a16z 谈保护 Little Tech：与 Matt Perault 讨论技术乐观主义 AI 政策议程

Matt Perault, head of AI policy at a16z, welcome to the Cognitive Revolution.

Thanks so much for having me on.

I'm excited for this conversation. Obviously, AI policy is a hot topic, and I've got a bunch of different angles that I want to get into with you. But maybe, just for starters, because I think we'll actually have a lot in common here, I noticed in my prep that you are a fellow at the Abundance Institute. I wanted to give you the floor for a minute to share your dreams for the AI future of abundance and get a sense for what you envision that looking like.

My primary affiliation is head of AI policy at Andreessen Horowitz, but I have a couple of side affiliations that I started as an academic. Before this, I was at UNC Chapel Hill, running a center on technology policy. As a fellow at the Abundance Institute, that's something I've been able to continue. I'm also a fellow at the center that I used to direct, which is now at NYU.

Those affiliations are awesome because they enable me to continue working with communities that I really enjoyed working with in the past. The Abundance Institute is a great organization. Christopher Koopman is the lead, and it's really focused on a regulatory agenda that can help unlock abundance. I've learned a lot from that group.

They have a great group of policy professionals and people who focus on economic policy. They do a lot in energy, which has been interesting and obviously relevant to my current job at Andreessen Horowitz. I've really enjoyed maintaining a close relationship with that group.

1. Policy Enables Abundance

The focus for me, in response to the question you ask, is going to be a little bit disappointing for you. People always ask tech people, regardless of where they sit in tech, to forecast the future and describe what the future of technology looks like. I would say that's something that I've never been particularly good at.

I'm not an engineer or a computer scientist, so the people who are really building the tools tend to sit on a different side of the house. I previously worked at then-Facebook, now Meta. I was on the policy team, not on the product side. At Andreessen Horowitz, we've got a lot of people who really understand the technology deeply.

That's obviously an important part of my job, to do that as well as possible, but my focus is on the public policy side. My orientation to the question is really: What is the right policy agenda to enable other people to unlock abundance? How do you think about ensuring that the people who can create the tools are able to do that?

That doesn't mean total deregulation. It doesn't mean being able to do whatever an engineer wants to do, but it does mean trying to avoid regulatory models where the cost of those models, particularly for startups and little tech companies, outweighs the benefits. That's really the day-to-day focus of my work.

I hear you on all that. You have to have some dreams, though, right? For me, the future is medical advances, equal access for everybody around the world to an AI doctor, and maybe one day a robot in my home that's doing my dishes and picking up my kids' messes.

Maybe one day people don't have to work if they don't want to, or if they don't have work that they find intrinsically motivating or fulfilling.

I don't know. I'm always interested to hear, especially because you're so invested in this. I would love to hear what animates you.

2. AI’s Human Benefits

All those things sound cool. Both my parents and my sister are mental health professionals, and when you talk about AI in medical care, I think about breaking down barriers to access to mental health treatment. That's been something I've had conversations about with my parents since I was really little. They were talking to me about the value of having mental health support in your life, and that's hard for people to do for a bunch of different reasons: It can be cost-prohibitive, it can be hard to identify a mental health care professional, and it can carry a stigma.

I think some of the advances there are really exciting and interesting. Obviously, it's important that it be done in a way that's consistent with providing a high-quality level of care. But I think breaking down access barriers is a really key thing. As someone who spends a lot of time writing and reading, I've been excited about the benefits that AI can bring to the analytical process and to the writing process.

It's a scary one because I think it's coming after my business model. My ability to have some level of writing competence has kind of always been a part of my skill set that has been important in my academic life, as a student and then as a professor. It's been important to my professional life, too. The ability of AI tools to develop thoughtful, clear prose is something that sort of corrodes that part of my skill set. You have to be a differentiator on top of that.

Another way to say it is that, at one point when I was working at Facebook, I said to a boss, “I understand my job now is to deliver a bad first draft to you.” His response was sort of funny. He was like, “Yeah, that's great. That's great. That's great.” And then he was like, “But every now and then, could you develop a good first draft?”

To the extent that the job some people in these fields have is to produce that first draft that enables people to provide feedback and critique and then massage it and hopefully turn it into a really good final draft, now we have tools that are able to, I think, do bad first drafts and maybe at some point do very good first drafts. So I think the fact that it's coming after a thing that's been kind of important to my professional life is a really interesting, exciting opportunity and also somewhat of a scary one. But it's one that I think is exciting because I think it will, again, play an important role in breaking down access barriers.

