe Consorti 08.00 You know, a real technology innovation. Joe Consorti 08.00 Yes.

Awesome. Let me do a quick public service announcement here. If you are out there and listening to us and you kind of think like, what’s about this? What’s about that?

Ask us. You have in your live stream the ability to ask us questions. I see that I got already a few. So submit them in the chat.

I will see them live and I will bring them in as we talk and as I see fit. Joe Consorti 08.00 And then we’ll see if we can get them to the next round. Joe Consorti 08.00 And we can if if we don’t get to everybody, which we most likely won’t, then we can answer them afterwards. So this is your chance to talk to Martin and me kind of live in an asynchronous way.

Now, let’s come back to the topic because what you just said is you pointed to something that you said there is an existing need. Very and I use, by the way, I for my own course, I use those no coding tools because, Joe Consorti 08.00 What I figured out is my course gets typical MBAs or business people and I’m trying to help them build to build stuff. So I give them background of technology and I’ve given them all those tools in this certificate program. So we developed our own copilot and like course I like style only a little bit more simpler for people who don’t really understand code and you see how they write code on the side.

Pretty, pretty nice. Joe Consorti 08.00 But you, Joe, you’re not going to be able to do that. So you’re going to have to go back and do it again. Joe Consorti 08.00 But you, Joe, you’re not going to be able to do that.

So you, you pointed to one thing here. You said, Hey, it has to help with a certain need. And because the need was there and suddenly the technology comes out. Can you, you as VC, you see many people who probably come up with a hammer looking for a nail where the fundamentals are not there.

Can you talk through this a little? Joe Consorti 08.00 Yeah. So, um, yeah, this is also a very complicated topic. So, so either.

Joe Consorti 08.00 That’s the reason why I’m asking. Joe Consorti 08.00 No, it’s a great, it is a very important topic. And there’s so little that’s actually like written about it that’s sensible. I’m just trying to figure out like, what’s the right way for me to like, so, um, many of the most successful high growth companies you’ve seen are actually in existing markets.

Like Google was third generation search famously, like Facebook was third generation social networking, actually cursor. When it showed up, there was a previous product from Microsoft called Copilot. Joe Consorti 08.00 That you probably heard of, you know? And so, you know, there’s always this kind of misunderstanding of markets that if that, you know, all these companies that showed up and were successful were like this new breakout innovation.

That’s almost never the case in almost every case, like something else has matured the market. The user behavior was there, the buying behavior was there. And then you just managed to like hit the right timing or the right product form factor, whatever, to take advantage of it. Joe Consorti 08.00 And that’s, that literally is the history of the tech industry, independent of what people say on category creation, et cetera.

Like that’s just the reality. And we see it all the time, right? There’s these great, and it’s not even like a hammer searching for a nail. It’s like, you know, the market exists or it doesn’t.

And if it doesn’t exist, you have to create the market, right? Joe Consorti 08.00 And so there’s an entirely separate motion, which is, you know, some people call it category creation. I call it market annealing, whatever it is where you do have something innovative. Joe Consorti 08.00 People may not even have a problem yet, but you want to go solve that problem that they may not have.

Right. And this is actually the journey that I did with my company, which is like, we created this technology that nobody knew they needed or wanted. And I see this all the time. And that doesn’t mean you can’t build a great company.

It’s just like, it’s much harder. You’re chewing glass. Joe Consorti 08.00 You know, you’re going to people and saying, no, I promise you have a problem, even though you don’t know you have a problem. And here’s the solution, even though you don’t even think you have a problem to begin with, it takes a little bit of time.

Joe Consorti 08.00 It takes much longer. It’s much harder. You build companies very differently if you’re going after category creation. And so the way that if I were, you know, listen, if our listeners here and thinking about doing startups or investing in startups, et cetera, I would, I would split the world in two ways.

