软件终于吞噬服务业 - Aaron Levie

The universal adoption of this as a consumer technology, and then its bleeding into prosumer, exceeds anything I’ve ever experienced. I think it will fundamentally change people’s daily patterns. This is all early adopters, and early adopters are very forgiving of mistakes, on purpose. When something is brand-new, a culture around it develops. People on the early internet didn’t complain that the internet was slow, right?

The more senior small teams that use AI are superhuman.

Yeah, it’s like they woke up and they were all Tony Stark. It’s unbelievable. Their productivity is insane.

First, I just want to comment that you posted in the group chat that the news around autism updates your p(doom).

Yes, it only works if you show the image, though. You’ll have to do the overlay to make them make sense. There are so many memes you can do with that Fox News headline.

Exactly. First, I want to get into the immigration news.

You really want to kick off with the fun stuff? Get the blood pumping.

Exactly. Martin, you mentioned this. What were your reactions to the policy?

It was interesting because it seems like anytime the administration touches immigration, there’s a huge, knee-jerk outcry, and we saw a lot of that from VCs, even. But it’s also very interesting that Reed Hastings, who is a classic lefty and has long been doing immigration policy for 30 years, said, “This is the right approach.”

That very much reflects my thoughts. This system has been gamed for a very long time. It’s very hard for startups to hire because of the lottery system, and it’s been locked up by the large companies, the consultants, Amazon, and Google. That has to change, and I think a very reasonable way to do it is to set a price. You’ve got a market, you need to allocate supply, and price is a great way to do it.

So I’m very, very positive on it. I’ve commented about that, and a lot of people seem to disagree. I think it’s an active discussion.

I think there are a couple of elements to this. First of all, Reed was ultimately responding to something that was no longer the actual policy. He said $100,000 a year was a great policy, and obviously the internet had moved on.

It’s not obvious to me that I would conclude the same outcome you just concluded: that the Amazons and Googles would probably capture the vast portion of the talent in this situation. It’s not clear to me that startups come out ahead or are better off from this particular implementation. Maybe Amazon and Google are easier to regulate, but there are a number of organizations that are consultancies and are price-sensitive that would be squeezed by this.

Given that they’re in the top 15 and make up 4 or 5 of them, that would be a significant freeing up for a higher-level labor pool. My thing would be: get all the people in the room who have an opinion on this topic, including practitioners in tech and, let’s say, the most right-wing people—though you can’t even say “right-wing,” because I don’t think this is even classic Republican—and ask, “What are we optimizing for?”

Are we optimizing for not wanting wages to go down? That’s an interesting thing. Are we optimizing for a particular kind of job not going to certain populations of Americans? Are we optimizing for ensuring that only the highest-merit people on the planet come here? Those are all totally different goals to optimize for. The framework and system you end up with should hopefully have a cohesive strategy behind it.

My strategy would be that we want the absolute best in the world here. It’s not exactly clear that there’s a fixed number for that. Some years there might be 5,000; some years there might be 50,000; some years there might be 80,000.

We probably want them to be net positive for wages. Let’s agree that, in any given industry or locality, wages should go up with this talent pool as opposed to down. You should have some market dynamic to that, and you shouldn’t be able to game and exploit talent pools by saying, “Now in Detroit we can wipe out IT jobs because we can offshore those.”

You could build a system that meets all of those goals while still ensuring that somebody who goes to a master’s program at, name your state school, comes out of it as an AI engineer. They’re not yet at the level where Meta is going to pay them $100 million, but they’re going to be totally valuable contributors to our economy. It’s all positive-sum. It’s not taking a job from anybody else; it makes us more competitive.

I think there’s a way to do that without overly putting constraints in the system that make it so a startup wouldn’t be able to economically participate. I think $100,000 per year would be a point at which startups would be directly impacted.

Working with a lot of startups, I’m not sure that’s still true.

Respectfully, the kind of startups that a16z sees are not all of the base of startups in the world.

Before we fall into quibbling about the number, is it $20,000, which Keith Rabois said and I thought was very sensible, or is it $100,000? I don’t know. But the idea that you—

Keith threw out $20,000.

Let’s just go with Keith’s number. If Keith said $20,000, I think we can be good with a Keith number.

I think it’s easy to fixate on the number, but you always have to look at what the number is replacing. I don’t think the average person having this debate, other than the people who really work on this, has any idea of the amount of productivity that is lost working this system.

The incredible amount of resources the bigger companies you mentioned have is enormous. They spend all of their energy, essentially, as lobbyists working this system. On the back end of that are all the justifications, management, and handling. Now they’ve deployed all their resources to manage in-house call centers to deal with getting their employees back to the United States.

You hit on one point that I think is really important to this debate and is getting lost. Within the big tech world, for the past 25 years or so, they’ve gone on this bifurcated curve: hiring for the people in the office while focusing on 25 or 30 university departments, and then deciding it was easier to hire huge numbers of people from 8 international locations and schools.

A lot of this is missing the fact that, if you look at Intel and where they all went to college, and if you look at the history of Silicon Valley, it’s all these people from schools in the middle of the country. None of those are the target recruiting schools for the main tech companies these days.

That’s been a place where I think the big companies have been somewhat lazy. As a person who spent decades flying to all of these schools and recruiting, I think there’s work the universities have not done to build better programs, and there’s work the big companies have not done to be clear about why they’ve stopped recruiting or haven’t seen the numbers. Making that change would be better for everybody.

Yes.

I agree. My expectation is that what gets impacted is an even different job set than that. If you go to Florida today and try to get an IT job for $100,000, you just can’t. That, I think, is the area most directly impacted by the large consultants.