Yeah. I feel that very much myself. Coding and even podcasting—I mean, NotebookLM—

Yeah.

—I feel like in many cases it competes pretty effectively with what I'm able to do. So, yeah, it's coming for all of us.

What is the balance for you between fear and excitement as motivators? As a writer, I think if there's something exciting, usually it's a person. Now it's a machine and people who are chomping at your heels, and I think that can result in being more creative and being more focused and devoted to parts of your craft in certain ways. It's also scary to feel like people are potentially coming for your work if you don't perform particularly well. I haven't thought about it as a podcaster. How do you think about the fear-versus-excitement balance?

3. The Existential AI Risk

I describe myself as classically ambivalent. The dictionary definition of ambivalent is contrasting or even contradictory, but strong feelings. That's always been very natural to me with AI. The upside is—

Hmm.

—legitimately thrilling, and when new models and new products come out, I'm always rushing to go try them and get hands-on experience with them. This morning, it's been Claude for coding up an app and Operator for doing some web tasks. I do genuinely feel a thrill from these—

Yeah.

—new product experiences. I'm less concerned, honestly, about it coming for my work—not because it isn't coming for my work, but because I worry about even bigger things, like: What does gain-of-function AI research look like, and do we really have the wherewithal to control this phenomenon in the macro sense?

Yeah.

I'm less worried about meaning or jobs or what we're going to do with our time. I would say I'm an optimist when it comes to people being able to figure out how to use their time and have a good life, given—

Yeah.

—resources, space, and opportunity to do that. But I'm a little concerned about just, boy, we don't really know how these things work. We don't really know why they do what they do, and they do seem to be demonstrating more and more of these sort of negatively thrilling behaviors.

As much as it's thrilling to see these things go about these tasks and learn to overcome obstacles and whatever, I also can't help but see that, paired with deceptive behaviors and scheming behaviors that we're seeing emerge, it's like, geez, I don't know. The dynamics of that just get to be really weird and really hard to predict.

But really, the only thing I don't expect is a business-as-usual future. I'm obviously a small player in history. I sometimes call myself the Forrest Gump of AI because I always find myself in these notable scenes, but usually as an extra. I just try to make whatever little nudge I can in a positive direction, and I kind of feel like, whatever those odds are, it's less for me about handicapping and more about whether I can put a little bit of shoulder into shifting the odds of a radically positive future just a bit. Hopefully we can all make a little impact like that.

That seems right to me, and I think that feels like more of the direction of travel. You probably track the macro conversation more closely than I do. I had initially thought, in the post-ChatGPT-release moment, that there was a lot of focus—not on the immediate, not on what was exactly in front of you, but on the infinite future.

It was about robots taking over the world, what happens in a world where people have less autonomy, and what that means philosophically—those big-picture things. Over time, it seems like, as the technology has improved and more models have become available to more people, people have gotten a better sense of how to use them. Our digital AI literacy, I think, has improved. People compare notes with their friends about how they use it, and then they integrate some of those use cases into their own lives.

The gaze has turned much closer to people, exactly as you're describing at the end: more about marginal value creation as opposed to, “I'm creating a robot that is going to take over my life.” And my sense, at least, is that when the focus is on that marginal value creation, it has connected people much more closely to a value proposition that's much less scary.

The question is less robots taking over the world and more, “I need ideas for recipes,” or “How do I deal with a kid who's not going to sleep?” Or, “I'm looking for activities to do on a family vacation in this place.” Or whatever it is. Or, “I'm struggling with how to get this piece off the ground. How can I structure the ideas?”

I've been using AI sometimes because I don't like writing conclusions. AI writes very good conclusions, or at least gives you a bad first draft of a conclusion. It can be helpful for that kind of use case. That, to me, has shifted the conversation and then also, I think, shifted importantly, given the nature of my work, how people think about public policy.

You want public policy to help us avoid harms, like long-term catastrophic harms. But I think you also want policy to give you room to unlock and experience the value. And I think thinking through what the right models are changes as more people have a closer connection to the value proposition.

4. Regulating Harmful Use

Maybe that's a perfect segue to jumping into some policy questions. I think folks will know a16z as a prolific, super-successful investor led by techno-optimists and thought leaders. A huge theme of everything that I have read of your work and of the firm more broadly is making sure that there's room for little tech, as opposed to big tech, to experiment, innovate, bring things to market, develop new use cases, and so on.

With all that in mind, you said it's not about opposing all regulation, but what is the first-order articulation of the policy agenda that you would advocate for today?

