Why AI regulation is every startup's business, with Andreessen Horowitz Martin Casado | TC Disrupt24

You know, it's kind of funny. I mean, this is such an exception industry. You look back and there's very few lessons you can actually draw. Like, as a VC, you speak in aphorisms, but there's very few actual simple one-liners that apply to startups. But there is one I think for me has been kind of a guiding principle, and I wish I would have known it sooner, which is Silicon Valley is so performative, right? There's a lot of like doing startups and doing the right stuff and being part of the culture club and doing networking and all of this stuff, and I don't think any of that really had a serious impact. I think the most important thing is just to do the thing. And there's kind of a corollary to that, which is the following, which is I think every cultural generation, there's this meme of people focusing on how much they—how they can be as productive as possible in a given day. And like, you know, they'd be like, okay, I'm going to wake up at 5:00 and then, you know, I'm going to like whatever, eat something and then run and then work out and then I'm going to work this time and I'm not going to sleep a lot and I'm going to like fast and what—whatever, whatever. And it happens consistently for the last 30 years. But in my experience with startups—I've done two companies—like, startups are so hard and you as a founder are so traumatized that I actually think you need to kind of do the opposite, which is like if you just focus on like staying sane and taking care of yourself, and if that means like sleeping in and eating fast food, like that's totally fine, because at the end of the day, these always take way longer than you expect. And I think the people that actually just focus on their well-being are the ones that survive. And just—just a last coda on that: when I started my company in 2007, there's probably five competing companies that did kind of roughly the same thing. Could you tell them apart? I mean, they're all kind of competent and good. I mean, if you ask me why did we end up winning, I would say honestly, you know, downturns are hard—2008 downturn—and I think a lot of the times these companies just don't hold together. And so if you can survive, like at the end of the day, you'll have a shot at winning. And so I would listen—it's good to hear about all like all the hustle-crazy stuff and like feel free to go ahead and think about that, but if you're doing a startup, you should really, really focus on your mental well-being.

Maybe I should just try and like provide like some context on my position just to kind of level set here, which is—um, okay, fair enough. Okay, listen, a new epoch of technology requires us to think about governance around that technology, right? So like, let's say for like the creation of the internet, it's like a new epoch of technology, and we did have to think about like, okay, so do we change and evolve our laws? And at the time of the internet, like it was a new paradigm. You had new types of risks, right? You had like internet worms, and you had this notion of asymmetry, which meant like if you leaned into it, you were more vulnerable. And we went through all of that discourse, and we decided that there's a few precepts we're going to stick to. Like we thought like open source is very important. We thought that engaging academia was very important. We thought about investing in being leaders was important. So the idea of like transformative technologies and regulation has been this ongoing discourse for decades, right? So the thing with all the AI discourse is it seems to have kind of come out of nowhere, out of whole cloth, where it just doesn't learn from a lot of the discussions that have been had over the last few decades, and they're kind of trying to conjure net-new regulations without drawing from those lessons, right? And so that was the case with California SB 1047, which is the one I was probably way too vocal about. Um, and so maybe just vocal enough. Maybe we'll—we'll—we'll see, we'll see. I mean, like, um, but my conclusion through all of this is if you're going to change the paradigm of how you do regulation—so we should—we all should regulate, regulation is very important—if you're going to change the paradigm, you have to have a notion of marginal risk that's different. Like, how is AI today different than someone using Google? How is AI today different than someone just using the internet? If we have a model for how it's different, you've got some notion of marginal risk, and then you can apply policies that address that marginal risk. But we don't have that today. So I think we're a little bit early before we start to like glom on a bunch of regulation to really understand what we're going to regulate. And again, we've been through this before with the internet, where we did have an understanding of marginal risk, and I think we did a good job on regulation. So anyways, so I think the good news with SB 1047 is that it forced a dialogue, and what a dialogue it was. And I think that academia, who had previously been silent, entered the discussion. And I think one of the conclusions that actually, if you look at, there's a proposal by Dawn Song at Berkeley, Fei-Fei at Stanford, they basically said, listen, the question of marginal risk is a research question, and we need to, from a policy standpoint, invest in research on marginal risk. And once we do that, we'll have a good notion of how to regulate, so we'll have evidence-based regulation. And I think that is absolutely the right conclusion. So I think like VCs like myself should probably step out and let the academics take over. For people in this room, though, who are actively building their companies right now—yeah—they're the ones that are going to be caught in the crosshairs of regulation, and it's all being done with academics or high above them or in Washington or whatever.

Yeah, so I mean, unfortunately, we're at a time where like this is just salient, and like normally when you don't have like some big shift, you can just kind of focus on building or doing kind of whatever you're, you know, you're working on right now. Like this is—this is a very salient topic where people are very focused on like, is AI going to come and is it going to hurt us and whatever. And so I think that you have to become part of the discourse, and I think you have to become part of the dialogue. And in particular, if there are voices that are policy experts that are telling you that this is going to be bad for the industry, you should understand what that is, and then hopefully you can become on the side of those that are actually doing the policy. So pay attention. Pay attention to who's—yeah—get involved. Who is saying what? Yes.