Do you mean there aren’t jobs that pay $100,000, or that you just can’t find a job?

They’re taken. They’re all out there; they just don’t exist. Any sort of IT administrator job, the services, the basic consulting gigs—all of that has been saturated. It’s very, very tough to get a job between $80,000 and $120,000 in much of the United States because of this.

This isn’t about a new graduate being a software engineer, because the reality is that the expected value of a software engineer over their lifetime is high enough that the market navigates that. It’s almost these lower-level, more IT-administrator jobs that have been squeezed out.

If we want to bring them back, and we don’t want to do this kind of arbitrage that a lot of these companies are doing, we’re going to have to change the pricing.

But wouldn't a minimum salary band effectively solve that problem for you?

Sure. Yeah, for sure. I think there are a lot of mechanisms to do it. I actually agree with you: we should talk about the problem we're trying to solve. I think we would all agree—

If you come to the United States, get a college degree, you should have a visa, right? Kind of.

Well, wait. Do we all agree on that?

Okay, yes, I agree.

Okay. But I don't think that the people proposing this strategy actually agree on that. Trump famously said this.

Yeah, he famously said a lot of things, for sure, and then he unsaid them. This wouldn't be an interesting debate if we all agreed on that and we had a policy that said if you go from IIT and then come to Kansas State, you get a job. That became part of gaming the system, because you would do the IIT thing, then get a company sponsorship for a master's degree at some school, so it didn't really accomplish the goal of investing in coming to the United States the same way that going to a 4-year school would have.

Wait, why is that?

What was the problem?

Because you ended up getting sponsored by a company. It changed the whole dynamic: were you seeking out the United States, or did a company pull you to the United States? It's a little different.

Okay.

But I also think that so often this discussion goes the direction this one has gone, where we focus on new-grad software engineers. I actually don't think that's what's getting impacted—

Right?

I really don't. I really think it's about what the consulting shops do. They do admin work and IT work—

And it's just a different salary band.

I think these are body shops, and an approach like this directly targets them in a way that's actually quite positive.

Yeah. I just think there's—then I would favor Keith's approach, because I think there is a number at which you're making other trade-offs in your business just to be able to—

The $100,000 number is—

Well, I do think that the dollars—but I do think, just to reinforce this, that—

The whole system can do without this immense cost and uncertainty.

Yep. And any solution should really, if you address that—

Yeah, then the rest of the dynamics will follow. But as long as it's a huge, complicated, expensive system, the big companies are going to continue to benefit from it disproportionately.

Yeah. And I will tell you right now, it is much harder for a startup to deal with the lottery system than it would be for them to pay $100,000—at least for startups.

But I don't think it changed the lottery system. It's a—

Right.

Right. It didn't.

It's $100,000 to participate in the lottery system.

Oh, no, I understand. Hopefully, a number of people will change the calculus for a lot of people who are in the lottery system.

I just think—

But I would prefer to remove the lottery system.

I think you can absolutely pull off a system that, probably in a shared definition for us and anybody else, would say this is clearly a high-merit job. It's going to increase the wages in this particular sector on average, and we want to make sure that we've got the best talent in the world coming in to do that. It's not going to drive down wages. We probably want as many of those individuals here as possible. I think some elements of this are intriguing in that they push the conversation forward on the dimension of the $100,000. That's a hammer to do that, and maybe there's a more nuanced approach that I would certainly prefer.

But again, it's important to note that the $100,000 will scale with the skill level, right? The higher the skill, the more that is amortized. You could argue that this is just a dial to get higher skill.

Well, once you put a dollar amount on it, you have to keep in mind that people are going to pay it. That's right. They're going to make up their own minds, and the skill might not be relevant to some people. That's the tricky part.

Yeah.

I want to segue from labor markets to labor productivity with AI. Offline, Aaron, we were talking about the METR paper, and the paper suggested that developers were actually less productive with AI. But that doesn't square with your experience talking to a lot of different startups and seeing how they're so much more productive. Why don't you talk about where you're seeing startups say they're more productive and why that's happening?

Yeah. I'll first just represent our own case study, and then there's the really extreme version. Our own case study is that we've adopted a few different AI coding tools, Cursor being a super popular one internally. As I talk to people, let's say, in the hallway, maybe they're trying to get me excited by AI, but I think they know I'm bought in. The qualitative answers I get from people—and then I'll give you our internal metric—are that some people will say they're getting a 20% to 30% productivity gain, while other people will say 75%. Interestingly, I have not been able to pinpoint the demographic difference in the answers.

Oh, but this is self-reporting.

This is self-reporting.

How happy are you?

Yeah. No, but we have internal metrics as well. About 30% of our code right now is coming from AI, so we've got the—

70%?

30%, 30%. So we have some of the pure internal metrics that show this.

What's interesting is that I have senior people saying they're getting 75% productivity gains, junior people saying they're getting 75%, and vice versa on the 25%, let's say. I haven't been able to quite figure out a pattern. Maybe, except for—and we've talked about this a little bit, but you see this online—the biggest criterion is just the people who actually push the AI to do more. It's this other new psychographic: who is willing to say, "You know what? I'm going to YOLO this task and just see what the AI comes up with." Your willingness to do that probably shows up in the ultimate productivity gain.

That's us as a relatively larger company. On the startup side, what's crazy—and this is the thing that just blows my mind—is that I will regularly talk to 3-, 5-, and 10-person startup founders who self-report that they might be getting somewhere on the order of 3× to 5× to 10× productivity improvements. The big difference is that a year ago, if we were to have this conversation, the conversation would be about AI doing type-ahead, and it could maybe add a few lines of code to your productivity per incremental unit of work that you gave it.