There’s there’s existing markets with existing user behavior with existing budgets and companies that go over those are going to have more competitive pressure. Execution is going to matter more, but they will grow much faster. And there, if you get timing, right, you get the good goes. There’s the other one, which is you really have this crazy innovation, but the market doesn’t exist and you can get iconic companies that way, but it’s just hard.

It’s just hard. And it’s a very different founder. It’s a very different journey. It’s like, and like, this is, this is amazing.

If you, you take those two approaches, you say existing problem, AI makes it better. And because AI makes it better, you have actually a chance of value creation. And then we have AI and we can do a value creation that we could have not done before. Now, there’s one question from one of, from the audience, which like Thomas Ward actually asked, how do you see the difference between single use case and multiple use case?

So single use case, like a mid journey, focusing on one area while open AI just goes broad and trying to address many different topics. Is that, is like both of them actually fall into the space of something completely new? Um, well, this is a very interesting question. It’s actually a great question.

Very interesting question, because if you actually look at open AI, to what extent have they actually gone broad? Um, like they were, actually it’s very interesting that you brought up mid journey. So, so open AI was basically the first with text to image, right? With DALI, you know?

And so like, but where is it with DALI? Nowhere, right? And so mid journey took it. Um, open AI was clearly the first with code, right?

You know, so with like Copilot, whatever, but like Anthropic seems to be the leader in code. I mean, uh, open AI was the best, I would say, production quality with video with Sora and yet, you know, like that market has kind of gone to others. And so, you know, the story of startups independent of this is choose a very large single market and go win that because, you know, a, it’s just too hard to walk and chew gum, but there’s also this kind of natural law of physics and company, which is if you have something that’s working and it’s a large market, why would you ever, not prioritize the next line of code or the next sales motion into that? And so I think that the story right now about, of AI in general, this gender of AI says these are incredibly large markets.

Like they’re, the market is growing incredibly fast. So we’re just seeing more fragmentation, right? It’s kind of interesting. Like, I mean, like it’s almost like every company is kind of finding its niche and this only happens when the market is, is expanding, right?

That’s when it fragments. And so I would say right now it, you know, focus is important. And even if you don’t do it, the market’s naturally going to do it for you. And it, you know, and, all of these things are sufficiently horizontal that there’s kind of a lot going on, but then when the market slows down, you will see consolidation.

And then you do have to start thinking about product portfolios and multiple areas and all of that stuff. It’s actually like fascinating, right? Ben, Ben Evans, the, the analyst, like he at one point in time worked for, for Andreessen Horowitz, actually. I don’t see, Benedict is one of the, one of the, smartest people I’ve ever worked with.

He is fun. So if, listen, if it’s a shout out to him, you know, we have no association. I mean, you know, he kind of moved on and he did a phenomenal job at Andreessen Horowitz, but if you can read his newsletter, I mean, he is one of the smartest analysts on the face of the planet. He is phenomenal.

He is phenomenal. And he is really like his newsletter is easy to read. I recommend him as well to all my students. And he lately said, um, OpenAI seems like a premature product.

It actually has not a good interface. And you alluded to this between OpenAI can code for me, right? But I need the IDAE integration in order to make cursor to make this a useful product. Can you maybe talk about the feature?

Well, there’s another, well, there’s another, well, there’s another way to view that, which is what OpenAI is good at is actually natural language chat and the products it’s very strong at are natural language chat. And the use cases of natural language chat, like retrieval question and answer. And I think that’s what’s really good about OpenAI is that it’s not just about the search and research, right? That’s kind of what was really working.

And that’s not that good. He showed that research is actually not their strong part in his latest post, but yes. But yeah, but no, but I mean, but these are the ones that people are using and, you know, I use deep research a lot and I find value out of it. Um, and so you’re just saying that that’s actually what they’re good at when it comes to code.

Anthropic just did better. And so like, and that’s become their focus and they’re very good at code. And there’s a lot of, I think, I think there’s value at the model level, but there’s also like, in order to, bring these more broadly, there is value at the app level. And this has always been the case with infrastructure, right?