Yeah, so I can frame what's baked into your question even slightly more explicitly. A lot of people think that we are extremely focused on being as deregulatory as possible. They think that I would have done my job effectively if we saw no AI policy. That's not the case. And it's not the case, I think, because it's connected in a really important way to the economics of the firm.

The life cycle of an Andreessen Horowitz fund is 10 years. That's actually significantly longer than the life cycle of a tech investment, or of a vesting cycle at a public tech company.

When I was at Facebook, it was a 4-year vesting cycle. I think it's a longer time horizon than you would typically see in private equity. We're not looking to juice stock prices tomorrow or produce massive returns tomorrow. We're looking to create a healthy, interesting, exciting, vibrant, abundance-oriented, big-dream-oriented—to get back to your first question—view of the world and a reality in AI technology over a longer period of time.

I think that's important for a bunch of reasons. One additional way that you could think about it is if people were really into AI tomorrow, really, really excited about it, and there was an explosion of interest in AI and an explosion of AI products, and then the market cratered a year later because of lots of examples of harms in our society, we're not going to get the economic returns that we're aiming to get. We're trying to build healthy, long-run, strong companies over a period of time.

When you're trying to do that, that means that you need to allow startups to get off the ground. So you were getting at this in your question: we're not focused on the health of big tech companies. Sometimes big tech interests overlap with ours and are aligned with ours. Sometimes they're not. That's really beside the point.

Our focus is on little tech companies. How do they have room to grow? But we have to have an ecosystem that's also providing a set of tools that people like to use and feel safe and secure when they're using them.

The core of our policy agenda is focused on trying to encourage policymakers to regulate harmful use, not to regulate AI development. In our view, that leaves room for little tech companies to build and grow. It's typically how technology regulation has historically been approached.

If you think about software development, there's not a lot of regulatory intervention in the building of software. But if you create software that's harmful for the world, depending on the specifics of what it is, in most cases there's going to be law that can take account of that harm and prosecute you and deter you from future activity where you're building something that's harmful.

So that's our orientation: to try to leave the process of development available to little tech and give them room to build. But, importantly, the second component is regulating harmful use.

Yeah. So let's dig into those, maybe one at a time. On the development side, obviously development can mean many things, right? It can mean me sitting down and fine-tuning a model and just tweaking it a little bit for my—

Yeah.

—purposes and baking it into an app. On the very high end, it means pushing the frontier of AI capabilities. While I'm definitely not a big analogy guy, I do sometimes think of this as being akin to biological research.

When I think of the canonical example, without coming down firmly on exactly the origin of COVID—

Yeah.

—there are these institutes out there in Wuhan and elsewhere that are doing this sort of research, where they're asking, "Well, geez, what would happen if we made a virus do this?" I think they have good intentions, to be clear. I understand the idea is that we're going to make these things so we can figure out how to treat them, so that if it ever does happen, we'll be prepared.

But we do have this history of lab leaks. So I think at the very frontier, there does seem to be a very material risk that's pretty analogous to the biological research example, where you don't necessarily have to deploy it. You don't even have to intend to deploy it. If you're developing something that's sufficiently powerful, there is this risk that it might get out of control.

I wonder how you feel about that risk, which I think for most—

Yeah.

—people in the AI safety community, it's honestly the number 1 concern.

Yeah, so I don't know enough about the regulation of biological research to know whether it's a good analogy or not. It feels, at least with my superficial understanding, to be a good one. We give broad latitude, I think, to researchers to research, including things that could be very harmful if they're released in a way that violates the law.

The research process itself isn't really good or bad. Typically, I think we want to allow researchers a lot of room to explore and not be super prescriptive. There are any number of different things that researchers can do that are problematic.

One that's closer to something I've worked on is that I was at Facebook when the Cambridge Analytica incident occurred. That was data going from Facebook to an academic researcher, right? That transfer, at least under Facebook's terms, was legitimate, and then the researcher misused the data. The initial transfer to a researcher was consistent with Facebook's terms.

I think in some ways—maybe most ways—you would think that that initial transfer is desirable. We want academic researchers. Typically, the research community is asking for more access to data, not less. That is probably a positive thing.

The abuse—the misuse of data once you have it in your control—is really a problem, and we would want to make sure that we have stringent rules in place to penalize misuse. Again, in your COVID analogy, it's not okay to develop a virus that could cause a global pandemic and then release it into the world so as to cause a global pandemic.