So listen, I actually don't know if Microsoft is a bad actor here, so I'd like—I'd actually kind of pull a little bit away about that. But here's the reality. So, you know, historically, open source has been a way to actually ensure you've got better security, right? Like open source has been a way to ensure like academics and researchers engage. And right now we're in this kind of strange period where a number of very vocal and very famous people in the industry are talking against open source, which has been actually—I can't think historically when this happened before, right? So Vinod Khosla has been very vocal against it. Eric Schmidt has been very vocal against it. And these are people that have lauded open source in the past, right? And so I think there's a question we should all be asking, which is, you know, is this different? Is this a different time, therefore we should regulate open source? I think it's a very valid question. Or is this like it's always been? Like, remember the internet was this kind of scary place, we were talking to strangers, we had new types of worms, and yet we decided open source was the right outcome here. So I think we should ask the question: is it different? And that question is a technical question, it's a marginal risk technical question. And if you talk to the experts, they'll say we don't know the answer to this. And so my view on the open source one is exactly what I said previously, is like we need to be investing in research and academics to answer marginal risk. And before we go and kneecap an industry and kneecap our ability as, for example, a nation to stay ahead, we should answer this question. And again, like I do think—listen, I sometimes I swing a little bit too widely, right? And like I just go to whatever's easy, like, oh, it's, you know, it's like the big companies or whatever. I actually think many of the big companies are good actors in this case. I actually think that there is a subset of people that do believe that there's an existential risk or may have some other guidance that are actually leading this one. But I just—I want to bring you back 20 years. Like this is very new, but this whole thing is very new in the sense that in the past when we had these emergent kind of scary technologies, everybody engaged actively. Like the startup ecosystem did, the academic ecosystem did. I mean, the forefront of EFF came out of like MIT, right? And so there was these kind of movements. And right now there's been relative silence from engineers, from builders, from founders, certainly silence from academics, and it's been largely carried by these kind of minor—like these minority voices that believe in kind of existential risk, and that's really kind of twisted the conversation. And so my goal is not—we shouldn't have regulation. My goal is to say when it comes to open source, is saying this is a valid question. Let's ask the real question and have all the stakeholders in the room. And I think that's what's happening.

I mean, I do think that it's—I think that these things are in flux and dynamic enough. I think it's very important that people here understand what's going on for their technology choices, but almost in the opposite, which is we could have export controls, like that could happen, right? Like they are talking about this. We actually could have liability on models that could impact you in the next two years. Like this was like very legitimately floated. You know, we could have restrictions on what you can do. And I wish I wasn't up here at a point where I had to say, you know, like this is true. But like, I mean, SB 1047 almost got passed. It almost did. And that could have very easily created an opening so somebody could have sued you as a startup. You know, listen, whether valid or not, and whether they would have won or not in the court if something happened based on software that you built. And that's a very new space to be in. And so I would say rather than making technical choices now, I would get very educated on what's going on at the federal level. I'd get very educated on what's going on at the California level. And then I would make decisions based on that.

There's another one that we haven't talked about yet because we talked about open source and marginal risk, which is copyright. You know, like this is probably of all of them going to be the biggest battle, and it's just incredibly unclear right now where it's going to land, right? I mean, if there's a generative AI company, it probably has a lawsuit currently for copyright. And this happens with the emergence of each of these kind of new waves, right? This happened during the early internet days, right? Remember like YouTube, right? You know, like I mean, it was totally loaded with lawsuits. You know, when it got acquired, then Google had to fix it out. So I think, you know, here's what I advise my startups: make sure that you actually have a lawyer on close call that can actually provide guidance on this. Be very, very informed on what legislation is passing through. And often these discussions are not on this existential risk open source stuff, but it's on like understanding the copyright.

Yeah, I think with these super cycles, the problem is there's just so much like confusion that it's hard to know what's actually working and what's not working. Does this make sense? And so I think either—listen, like super cycles don't obviate old cycles, right? The internet didn't get rid of computers. It didn't—computers—they normally actually accelerate it, right? So like just because there's AI doesn't mean you have to do AI to get funded, right? And in fact, I think the companies that end up getting killed in a super cycle, like they don't do the thing that we know that works and they don't do the AI-native thing. They always do something kind of in between, kind of they AI-wash it or they do AI hybrid or whatever. And it happens every cycle. Like do you guys remember in the move to cloud, all the hybrid compute stuff? Like that didn't work. Like hybrid mobile, that didn't work. And so I think if you're in the space, either be very native and like understand why the super cycle exists, or like keep doing the traditional stuff because it tends to be buoyed on that.