Like clearly there’s value in computers, clearly there’s value in operating systems and clearly there’s value in apps. And we’re seeing the same thing play out here. Absolutely. Yes.

Perfect. Now, if you, if you like, let’s, move on. There is another heart from, from, if we move on from Coursera, there’s another height, which is going on. And like, you, know, Sam Altman says we know that like how to build AGI and I publicly said several times, I don’t think so.

But the whole point here is that we do this iterative loop. We have an AI and that spits out the most common next thing. Life is like the box of chocolates. So yes.

And then we critique ourselves. Maybe there are other ways what life could be like. And then we kind of make a list and then we follow through. This iterative loop became like something which we call them deep research or research or whatsoever.

And it became the basis for agents. Now we like Anthropic did a lot of co-pilot or like co-pilot as the brand name came from Microsoft. And now we do this critiquing and following through and this becomes agents. Can you talk a little bit about is this a different category?

Is this just an improvement of something? In which of your world does this fall? Yeah, so I think agent is one of the most diluted terms almost to the point of being useless. But there’s actually three working definitions.

So is AGI by now. Yeah, well, yeah, that’s misdefined. Yeah. So, I mean, just to give everybody listening kind of context.

So my team, so I run the infrastructure fund at the billion dollar fund. I have a team of let’s call it 12 people. Very kind of, you know, like, you know, many PhDs, many founders, like real smart team. My team sees about 3000 deals a year or something like that, of which we spend pretty serious time with probably 500, you know, of which we do really, deep work on about 50.

Like, I mean, it’s a very real funnel. So we just kind of have this kind of broad view. And I think the most I only say all of that to say, like, the most confusing things right now is this term agents because it doesn’t really mean the same thing to the same people. Like in the same meeting, you’ll see two people talking and they have two different ideas in their head about what agent means.

And so they can’t communicate. So let me let me just give you the three definitions that I see most commonly. Definition number one is an actual like agent representation that someone talks to. Like I’m a virtual agent and I’m going to help you with your support ticket.

That’s one definition of agent. And it’s this very product anthropomorphized. I’m an AI agent. That’s one.

And the value is another term. There’s another term. Just because like. Let’s let’s put value creation next to it.

But those the value is it’s an existing problem. And the agent does this existing problem half the way, fold away and is cheaper. So that’s a value. That’s right.

Yeah. Yeah. And it’s actually it’s actually an agent. Like it’s like I’m a sales agent.

I’m a support agent. So that’s one. There’s a second one that’s more of a technical definition. And it does hold water, which basically says you close the control loop with AI so that the output will become the input.

And you can do multiple iterations without the need for a human to be involved to solve higher level tasks. Right. So if I say go book me a ticket, there’s many steps that are involved. And so it can do this by reasoning about what to do next.

And then it will do multiple steps to do that. So that’s that’s one definition. By the way, those are kind of not working. The tech isn’t quite there.

Like these elements are very unpredictable. It’s very hard to kind of nudge them in a way that you can actually get predictability for complex tasks. And then the third definition of agent, which is probably the most confusing, is if you if you’ve heard like NCP, like, you know, like the model context protocol and like agentic this and agentic that and tool use and whatever, that’s kind of generally used for any time an LLM calls out into a tool like I’m going to like send an email or I’m going to do this. And so like they call out.

This agentic, but it’s not really the formal. It just means any time to do it. That’s cool. So I will say like the term itself is almost useless.

But like the things it’s talking about are very real and they’re very real trends. Yes, I love this. By the way, this call out thing, if you would take that description, then the fact that chat GPT for actually could do math questions, that was the first agentic call out because it actually is. It’s a calculator, right?

And so exactly. Exactly. Yeah. And then for that, was not saying agent in the in the sense of like the closing control of it.