Now, obviously, the big phenomenon is background agents, where I give it a very detailed prompt, send it off, and it comes back. People talk about it as a slot machine: some percentage of the time, it's not going to come back with the right thing, and you have to decide which parts you actually pull in from it. But the startups getting real multiples of productivity gain are fundamentally engineering in a different way: they're sending off a task, the task goes off, comes back in 20 minutes, and they're really in the business of doing code review, not code writing.

It's obviously going to change quite a bit of what computer science looks like in the future. The only question is what all the things are that that's good for, where that breaks down, and what kinds of teams can actually evolve to that state. But that has been blowing my mind the most recently, and I think it fundamentally changes what the future of engineering looks like.

I think what you said is super interesting. Let me ask you: I think there's an overlay that goes beyond junior and senior, and it's that we're all talking about characteristics that a solution has—2 really important characteristics right now. One is that it's engineers doing stuff for engineers, and they understand the domain super well. I think that's a really important part and a really big thing that people aren't talking enough about, which is: maybe what's going on is that you have AI accelerating people who work in the domain and are very smart.

Yes. And then the other is, we shouldn't forget—and this is to your self-reporting a little bit—but also that this is all early adopters, and early adopters are very forgiving of mistakes on purpose. It's a super interesting dynamic where, when something is brand new, a culture around it develops that just lets anything happen.

The early internet people didn’t complain that the internet was slow, right?

The only people who complained that the internet was slow were the late adopters, who were like, “Wow, this is so much slower.” Take online video, which was like, “I’m watching this tiny, tiny postage-stamp video.” The early adopters were like, “This is the coolest thing I’ve ever seen.” Everybody else was like, “Why would I want to watch anything like that?”

The same was true of downloading music and everything else.

Do you remember that you guys made this watch—the SPOT watch?

The SPOT watch?

Yeah, the SPOT watch. So, in 2003 or 2004 or something.

So I bought one, and it used unused FM radio bandwidth.

Yeah. You could get stock quotes that were 45 minutes delayed if you were outdoors in a field.

Yes. I had one, and I was in this camp that thought, “This is the coolest thing in the entire world. Obviously, this is going to be the most mass-market product of all time.” I was 20 years too early with the Apple Watch.

Turn-by-turn GPS—the first time you could put turn-by-turn GPS in your car. It was super cool, except for the fact that most cars moved faster than the ability of the computer in the car to calculate when you were going to turn. You just drove around. It was like a U-turn machine.

I think it’s so interesting because people use these AI tools today and assume, “Well, I’m not a doctor. I don’t know anything about being a doctor. Let me ask you to cure me and diagnose me.” It’s like, whoa, that is the worst.

All the people who see failure—they just weren’t great programmers to begin with, right? They didn’t know how to ask, and they didn’t want to review it. Whereas great programmers, or just professional programmers, know that code review is really important. That’s just how you do it.

I think there’s an aspect of this that makes it very difficult to measure. One of them—and I don’t think it’s just an early-adopter thing—is that these models are so magical that you get dazzled. Even if it’s not what you want, you’re like, “It was great.”

I think it’s very easy to conflate that with being productive: “It’s not what I wanted, but it was amazing, so therefore it must be great.” Maybe, over time, we just abdicate having an opinion and the model does everything. But right now, I see this a lot: People are enthusiastic about using AI, but it really hasn’t impacted their output. They’re just enthusiastic. The second one is that I feel like there’s almost shadow productivity.

Well, sorry. How would you verify that with a 5- or 10-person company that empirically is operating like a 50- to 100-person company? You can see the sheer scale of their code and think, “Okay, you could not have done that 10 years ago.”

I actually agree with everything Steven was saying. Anecdotally, I work with a lot of companies, and the more senior small teams that use AI are superhuman. It’s like they woke up and were all Tony Stark, and it’s unbelievable. Their productivity is insane, but they’re all super senior.

Look, I don’t want to take anything away from them, but those companies were also incredibly productive relative to a 10x-bigger company because there’s no code and they’re starting from a clean slate.

No, they really wake up. Yeah, for sure. They’re very senior, but they’re also, almost to a person, AI skeptics to begin with and incredibly sober about the value. They just use it in these very pragmatic ways.

There’s one other category that I’m seeing, because I want to be intellectually honest about the full spectrum. I’m seeing these 19- and 20-year-olds. First of all, I don’t know what is in the water at Stanford, MIT, and so on right now, but everybody’s dropping out. People are literally going there for a week just to drop out.

There is a tendency in that cohort. They’re the 9-9-6 people, and there is this tendency: They would have been 10x engineers in a prior world, but now they’re 100x engineers. They’re senior in terms of their own relative cohort, but the way they’re building their startups is completely different.

If I look at how these companies run versus today—so we dropped out of college 19 years ago, which is scary—it’s the biggest change in how you start and run a company that I’ve ever seen.

And I think if you looked at 1995, when you were dropping out of college, versus 2005, when we dropped out of college, the company-building process wasn’t all that different.

The internet fundamentally changed that.

No, no—post-internet. Post-internet. Okay, so you’re right about the internet, because that was a—

Yeah, 100%. We’re not talking about 1985. At 1995, you’re dropping out to do an internet startup. By 2005, other than the fact that our resources were in the cloud versus having to go to a data center—in our case, we actually still went to the data center—not that much about the company-building process was different.

Today, in 2025, everything about how you’re starting your company is completely different.

I think the key through line in all that is velocity.

Yes, exactly.

I think that the internet increased velocity. Yes. And I think that's just super important to what's going on. If you go back to the early internet, when you started a company, there was still a lot of old-school thinking: “What’s your business plan? What’s your plan? We’re going to be in stealth for 2 years.” There was all of that stuff, and it was really Marc and Ben at Netscape who changed the velocity of how companies worked.