I don't know the legal reasoning, but my assumption is that that is unlawful and therefore should be penalized as such. The research, I think, is a different question.

Again, mapping it back onto AI development, the development itself is just science. It's math. The concern that our firm has had is that when you look to criminalize the math, when you look to make it harder for scientists to do science, that just has the impact of slowing development and actually doesn't really do much to address potential harmful uses.

If we're concerned about harms, we should try to make sure that they're punished. I think the concern that I have, and maybe it's why so many people think we're purely deregulatory, is that people spend a lot of time hearing, thinking, and absorbing, "Don't regulate development."

Then, when we say, "And regulate harmful use," it's like the ears close up. I think maybe the idea is that it doesn't feel like there's a clear pathway there. Maybe the view is that it's just a diversionary tactic: we're trying to get people away from regulating AI development so that they do nothing.

But actually, we think regulating harmful use gives lawmakers and enforcers a lot to do. There's a lot there that's really significant and meaningful.

When I was in law school, I spent a summer in the Criminal Section of the Civil Rights Division at the Justice Department, and that was an interesting experience for a bunch of reasons. Primarily, people go there because they want to litigate cases, I think.

When you're there every day over the course of a summer, what you see is that, especially as an intern, you're not traveling around the country going to different courtrooms. You're in the offices in the Civil Rights Division, watching what the attorneys are doing when they're not in court.

What they're doing is building cases, and that's really hard to do. Cases don't just come with clear fact patterns where you can really understand exactly what's going on and where a violation occurs. They don't come to you with that clarity. You have to build the case, and you have to identify the violation.

I think the same thing will be true in AI. We're not going to know when AI is used in a violation of antidiscrimination law. I should say it this way: it's not necessarily going to be obvious. You will need to build cases.

You need to identify the harm that's occurring and the AI tool being used for that harm. That, I think, requires work and investment and probably lawmaking in some cases.

You need to have enough people to do the enforcement. Those people need to have the resources they need in order to do the enforcement. They need to have the technical understanding of the technology in order to understand when laws are being violated.

There may be cases where existing law isn't sufficient to account for harms created by the use of AI. In those cases, we may need to think about tweaks to existing law to make sure that you can take those harms into account.

So when we say, "Don't regulate development," that means something. And when we say, "Regulate harmful use," we consider that second component to be a really active part of the policy agenda.

5. Frontier Safety Thresholds

Let me do one more beat on the development, and then we'll circle back to the—

Yeah.

—harmful use. On the development side, obviously a big pattern of bills that have been put forward—which I think the proponents would say is their effort to meet folks like you in the middle on this, or even be quite compromising—and also I would...

I think they would say it reflects a genuine shared concern: They don’t want to—

Yeah.

Quash frontier research, certainly unintentionally. Yeah, a pattern has kind of been emerging recently, and this is now happening again in New York with a bill that I just did an episode on, the RAISE Act, which I’m sure you’re tracking.

Basically, the idea is: We’ll have some threshold, and I know you have thoughts on thresholds, and if you are above that threshold—and typically they’re trying to design the thresholds so that high single-digit, maybe low double-digit numbers of companies would be caught by them—then you have to develop a safety plan, publish that safety plan, and share it with a regulator, maybe publish it publicly. Those little details vary. Then you have to follow it, and there’s some variance on the provisions, like, do you have to have an audit, or do you just self-report, or whatever exactly, right?

Yeah.

My impression is that you don’t favor those, but it does seem like that is pretty consistent with the general principle of giving people pretty wide latitude, because most of these proposals, at least the ones that seem to be getting the most traction, are like the companies get to define their own safety plan, and they’re basically grading their own homework for the most part.

We just want to make sure, speaking on behalf of the sponsors of the bills, that they’re actually doing it in a serious way, because we do think this stuff is pretty serious. We’re not going to tell every biologist exactly what experiments they can and can’t run, but there is a biosafety level 4-type concept where, if you’re—

Yeah.

—not implementing certain standards, then we look very unkindly on that. So how do you feel about the—we can bring in thresholds here—but also just this kind of very top-tier, grade-your-own-homework, but we really want to at least know that you’re doing it kind of vibe?

Yeah. So, just as a threshold matter, you agree that those regulations are targeting AI development, right, not harmful use?

Yeah, I think so. I would say “frontier development” is usually the phrase that I—

So, there are a bunch of ways to get into this. One, just at the top-tier level, is: What is the problem with targeting harms? If what we’re concerned about are AI harms, why don’t we put the emphasis on ensuring that we can address harms when they occur?