Just it could actually go out and it had some agency when talking to something else, basically. Now, we have two questions which actually want to funnel here because we have just mean and rocky Russell and both kind of saying, OK, when we are a non-AI company, we hear that agents are. Like our big savior, right? Like because we can improve workflows, we can put them in and similar to the cursor discussion.

We had a workflow and we had a model and tropic, by the way. I mean, yes, I’m truck because better and writing code. But curse, I got all the fame because it was integrated into a workflow. What is your advice to enterprise transformation, enterprise companies if they say, oh, I see all of the things happening?

And Martin told and tells me that there’s a huge shift. But I have my workflows. I don’t know how to participate in this huge shift. I mean, this is the the CEO issue of all CEOs questions, right?

Like the DC is trying to answer that. Then you’re in trouble. I will say I will say one thing was I sit on a lot of boards. I’m on like, what, 24 boards or something like that.

And like guaranteed some DC on the board is going to be like at AI. We need to. We need more AI. Right.

Like without actually understanding like the market. This is like this is like. This is the most founder centric question that ever exists, which is you’re in a space. The space is being disrupted.

What do you do? And the first question you have to ask yourself is like, are you really being disrupted? Because the VC has no clue right on your board. If you’re being disrupted or not, there’s plenty of companies that are doing great that are adjacent and are drafted on AI.

And they don’t add any. There’s plenty of those. So you first ask that question. The second is, if you are being disruptive, then you need to basically redo the product journey again.

So I view all companies go through three phases of the product phase, the sales phase and the operations phase. You can’t skip phases. And sometimes you have to go back to the beginning. And listen, I mean, we’ve seen that we’ve got great class examples of this.

Netflix is probably the best ever. I mean, from DVDs to total self disruption. Like we’re in Matter when I sit in these board meetings and the VC is like, add AI. It just drives me nuts because it’s not actionable and it’s often just not even true.

Yep. Yep. And I like, so for a while I was advising the Deutsche Bank and this was like, so it was every week, one day, every second week, one day over in Frankfurt and talking to the CEO. And we had a lot of discussion about how to use data and AI and it’s not sprinkling completely, right?

Because it comes down to the fundamentals. And as a, I think the big advice you give here, and which is really helpful, if you use AI, you should always ask yourself, how do I make either something better? Because you are in the existing product bucket. You have customers, you have a workflow.

How do I make this better? Or how do I save costs? These are the two things you can do. That’s like, such a good question.

Yeah. Such a business. Yeah. But I also think you should ask, am I being disrupted?

Like you have to ask that hard question. And like, if you’re being disrupted, like adding AI doesn’t help you. It requires sometimes these very fundamental changes to the company. And again, we have these very iconic studies of this.

Like Netflix is my favorite because the cannibalization was so brutal and so complete. If you read the book, they’re like the existing exec, like the old execs were so pernicious. The white blood cells were so dangerous. So we literally had to kick them out of the room.

Right. I mean, and again, it just all, it just all, it all depends on the nature. There’s no answer to this question and people want to reduce it to like a simple thing. So I love that this question came up because it is the most important question.

Unfortunately, it’s the one that like makes founders, founders. Really? Like, it’s so funny that you said, like, you’re doomed if a VC needs to tell you that. I’m totally with you.

No, there’s another question, which a lot of like, I’m counting now here, three people who asked me this. And if you just tuned in, ask questions, I will feed them into our conversation. But like, and you do this in the live stream, but one of the questions, like three, three people asked something like, what can we do to not be obsolete? How do I know that I’m not going being disrupted?

And I learned something from you, Martin, in our prep work. I should have known actually, but I learned, it became actually clear once we talked, that you’re saying, I’m not predicting the future. I’m predicting trends by looking at very good founders where they are going. Now, which industries like do you think are being disrupted at the moment where people should worry?

Actually, they shouldn’t worry, they should change companies. I mean, this is like looking at a hurricane and trying to predict the past. Which you kind of can, but you really can’t. And locally, you definitely can’t, like you don’t know how it’s going to stand, which is not going to stand.