The cloud was an accelerant to that acceleration in velocity, and AI is a refactoring of how velocity works.

What was interesting is that even the cloud was this great virtualizer of the physical stuff you would have to deal with.

But that was a 2-year buildout. You could have customers as soon as you had code, which—

I mean, PLG and a lot of these high-velocity startups actually started pre-AI, where you had pretty phenomenal companies. Think about GitHub, Slack, and Figma. You had some pretty remarkable companies that came pre-AI and are drafting on it.

A lot of that was basically cloud, then consumerization, and then unlocking different go-to-market motions. Zoom was a great early example.

The thing is that AI is an accelerant in building the product, whereas what the cloud did was accelerate getting paid—which used to take 2 or 3 years.

I would argue that the cloud also made it much quicker to build a product, because you had all these big—

Well, it made it much quicker for customers to all have the same one. I don’t think it made it 5 times faster.

Fair enough. I think it was like—yeah, you could see your website a lot quicker, but I could see a website in 1998 pretty quickly.

Yeah, but maybe this is an infrastructure thing. Building a big, distributed service was really hard. I wasn’t doing distributed infrastructure in 1998.

Yeah, but I think AI productivity is hard to measure for 2 reasons. The first one I just mentioned is that it’s really dazzling. People are like, “Oh, it’s amazing.”

It’s shocking that people measure the wrong thing in productivity.

That’s right. Literally, it’s the history of productivity measurement.

But also, what happens here is that you have the board saying, “We need more AI.” So they go to some CTO or some innovation lab, and the innovation lab does AI. They build some internal tool, and it fails. Of course it fails, right?

The reality is that this AI wave is so personal. Probably most people in the company are using ChatGPT. There is probably some personal assistant, and they’re using Cursor or some other coding tool. That’s much, much harder to measure because it’s not advertised.

If you look at reports on enterprise projects failing, and you look at what they were measuring, it’s like, yeah, clearly some internal project pushed down by the board, where they hired some consultant to do it, is going to fail. It will always fail. But that’s not actually what’s going on. The movement that’s happening is very secular.

Well, this is the next time that bottom-up adoption is really changing the productivity equation.

And that's something that defies control. Big companies do not know how to deal with that because they want—they need—to control it. They worry about safety, security, privacy, and all of their corporate rules. And then the other thing I think to overlay on that is that AI is a very unique—if that's not a bad way to say it—innovation in that it's non-deterministic.

Yeah. And so all of a sudden you have this very personal and non-deterministic thing, of which the real problem in a large organization with all these pilots and AI projects is that you can't—not just measure it, but you don't even want to put out there a non-deterministic solution, because your whole thing is, “Well, we operate at scale and we have 60 countries.” Yeah, the white blood cell.

We can't put a customer support solution out there if 5 agents in 5 languages all have different answers based on the way that the customer or the CS agent asked the question. And I really feel like that is going to be the hugest challenge in large organizations figuring out how to adopt things: they're going to just get all bunched up over non-determinism.

Yes. I think that we will have to have some new form of measurement, probably in general on this, because so much of the improvement in productivity will sort of be this subtle change of, “I used to go to Google for that, and now my EA is writing emails using ChatGPT.” Where is that showing up, other than just—and we talked about this, I think, in the last podcast—just that we're going to start to do higher levels of work, and that will just end up looking different?

But you won't be able to be like, “Okay, how do I measure what the productivity was?” when it's just like I'm working totally differently.

Yeah. Well, but I mean, code, I think, is such a great example. What do people—in my experience, the more senior folks—actually use AI for in code? It's documentation, writing, testing. I mean, it's a lot of the other stuff that may not actually increase the shipping schedule, but you get a lot more robust code, a lot more maintainable code, a much better architecture, and a much better future-forward architecture. So we could be building tremendously better software but still be shipping features at the same velocity, right?

We might need to just measure quality of life as another metric. Literally, I'm just happier. I don't have to—

Developers don't want to write documentation. Exactly. Yeah. But this is—I think it's so easy to get accused of hyperbole and overstating it and things. But what I think is so key to what's going on is this: it is what you were talking about, which is that the whole notion of what you're going to do at the job is going to be really different.

You know, this is my obligatory super-old-person thing, but pre-spreadsheet/post-spreadsheet is a perfect example of this.

Classic example.

But before spreadsheets, you would be a banker and, you know, you would be like, “I'm supposed to help this company get acquired,” and you would come up with a financial model. You'd have 50 recent MBAs all churning away with their HP calculators, figuring out the financial model. And then you'd go, “Okay, let's do this again, but if the interest rate changes—”

Or if their sources of funds change, and you're like, “Okay, well, that's like a week.”

And everybody has to do it all over again. So the quality of your decision-making—

Was really bad.

Yes.

And so what happened was that, just in 1985, that job completely changed. By 1990, instead of asking the recent grads to do it, you were doing it yourself. I actually remember this absolutely crystal clear. My 2 cousins were University of Chicago MBA grads in 1985. They did not use a computer when they got their MBAs. And when I was talking to them about going to Microsoft 2 years later, they were like, “Well, you know, we have these kids who use a computer for us.”

And they ended up using computers and stuff, but their whole notion of banking was defined by this multiweek turnaround, and then all of a sudden it was just more interns doing their Lotus 1-2-3 thing.

And I think this gets to what's going on with code and startups: there's just a whole different mindset over how much you can do, how soon, and how you iterate. Figma and Dylan, they're always talking about this. Figma is now going to change the trajectory of a design idea from, “Oh, let's iterate over here,” to, “Let's just do it.”