So I think it's worth actually saying, so what actually works? So here's what we know about the AI cycle. So the way that I like to talk about it is like the marginal cost of something is going to zero. So if for compute, the marginal cost of computing went to zero, instead of like calculating logarithm tables by hand, you know, a computer did it about 10,000 times faster, you had the compute revolution. The internet, the marginal cost of distribution went to zero. So instead of like sending boxes, you know, go over the internet. Okay, we have the internet. So there's two things that we know work very, very well in the AI world, right? So the marginal cost of content creation is going to zero, and it tends to be content creation where there's no formal notion of correctness, right? So these are like images and videos and music and fiction. Like any company that's in those spaces is doing incredibly well, and not just incredibly well on growth, but I mean like real profitability, efficiency, etc. So there's great, and we're just getting started on that. So that's definitely working. Language reasoning, we know that one-shot language reasoning, the marginal cost is going to zero. That's definitely working, right? So like summarize an email, you know, help me draft something, think through. So that's going to zero. This is always a funny one to talk about, but we actually like—we've never been able to build computers at emote in a way that like humans find convincing. Like that's kind of never been a thing, but we've cracked that problem. So this notion of like emotional connection or companionship is definitely working. Now from an investment standpoint, it's always a little bit dicey because it tends to like maybe butt up against like not-safe-for-work stuff or like, you know, it tends to be very fragmented. But those companies work very, very well, right? And this is a new paradigm. Like this is a new user behavior. We don't know how to think about it well in the companionship stuff right now. Teens are really into it, and pre-teens are really into it.

100%. No, I totally agree. So just like with social networking, just like with the internet, like I mean the internet raised many, many questions of like you're talking to strangers, like you're getting to know them, like etc. And like these are things that we've been grappling—we actually have a policy framework to deal with, and we 100% should apply it here. So the pit is one, like this is a new user behavior, so a lot of companies like they don't know how to build in this space. But if you take a native approach, the companies are working. Then of course, as we all know, code is working quite well, which is, you know, kind of a more an emergent space. And now we're just moving into B2B, and we're a little bit early for that. But like I think it's quite easy to see just like how the internet started as a consumer phenomenon and went to the enterprise, the same thing is happening with AI.

Yes. I mean, new companies really can't do new user behaviors. They just don't know how to do that, right? So I mean, there's always been this question in AI, and this is like the question going back 20 years, which is it's this magical technology. We've been talking—this is before the generative stuff—and we've been talking about it. And clearly like Meta is making a bunch of money from it and Google. So why haven't we seen disruption where like we had a whole wave of AI-native startups? Like that's never happened. And it's always been this huge paradox in investing because we were like AI, AI, AI, but all of the money went to the incumbents, right? And the answer to that has been the economic suck, right? It's like it's a great way of optimizing existing things. Like I can, you know, whatever, I can like target ads 10% better, right? I can convert you 20% more. But it's like then these kind of marginal gains that take a lot of money to do. So if you're a monopolistic incumbent, of course you can take advantage of that, but as a startup you can't. Probably the number one thing of this generative wave is like that is different. Like the economics are phenomenal, right? Like if I were to like create an image of myself by drawing it, it would cost what, a hundred bucks in an hour. Like the inference cost is a fraction of a penny. So you have a dislocation in economics, and it's a new behavior. We just didn't have this behavior. And so incumbents just don't know how to go in like entirely new areas with economic dislocation. And this is why we're seeing companies that you would naturally think would be like in the domain of a Microsoft or a Meta or an OpenAI being quite successful right now. And so this is the wave if you want to go against the incumbent now, because you have all of the foundational properties in order to do that, which we just didn't have in the past.

This is a great question. So I think a few things we should all agree on. There is a robust regulatory regime that exists in place today that's been developed over 30 years. Two—well, there is—well, there is a regulatory regime. Two, whatever we do, if we got it wrong in social media, you can't fix it by putting it on AI, right? And like by the way, one of the I think biggest kind of logical fallacies that happens from like the AI regulation people, they're like, oh, we got it wrong in like social, therefore we'll get it right in AI. I mean, which is—it's a nonsensical statement. Let's go fix it in social. So I think 100% we should have a broad dialogue that includes things like social media. And then we should say, if we broke—we didn't get it right in social media, let's get it right. And then B, if it's different, let's understand how it's different so we know we're regulating the right things. So I'll give you a very specific example where this line of reasoning I think is wrong, which is if you don't know where the marginal risk is—which we don't—and you create a policy based on that, you're probably not going to regulate what you want to regulate because it's going to be the wrong thing, which we actually saw with the executive order. And if you're worried about social networking, you've actually taken a lot of energy and effort and discussion away from that problem. So my answer to you is, if there's a problem with social media, let's focus on it. Let's include that as part of the AI problem, and then let's kind of work towards all of these. And if there's a marginal difference, let's figure out what that is and then apply policy to it.