I mean, these are very complex, chaotic systems. Let me say a few things. Listen, I’ve been, I’m an old guy, I’ve been boots on the ground for major disruptions, like major, like PC to SaaS, like internet, like cloud, like mobile. We almost never remove layers of the stack, we just add it.

The TAM just gets bigger, it doesn’t get so much bigger. It just gets so much more complex. It just gets so much more smaller. And yes, like what is in demand, you know, like grows, but it’s always adjacent, right?

So it’s not like there’s just going to be mass, you know, shifts, at least in tech. In tech, it’s still software. You’ll still need software developers. You’ll still need designers.

Like, I just think that this grows the TAM of tech. There are clearly areas outside of tech that will be impacted because of the nature of these models. Like, so for example, I’ve got two cousins, you know, live in Berlin, and translators, pretty high-end translators, and they visited me recently. They’re like, listen, we’re looking for different jobs.

And it’s not because our jobs got removed, but like, they want us to double check AI. And I don’t want to do that because I’d have to rewrite everything. And like, it’s a soulless work. And so I don’t want to trivialize the impact to on-the-ground work.

And I do think if you’re in design areas, you know, the world will change. And I think if you’re in language areas, like the world will change a bit. And I do think that there is like a societal imperative to step in and help with that. And that happens every one.

It happens with the compute revolution. You know, so I think that. Photography made paintings different, right? It doesn’t mean that we don’t have artists anymore, right?

Right. We see it all the time. But I will say, I’m going to just answer, like, I’m not a sociologist. I’m going to answer strictly from the perspective of tech.

I mean, from a perspective of tech, this is value accretive. It is TAM accretive. It’ll be larger. I definitely think that you need to, like, if you’re a founder or if you’re running a company, you need to do like you do for any sort of disruption and understand the impact.

And you need to navigate that. But just realize, like, the future is way brighter with this stuff. And that’s for sure. By the way, the internet, people are even more scared of the internet.

Like, oh, there’s going to be one company and it’s going to be Microsoft. It’ll have all of the apps. It’ll have the internet. It’ll have the entire web and blah, And like, none of it’s true.

Nobody forgets how scared everybody was. Like, totally same vibes today. I love that reference. Yes.

Pretty amazing. Now, we discussed workflows and, like, the value. We discussed those different, either improve something or create something new. We talked about the importance of fundamentals.

There is one topic which you very often bring up when you talk publicly about startups. And it’s a little bit the go-to-market strategy. That, like, having a cool idea in itself, as you said earlier, and it’s true, like, if you do a category, it’s chewing glass. Like, talk us a little bit through go-to-market strategy thoughts and maybe take a company from your portfolio.

Well, a very interesting thing that happens in these super cycles is the enterprise does not know how to consume them. And so, you know, like a traditional, infra company, B2B company, your life is dictated by sales, marketing and sales, right? Just because, like, you know, the enterprise is buyers and that’s the majority of the money. And, you know, like, and I spent my entire career basically cutting my teeth in enterprise, you know, sales.

But these super cycles are so different that the enterprise doesn’t even know how to think about them, right? It was the same thing with the internet. I mean, people don’t remember, like, the PC was a consumer phenomenon. Like, the smartphone, other than, you know, the internet was a consumer phenomenon.

The reason it was a consumer phenomenon is not that it’s not useful to business. It’s that business, like, can’t figure it out yet. Like, early on, there were companies that banned the browser. Like, some microsystems famously banned the browser for a period of time, right?

You know? And so, I would just say, like, the meta- The meta point on this topic is, if you look at many of the companies that are being very successful right now, they actually have these prosumer companies or they focus on individuals. Like, Cursor focuses on individual developers, Midjourney, consumers. Even, like, ChatGPT is very much this consumer thing.