Yes. I mean, this is a weird thing to think about: the percentage of my productivity, as an example. My day isn't representative, obviously, because I'm bouncing between too many different things, but there'll be so many times when it's 10 p.m. In a prior world, I would have sent off a task to somebody—an analyst or a chief-of-staff-type role—to research something. It would come back 3 days later, and then you'd find the answer.

Now it's obviously just, “Kick off a deep research, go to Cursor to generate a prototype, do some kind of analysis,” and you have it back in 10 or 20 minutes. I've just compressed whatever that was going to be, as a serially connected task, into something fully compressed. By the morning, you're kicking off whatever that project was.

It's nearly impossible for me to peg what that is as a number. It's just a fundamentally different thing in terms of what work looks like, because you compress so many different steps of a workflow into a single action. So it's a completely different way of thinking about work. Where does that fit in for you in what we were talking about—what I was asking about earlier—which is, how does your expertise really contribute to that?

And in particular, I think it'd be interesting for people to understand: when you talk to customers, how do you help them avoid trying to get AI to make them do jobs they couldn't do in the first place? How does that—because that's an easy point of failure.

Yeah, I mean, it actually is this really counterintuitive thing. You've talked about specialization on the last one. The biggest gains of AI go to people who have some degree of expertise in an area to know what is actually true, what is not going to work, and what they should integrate from the output of this AI. What are the 2% of things that maybe are hallucinations or took the data in the wrong direction?

If you don't have a deep understanding of your particular space, field, or domain, you aren't able to have the right judgment to make all those decisions. So I think the experts just get more powerful in this world. And that's why I'm not even convinced that you can tell a college student to learn anything different than ever in any other period in history. Be really good at a particular field, and then AI is merely a turbocharger of your capability in that particular field.

If I didn't generally know the things I know about SaaS—which is obviously a really weird expertise, but I'm okay at understanding SaaS—then the things I give to a deep research agent that I then go and incorporate back into work wouldn't make sense to me. I wouldn't have all the context for that one thing that it mentioned. How do I form that into the overall strategy? But because I have some understanding of this particular industry, that just makes me way more productive. So I don't think expertise goes away at all, and I think any of the experts in their particular area just become more powerful.

So we actually have a fair bit of anecdotal market data on this. It's very interesting. If you take a lot of these—let's just take a non-extreme image or video—and look at the customer base for any popular platform, it's very interesting. If you draw a dollar at random that's monetized, it's from a professional, for obvious reasons: they can produce. If you draw a user at random, it's casual and in the tail, right?

There's a clear prosumer movement from monetization. I'm associated with a number of companies that work with professional designers or professional creatives. They spend just as much time on the AI tools as they would on traditional tools. It just turns out the output is far richer.

Yeah, but it's still human taste. There are still very specific requirements. And so I think that, if there ever is an ads model that shows up for AI, there will be a long tail of people who want to use AI but don't have the financial incentive to do it, or it isn't tied to their actual job. We're already starting to see that bifurcate out, and there's going to be another subsection that does.

And my sense is, let's say I'm a casual developer and I'm writing a 3D game, and I want to have a 3D asset.

I’ve got 1 of 2 choices. I can have AI create it for me, or I can contract a professional to do it.

As a developer, I’m not going to create a great 3D asset. Even if I use AI, it’s not going to be a great 3D asset, right? And so I think you’re going to have the same mindset that you have today: either you can hack something up yourself, or you can go with a professional, and the professional will be using AI.

And I think what’s super exciting is that AI has created a third category, which is: I’m not trying to do this as a job. I’m never going to monetize this, but there’s some product utility gain to me that’s worth $20 a month. So my ability to now generate prototypes when I’m by myself, just brainstorming—again, I’m not going to do anything in the ultimate delivery of that, in any functional code—but it’s worth $20 for me to be able to realize the thing that I’m thinking about.

And so there are just all new ways of capturing TAM because there’s all this utility that gets unlocked.

Right. And there’s one more thing that’s worth noting, which is that people who are interested in an area get there through AI right now. This is the number-one opportunity for somebody who wants to enter an area to do it as an AI-native, because the tool can actually teach, and there’s just such a paucity of talent out there that you can fill the gap.

This is also the history of productivity in general: the more tools that you have available, first the experts use them, and then more people are able to become experts.

I think that’s part of managing your own career. I love that point you made about how you still have to be really good at something. I think that if you want to be good at finance or at sales, you should become really good at it and assume you’re going to use AI for that, and you’re going to be better than the people pretending.

If you just take—we were going back and forth about, “Oh, it makes a PowerPoint slide deck.” Well, it turns out that to make a good PowerPoint deck, it’s still a skill, and people still pay McKinsey huge amounts of money for better PowerPoint decks with better pictures.

Yeah. So listen, I contract a lot of work for videos and art, and I have for a very long time, as part of different companies or even here at a16z. I have seen the shift from contracting someone who uses traditional techniques to using AI. The dollar amounts are the same, right? And so, again, this is Jevons paradox all over again: you spend just as much time. It’s just that the output tends to be more dazzling, or whatever it is.

And you should see more versions of it. You should run more simulations. You’ve got more iteration with it, and you’ve got more control. I can have a video where I want a dragon to fly out of the sky, so you have a lot more control as the customer. But for sure, this is not somehow dropping the cost of output.

I wanted to—were you about to jump in? I want to circle back to a point you made earlier about there being 20-year-olds who are building companies in new ways. Remember a few years ago, I think Patrick Collison and a few others were asking, “Hey, where are all the Gen Z super-successful founders?” Remember that? Of course, there was Dylan Field and Alexander Wang, but their companies took a few years to really work.