And the reason is, is because, like, you know, you go to the enterprise and, like, I don’t know how to think about it. How do I operationalize this? What budget does it come out of? Like, who, you know, like, whatever.

Like, it’s very, very complicated to, like, get them to think about it. Like, you know, you’re in Matter and as budgets get created and as organizations shift within large enterprise, then the most important thing is the sale, at least for B2B companies, is the sales cycle. And you really have to become a student of go-to-market. And so probably the steepest learning curve I’ve ever had in my entire life, in my entire career, was nerdy, poorly dressed PhD with a company that had to learn enterprise sales in Japan, right?

And I mean, it’s like, you know, this is so terribly different. And so it is a very, very core aspect. But in this stage of the super cycle right now, we’re just not really there yet for many companies. Some companies are, but most companies are not.

Fascinating. Let me challenge this. Like, I’ll push back on it. So when I see the consumer behavior, yes, and like, I’m with you on Cursa, I’m with you on OpenAI, I’m with you on MidJourney.

However, I would claim that in the broad application, this comes down to our agents, where you said, like, look, they are hallucinating. It’s hard for them to keep track. It’s hard, like, reasoning can get you anywhere. That it actually has to go through a confined space, a confined aerospace, or a confined area where you cannot go too much right or left, where you have a path.

And that would be the enterprise. If I want to book, like, you know, OpenAI, like, put on their website, oh, this is so cool. You cannot book a trip to Japan. Really?

No, you can’t. Now, if you are somebody who is only booking trips, I can have an agent to actually check certain things which are in my workflow. Wouldn’t that mean that enterprise adapts first? Well, I mean, let’s just go through the companies and the numbers.

Name one top enterprise company right now in AI. Mistral? I know. Yeah.

Yeah. I mean, this is open source software developer-y. I mean, clearly they’re doing some enterprise sales. I mean, I will say OpenAI, predominantly consumer, of course, that enterprise is turning on now.

You know, 11 Labs, you know, private consumer, of course, enterprise is turning on now. Ideogram, totally consumer. They are, like, Harvey, of course, sells to legal. So, like, if you’re selling to, like, an existing motion budget buyer that’s non-tech, maybe, right?

Yes. You know, or something like that. Yeah. Or something like Glean, which is enterprise search, which is an existing category, great product, maybe.

But, like, these phenomenon, these high-growth phenomenons you’re seeing, a lot of them are – it’s not that they’re a consumer. Like, these are used by professionals. Like, Ideogram, which is a text-to-image company, like, a lot of the users are designers, like, legit professional designers in the profession. But, like, it’s a credit card from an individual because there is no central buyer for this type of things.

And so, this is a natural. It’s a natural evolution of these super cycles. The internet is exactly the same. There are, of course, exceptions.

Absolutely, there are exceptions. And over time, the bulk always ends up in the enterprise. And that, you know, that’s the direction it’s going. But we’re not there yet.

Got it. Understood. Yes. And this is actually – I mean, I would – the idea of, if you take, talk Harvey, like, all the – if you go the consumer route, that’s probably the biggest price you can get.

And, yes, these are the big names. Harvey is, I would say, doing good business, right? Great. It’s great.

It’s great. Optimizing on one area. Yeah, yeah. But also, it’s also a non-technical – I mean, I would definitely caveat what I’m saying for, like, non-technical use cases, like, you know, whatever.

Like, you know, debt recovery, legal, you know, or existing buyer behavior, existing product categories, like enterprise search. Like, Glean is doing great in enterprise search, right? Like, and, you know, whether it’s AI or not, I mean, like, it’s doing a great job in enterprise search. Yes.

But for these new behaviors, which is like, you know, the magical creation of stuff. Like, it’s – by the way, if you haven’t read Chris Dixon’s, you know, the next new thing will start as a toy blog post. I think today still is the most prescient blog on organic adoption of new technology. And it basically says what nerdy hobbyists do on the weekend will always look like a toy, but that’s what becomes the next new thing.