Now we’re seeing the Cursor founders, the Mercor founders, get to massive scale in a very short period of time. Maybe the foundation-model companies required a certain level of experienced founder because of the fundraising amounts, and maybe the applications are more conducive to younger founders. What’s your reflection on this?

Well, I don’t remember exactly the date at which he mentioned that, but I do think there was a period, in the sort of mid-2010s to early 2020s, where we were actually in a bit of a lull as an industry. The reason for that was this:

We kind of did check off a lot of boxes of the core things that people needed in the world. Once you have Slack, you don’t need 5 other chat tools. Once you have Zoom, you don’t need 5 other videoconferencing tools. And so it gets kind of derivative past these core platforms.

Once you have SaaS, you kind of check off all the major things you do at work. In the consumer world, we had ways of delivering food, listening to music, and watching videos. There’s not an infinite set of things that we do as consumers. Then what is the 20-year-old founder supposed to work on?

They’re going to have pretty finite opportunities as compared to the mid-2000s, let’s say, when the whole world was open. You could start anything because every single category had to be reinvented post-mobile, post-cloud maturity.

So now we have that era in AI. That is why I’m so unbelievably pumped up. It’s because you have a complete reset of the landscape, where there is incumbent advantage in distribution, but that is it. There’s no other real advantages.

There’s a bunch of disadvantages.

Yes, and then there’s a bunch of disadvantages. Yeah, go ahead. Sorry.

No, no, no, but I mean, you know where I’m going. You have the exact makings of a landscape where new startups can come in and do things that incumbents either can’t, or where there’s no obvious incumbent to even do that thing. Again, you’re taking maybe services and turning them into AI labor, and there was no software incumbent previously to even attempt to do that.

Then you have incumbents that have a whole lot of complexity in terms of their ability to execute in some of these spaces. They’re not going to retool their entire internal engineering workflows to move at 10x the pace. A brand-new startup can do that and then instantly get the scale of a larger company.

It’s the first time in history where you have none of the disadvantages of a big company. The traditional advantage you have as a big company is scale and distribution. Scale because you can look at a feature and say, “We’re going to go build that next month.” Obviously, it’s harder because there’s just lots of complexity to that, but at least you have the human power to go do that.

Now, as a startup, you instantly have scale because of background agents, et cetera. So then it’s a distribution game, and a lot of these pieces of software can go viral now in a way that wasn’t possible 10 or 15 years ago. We’ve kind of neutralized a lot of the incumbent advantages, and so it’s a ripe opportunity for brand-new startups.

Often, it will be people coming right out of college saying, “Hey, it’s my first time building a company.” They’re crazy enough not to know how hard it is, so they’ll jump right into markets that otherwise we would assume are already solved for—where there’s no way you’re going to build a company—and you’ll just have new startups that actually go and do it and produce real companies in these spaces.

Yeah. Because this is just so critical. What’s really happening is why it’s an actual platform shift. Silicon Valley has seen this movie many times before, and that’s why there’s often a lot of, “Is this crying wolf or not?” Everybody knows that when there’s a platform shift, that’s the moment in time when startups are at an advantage.

And so each time there’s a platform shift, everybody’s like, “Oh, this is it. Is this going to reinvent everybody?” Then it doesn’t, and people get really like, “Oh, it’s always incumbents.” But historically, the advantages to incumbents are wildly overestimated.

And really, this is the one where—I mean, was the internet undoing Microsoft, or not undoing Microsoft? It’s a super interesting thing because, of course, there’s a $3 trillion company now, but not on the internet in a way that you think about the internet. None of the consumers, none of the platforms, none of the assets that we had in the ’90s—

They became internet assets. I mean, even if you look at Azure today, it’s an amazing accomplishment. It’s not running Windows anywhere.

Right?

And I think that’s why it’s not crazy to go, “Wow, is this going to be good or bad for Google?” There’s a bunch of stuff that becomes really difficult if you don’t make the transition. And then it turns out historically, even if you do make the transition, you really didn’t, and you just have to wait for time to pass.

Well, this is like Intel with the GPU. They missed the GPU in 2005, right?

And they missed the opportunity to buy the company, to do the work, whatever. They kind of missed the data center, too.

It just took a longer time to figure out that they missed that as well.

Well, and we have a pretty narrow definition of disruption, in the sense that we expect the incumbent has to lose for this new startup.

And that never happens.

It never happens. And so it's the whole radio, TV, theater, movie analogy. It turns out that Microsoft can be a $4 trillion company. You can have all these new categories emerge that maybe Microsoft should have owned if everything were perfectly analogous to the desktop days, but they just don't.

It all works together as one sort of ecosystem because it turns out software did eat the world. These markets are actually 100 times larger than what we realized. Incumbents can grow, and then you have new disruptors that emerge along the way.

Yeah. Anytime you bring in a new technology that brings the marginal cost down, the market's going to expand, and the incumbents can do it. I will say incumbents are very bad when new user behaviors and buying behaviors show up, particularly when they don't know how to cater to them.

AI is definitely a new user behavior and a new buying behavior. So this is very much an advantage of startups, just because to change a large company around a new user behavior cuts across the entire company—everything from marketing all the way to support and the back end. That's just too much of a lift.

I mean, if you look at the best practices of even how you would create an agent in the last 18 months, I think we've gone through 2 or 3 architecture-pattern changes.

And so I can barely keep up as a midsize company. I can't imagine if you had so many more people you had to organize around that.

Disruptive new technologies require people to understand how to use them and consume them in different ways. That evolves over time as you develop best practices. That sort of flexibility can only come from startups.

It's actually a very interesting question how Microsoft did Copilot to begin with, because it was one of the first of these products to be very successful. I learned recently that it was created by OpenAI, so that kind of explains it.