And it always looks kind of weird, and it always is something that, like, starts with these kind of fringe groups. And it’s very different. The kind of innovator’s dilemma kind of, like, crossing the chasm view, it really is about these organic phenomenons. And we’re witnessing an organic phenomenon up close.

Now, give me a second to dig into Glean, because Graham Pearson talked in one question about the value of data in terms of – and he called this outside-in versus inside-out, meaning trained with external data versus trained with internal data. And he listened to us. And, like, his question is, how do you value data assets? Because ChugGPT didn’t come really up with ChugGPT 5.

There is a kind of a capping out of what we understand of human language. Like, so where is data valued, and will that change the equation on enterprise versus consumer? I mean, the marginal value of data is very low. And it always has been.

It’s never not been. Actually, let me dig into it, like, just to make – like, the marginal value. Yeah. Yeah.

Just make sure that people understand this. If you want to get paid for a Facebook post, like, one single Facebook post doesn’t help the model to learn anything. This is one of millions data points. Therefore, the value is zero.

So I’ll be – yeah, so I’ll be more precise, which is once you’ve collected a lot of data, the next bit of data you get, the value of that converges on zero, which means that the leaders will always be moving slowly. Right? They’re not moving slower than the challengers, which is a perverse economy of scale, right? It’s the opposite of a network effect.

So what is a network effect? A network effect is if you actually have an interconnection of pairs, right? Like an actual network or a social network. The thing with a network effect is the number of connections grows super linearly with the number of endpoints, right?

So, you know, like, the more you add, the more connections you have. And so the marginal person, like, the value increases, right? So they’re – like, the value increases, right? like the price to add the next person, that value converges on zero.

And so in network effects, things like social networks, like LinkedIn, you know, Facebook, like physical networks, you have great defensibility, great economies of scale, and you have these monopoly businesses. With data, it’s always been the opposite. Perverse economies of scale, the leaders slow down. There is no such thing as a data moat.

That’s non-contractual. Of course, there’s contractual data moats, which is I have the data that you don’t have, and you’ll never be able to get it. But one thing that people misunderstand about data, and it’s such an important point, I’m just going to take a moment to say it, is data sublimates into the ether, even if like you don’t want it to. Like, let’s just say you have these textbooks that only you have the copyright to, and nobody could ever read these textbooks for you.

Only you have those. Okay, so people can read them, but they can’t copy them. They can’t do anything. They can just read them.

It’s in a library somewhere. So you can just check it out, but you can’t copy. You can’t scan it. The reality is like people will read these books, and they’ll write blog posts, and that data is in their head somewhere, and that would be written down somewhere.

And so the actual knowledge is finite, and it just kind of sublimates away from the sources. And so over time, there’s a fixed set of knowledge that everybody ends up getting, and it all ends up becoming public. And so it’s just not the right strategy for a company, in my opinion, to try and believe data is going to save you from an economic moat. It’ll save you in many ways, just not that way.

It’s not the right moat. Yes, I’m with you. And we saw actually this levering out so many times, over and over and over again. The reason that Netflix is amazing is because they control the go-to-market.

They control the access. They have the right- They’re a brand monopoly. They’re a brand monopoly with a distribution monopoly. I mean, it’s like- Yes, I started- But also like people will use examples like Google.

Like Google was an algorithmic advantage, and then it was a two-sided monopoly. It was a two-sided marketplace between people doing content and people doing advertisement. Now- I start my courses actually with exactly that point. I say like, guys, you came to me to learn about AI and data.

Let me tell you, that’s not what makes you rich. That’s not the business model. Yeah, No, And since we are going almost to the hour, and I could do this forever, let’s talk about the super cycle. You mentioned it.

You said it’s a perfect storm. And we have seen perfect storms before, and they can lead to unexpected consequences. And I have a list here from a few folks. David Meyer was watching us.

He says, well, what about ethical AI? How do we deal with life sciences? Or Carissa asked about what if we kind of use LLMs now on Wall Street? Will we not create an LLM bubble, an economic bubble?

So what keeps you up at night? What do you think could go wrong? Could there? Well, let’s do the ethics one.

This is a very important question. What’s so strange is how the conversation has become so irrational. Like, I remember the rise of the Internet. Like, we had like full on like net new marginal risks that were demonstrated very early that took down critical infrastructure.

And like, in a way, at that point in time, we weren’t scared of our shadow. Like, we were like, okay, let’s identify the marginal risks. And then let’s go ahead and adjust the policy infrastructure here. And what’s different with AI is nobody’s identifying the marginal risks.

And in fact, if you ask many of like the experts, if you ask like Don Song, right? Like, you know, Berkeley professor, you know, MacArthur Genius Fellow. You’ll be like, we don’t, you know, like AI safety research. She’s a safety researcher.

She’s like, we don’t know the marginal risks. And like, we don’t have the same examples that we do. We did in the Internet age. And so I feel my big concern on a lot of the ethical stuff is somehow there’s this anthropomorphic fallacy that people apply to AI.

And we think, oh, it can do anything and all this terrible stuff. So let’s go ahead and get ahead of it and put the stuff in place. But if you don’t understand technically the marginal risks, you may end up putting the wrong policy for the wrong thing and making us less safe. And it’s a very important point.

I’m not saying, I’m not saying, listen, I’m not like some libertarian. I’m not saying, you know, whatever acceleration is at all. Like, I’ve worked in policy. I’ve like taught, like I taught a Stanford course on policy.

It’s very important. But you have to know what you’re policing. It’s so important. Otherwise, you may have ineffective things and people get hurt, right?

You don’t want people to get hurt. And so I think ethics is super important. I think we’ve got great frameworks to think about it. I think these frameworks have evolved over every, like, you know, technical super cycle ever.

And we should apply them here, but we should apply them sensibly. And we’re not. And I think that’s a huge issue. Yes, that makes perfect sense.

I mean, one of the discussions, which is very dear and near to my heart, is you need to understand the technology. If you do not understand the technology, then you can’t make money or business. But you cannot also not regulate. And we had this whole idea about AGI is coming.

They will take over. Everything, like CEOs will be replaced. Everybody will be replaced. And it’s such a, I can’t say the word.

We’re on a live channel here. But, like, it’s such BS, right? Like, you don’t. This is a smokescreen which leads us in the wrong direction.

And I love it the way you describe it. We should look into fundamentals. We should look at how. Marginal risk.

Understand the marginal risk, which is, like, we have an existing policy framework that we’ve put in place. And the discourse has been around for 20 years. And we’ve evolved it during that time. And you want to make sure that that discourse, which is hard won, is applied to this technology.

And then if there’s a. You have to understand what that shift is. And you have to evolve your policy framework to that. Unfortunately, that’s just not what’s happening.

I think it’s like goes back to the Promethean legend. I just think we’re scared of things that sound smart. I don’t know. Actually, that sounds human.

I think the. That sounds human. That’s right. The thing which scared most people is that that thing suddenly.

I always joke, like, ChatGPT is like your loved partner at a consulting firm. Sounds super smart. Doesn’t know a lot. Right.

So the whole. Yeah, we’ve got a weakness. And as humans, we have a weakness for this stuff. I mean, it’s like it goes back to Prometheus.

I just feel like we’re kind of scared of these things. And instead, we should be incredibly rational and do the right thing. Yes, I think this was an amazing discussion. We had actually some listener actually wrote even a nice comment.

Is that like this was one of the most informative conversations I have heard since the 1960s. Well, you should come to this. Channel. That’s lovely.

I appreciate that. It’s very. So thanks, everybody, for joining us. Thanks for all the questions.

And Martin, big thanks to you. This was exciting talk. No, absolutely. OK, thanks, everybody.

Bye.