Well, but you had a startup person, and conveniently, he was a startup guy. Even then, it's remarkable that it came out of Microsoft.

What's always remarkable is when something new and defining comes from these big companies. Basically, it comes from skunkworks. They had nothing to lose.

And it didn't interfere. The iPod is a classic example, or the iPhone. People always talk about how brave it was, but Apple's computer business was dead—dead. It had 3% share and was going nowhere. It was a Hail Mary. The iPod was a Hail Mary, and then the phone came along. They weren't in the phone business, so it didn't matter.

The fact that they made a phone that was a little computer was what made Nokia say, “What is that?” I think people really need to wrap their heads around the fact that, to your point, the big companies stay around for a very, very long time.

This is something I've seen a bunch of people say in the past couple of weeks: “Oh, it's basically Clay Christensen's The Innovator's Dilemma.” Of course, nobody's ever read the book. Clay was a great guy—he was down the hall when I was teaching there—but the book is about 2.5-inch disk drives and a bunch of crazy things. I really don't think it applies to modern stuff.

The thing that he missed, aside from the cost and the low-end and high-end issue, was also that the companies don't evaporate.

And your point about that is really important. There's this shadow that everybody is worried about, and so you see it constantly. When a company says, “We're not worried about Google doing this,” that's what you want to hear, right?

You've lived this because you were the classic case where Steve Jobs said, “You're a feature.” And now there aren't just 2 whole companies that do this stuff.

Fortunately, he only said that to Drew, so I think I got the Gates version of that one.

The one thing that is very timeless about Clay would be the thing that transcends floppies or whatever: the incumbent doesn't want to do something that's against its business model. That part is fully timeless.

To your point, it's actually very rare whenever we say these companies disrupted themselves. It's almost never the case that they disrupted themselves. It's the case that they went after a market where they had no actual market share, and it just worked.

Which was the development tools for Microsoft, because there was no business in Microsoft development tools anymore. Nobody was writing Windows programs. So really, it was, “What could we do for the cloud?”

Honestly, it's even a little bit of a silly discussion to have with AI. AI is very disruptive, so we're like, “Okay, well, then it allows startups to work against incumbents.” But even in nondisruptive technologies, startups often have a play against incumbents.

AWS launched a service on us.

And they're onto my open source; they're competing with it. It happens every single time, and I'm always like, “You know what? You'll be fine.”

I can't think of 1 company AWS has ever put out of business by launching a service. You know what? They've all been fine. So in some ways, even in the normal state of business without massive disruption, startups still have a play.

Yeah, it was: “Don't build a SQL Server and compete with Oracle directly. Don't build a word processor.” Maybe those were pretty evergreen for a long time.

And nothing to do with a spreadsheet. There are a few things that—

Yeah, there are some categories where people should call us first. They won't change.

There's a Far Side comic that we should definitely show people. It goes, “Can I have 10 cents?”

“Why do you want 10 cents for your startup?”

“So I can buy a loaf of bread and beat you over the head with it for such a dumb idea.”

It's a good one.

We should have a call-in on this thing, right?

Exactly. Oh, yeah, yeah. “Here's my startup.”

The other thing that I just don't think we've had, at least I don't know of, a modern case study for is this opening up of non-software TAM for software. There aren't even incumbents in the classic sense. The incumbents are really just professional-services categories of work.

For the first time ever, you're packaging up intelligence for a particular domain and workflow. There's no software company you're competing against.

But it could be the vertical company, right? You have to become an ag company.

But then the vertical company will probably also be your customer.

Yeah, but they're also your customer. So it's actually this amazing thing where the people you're probably disrupting on paper are actually the primary users of your technology.

Take advantage. Yes.

And so then there's really no inherent competition until eventually more companies flood that space to pursue that idea.

So this plays out in practice. If you have an AI company that goes after, say, agriculture or construction, they end up realizing that the competitive set is agriculture and construction. The buyer knows how to price things in agriculture and construction.

They end up basically becoming agriculture and construction companies, and then they end up doing exactly what you're saying: selling to agriculture and construction companies because those companies aren't good at that.

There's a whole world. In fact, the earliest PC software was extremely vertical. If you actually look at the TRS-80 catalog from the early 1980s or late 1970s, it would be, “This is crop-rotation software.” Literally: “This is what you should do.” The salesperson for Tandy would show up in Nebraska and sell crop-rotation software.

Then there was, “I run a dentist's office, and this is scheduling for a dentist's office.” What happened—and this is what I think is going to happen—is that these professional-services organizations are going to include some that are computer-savvy. Today, they're really good at using existing tools. They're just going to go, “You know what? We should just build a company.”

I think it's an incredible time if you're committed to not building a software company. Maybe you don't want to do that, or you don't have the team to do that, so you're going to build a real-world company. It's an incredible time to start from scratch with AI as your foundation.

If you wanted to be a new systems integrator, and your whole point was, “We are a systems integrator, but we use Claude Code, Cognition, or Cursor to get the output,” you would have such an advantage over any incumbent because the incumbent isn't going to be able to rebuild that.

I've seen examples of people building new ad agencies because, obviously, if you can do a $1 million ad-video campaign for $5,000, somewhere between those 2 numbers you can charge the customer.

So, there’s this incredible time where you can just be building all new kinds of companies from the ground up, leveraging the breakthroughs that we’ve now seen in AI. That actually happened in the early internet, particularly in the advertising space. There were these digital-native ad agencies that just knew how to use Flash, and they would get bought for $1 billion.

The same thing happened with social, by the way.

Yeah, exactly. What did you think of the survey about how many people use AI every week? I thought this was pretty interesting. What was your knee-jerk response?

Well, it turns out the number of people using it is up to 75% of adults, many times per week. Of course, you just see it when you use Google Search. It’s all self-reported, so you don’t really know. But it was pretty interesting because I pulled a 1999 Pew study on internet usage.

As you would.

As I would.

Basically, in 1999, half the country owned computers.

Yeah.

And they were all online. Even 4 years post-Netscape, it was still half the country. And so you look at that as being slow or being fast.

It was fast back then. I mean, that felt fast, actually—buying a computer.

But still, if you discount the search AI, there’s still activation energy to figure out what this new thing is. Nobody knows how to ask questions to a blank edit control: “Make me smart about something.”

I think the universal adoption of this as a consumer technology, and then bleeding into prosumer, exceeds anything I’ve ever experienced. I think it is—

It will just fundamentally change people’s daily patterns. My sister, not in tech at all—she’s a teacher—was in town, and she was like, “Yeah, I was asking ChatGPT this question.” I had to do a double take. I was like, “Chat?” Then I was like, “Oh, ChatGPT.” That’s what normal people call this, and it’s just completely pervasive as a standard technology.

So, to me, it’s like, okay, we now have the conditions laid for the next phase, which is—and we’ve now seen this for a couple of decades—that consumer adoption goes first. Then it gets pulled into the enterprise because you go to work and you’re like, “Why can’t I ask questions of my enterprise systems the way that I can ask questions of everything else in the world? Why am I not getting that same level of productivity gain?”

And then you again have the kids coming out of college who only know how to do homework with ChatGPT. They come into the workforce and they’re like, “Why would I spend 2 weeks writing this report when I just came out of writing essays in an hour?” Obviously, something has to give on this.

So, this just lays the foundation for why we’re going to see a massive upgrade cycle in the enterprise. Also, to build on your earlier point about distribution, the reason distribution is not the advantage it used to be is because it already exists on 7 billion phones.

Yeah.

And so, at every other platform shift, there was an upside to getting new distribution that didn’t exist before, but you had to overcome that. You had to get the internet to people who didn’t have the internet before. You had to get SaaS to people who didn’t have it. Now everybody has all of the ingredients.

Anytime your business strategy relies on Comcast showing up in a neighborhood for you to get distribution, you’re going to have some problems.

So, this is a very different trend.

Honestly, the last quick point on this is that, for the first time in a very long time, we’re seeing brand effects with an early technology. What I mean by that is, if you look at the major model providers, how much better are they from each other? Maybe a little bit, maybe not. It changes all the time. But you actually see clear leaders if they break out early, just because people learn them.

Household names—people know Midjourney. People know OpenAI. These markets are so big and they’re growing so fast that if you become a leader in your segment, people will just adopt.

But don’t discount the early leaders of search, like Excite and Yahoo. This is just to not discourage people. It’s so early that names you’ve never heard of existed before Google, and that’s going to be really important. It’s never the first people.

When we look at mobile, there were big companies built, like Uber, WhatsApp, Instagram, and TikTok, but the biggest beneficiaries were Facebook and Google. In AI, do we think it will be different—that the biggest companies in the world 10 to 20 years from now will have been created after ChatGPT—or will it be similar?

What’s your time frame?

You know, post-2019, I don’t know.

No, no. How many—10 to 20 years, you said?

Oh, yeah, sure.

Okay, okay.

So, we can’t know if we’re wrong until we do this podcast in 10 years.

I think—this is so boring—but I think it’s going to look like what we saw in something like SaaS or cloud: the incumbents get bigger, but then there are all these new categories that we wouldn’t have been able to predict.

There are lots of $10 billion, $20 billion, $50 billion, and $100 billion companies that also emerge. Over time, those will continue, similar to mobile. Some make the transition and some don’t.

Some will go down on a relative basis because their market wasn’t as ripe for agentic workflows. But I think you can say that if you have a current system of record with a set of workflows on it where agents make sense to make that workflow much more powerful, that’s a good position to be in.

But I bet that if we look back in 10 or 20 years, the vast majority of things agents do won’t relate just to the things we’re currently looking at, because there are so many more fields that are now open. For all of those use cases, I would favor the disruptor or insurgent. In today’s spaces, I would favor the incumbent on the margin. But the markets are so large that you’re going to see growth in all of them.

There’s a key attribute across all of those, which is sort of thought leadership: who is really setting the agenda for what people are talking about? I think that’s the thing that really changes. The incumbents become bigger, but nobody wakes up in the morning wondering what they’re up to.

Nobody starts to wonder, “Well, if they’re going to do it, we need to understand it.” And that’s the shift. You could think of it in the enterprise space or the business space: what do the CIOs wake up thinking? It just becomes, “I use ChatGPT at school,” and there’s nothing you can do about it as a company.

Yeah, I think the more provocative question is: are there any laggards that will use this to get ahead? We’ve seen this in the past, right? Will Cisco do something interesting? Oracle is making some kind of crazy moves. Are we going to see those that missed social—

No, I think we all missed Oracle as an example, right? From 3 years ago to today, you would not have been like, “Oh, definitely next.”

Microsoft had its moment where you didn’t know what the future was, and then they used cloud and Azure to come back. So, this actually is an opportunity for laggards that are behind the curve to come back.

Yeah. To the Cisco point, data centers are sexy. It turns out that we’re just going to be building out lots of AI factories everywhere. So, you’re going to get more scale from parts of the stack that we stopped paying attention to, like Broadcom. Again, people were not—

Hock Tan is going to end up running everything. Everybody looked at Jensen and maybe somebody else.

We’ll table the rest for the next conversation. Thank you so much.