In terms of the development side, which I know is what you want to focus on, I think there are a couple of different aspects of it. One is who the bills cover, and the second is, for the people and organizations they cover, what does it ask you to do, and is that, as you put it, reasonable? I wouldn’t use the term “reasonable.” I think I would say: Is that going to encourage the kind of AI ecosystem that we want to see?

On the who side, I think you actually, even in your question, said the thing that, as you framed it, sounds like a compromise. It sounds totally reasonable, and it’s the thing we fear most, which is that this is just going to apply to a handful of companies—5, 10, 15 companies. It’s just going to apply to a small number.

That’s the thing that we’re worried about, because the other way to say it is: This is going to be a concentrated market, and we’re going to have a regulatory model that is not only going to be okay with a concentrated market. We’re actually going to have an approach to regulation that is going to create concentration. You can only build at the frontier—there are only going to be a small number of companies that are able to build at the frontier. That’s not what we want.

We want to ensure that startups can build at the frontier, and that if you’re a founder and an engineer, a founder and 5 engineers, or a founder and 10 engineers, those kinds of companies are able to have the ambition of building at the frontier. When you said in your opening, “What’s the dream of abundance?” I said, in a roundabout way, that the way I think about that is: What is the regulatory model that allows smart, creative, ambitious, skilled entrepreneurs to build great businesses? This is what I’m talking about—the capacity, the ability to build at the frontier.

And it’s just going to be 5 to 10 companies. I thought, before I started at Andreessen Horowitz, that it was an exaggeration—like, it was a story that I’d heard people tell, but I didn’t think it was a real thing. And now it is. I mean, I’ve heard it a couple of times. I heard it on your previous podcast. I heard it in a conversation with a member of the European Union, someone who’s working with them on their AI regulation, so I don’t think it’s an exaggeration.

I think it’s a widely held view that, look, this kind of development is just going to be a small number of companies. For a lot of people, they’ll hear that as very reasonable, and I think it’s really important, whether or not you agree with us, to see that as very strongly opposed to the vision that we would have of what a healthy AI ecosystem would look like.

In some ways, I’m very sympathetic to the idea that concentration of power generally does not appeal to me, so I do find this, for what it’s worth, to be the case among many people who would call themselves AI safety-minded people. Broadly speaking, they are techno-optimist libertarians throughout their lives, until this one thing.

I think the thing that makes it different for people, which answers the question around, “What’s wrong with just regulating harms?” is that there may be some things where, by the time the harm is done, it’s just too late, right? The COVID pandemic is an instance of that. Again, I don’t take a position on exactly where COVID came from, but if it did come from a lab and now we have 10-plus million people dead globally, it’s really hard to go to that lab and say, “Now you have to pay a fine,” and act like that is an adequate response. I think it’s a perfect example.

How many constraints have we put on biological research in the wake of COVID?

Well, as far as I know, gain-of-function research continues, right? I mean, I don’t think any actual lines of research have really been shut down. Have they?

That’s what I mean: What if the response to that, instead of focusing on all the problematic incentives and the geopolitical issues and stuff related to how COVID was released into the world, was, “We’re going to make it really hard for every single researcher working on biological research to do their research”?

You’d hope to be more targeted, and that does obviously connect to the number of companies. But there is something to be said, I think, for—I’d love to fix the world. I’d love to have more wastewater monitoring, and I’d love to have air scrubbers installed in all the schools, and there are so many things that seem like honestly pretty obvious hardening. Certainly, this is a theme in the AI safety discourse too: What can we do to harden society against these potential future assaults?

But there is also the notion of choke points, and you can say, “Okay, there’s one thing that seems to be a really bad problem, and that is if you create new viruses that don’t exist in nature that have super-transmissibility or super-lethality. That seems really bad, or at least potentially really bad.”

And it doesn’t seem crazy to me for society to come in and say—and I don’t think we have, but I would support it if society got together and said, “Don’t create—

Then I think that we haven’t. We lived through the pandemic. We didn’t create civil or criminal liability for what you’re describing. I’m not a biological researcher, so I have no idea. I don’t have a sense of why we didn’t do that or why we, as a society, might think that risk is one worth bearing.

But my guess is that kind of research on some of the most virulent, potentially problematic viruses is the same thing that will prevent those viruses from spreading over time. You have to do research on them, I assume, in order to figure out how to combat them. That doesn’t mean it’s okay to release them into the world.

I think there are lots of reasonable things that you could do in that situation to try to ensure that there’s a deterrent for harmful use, but we don’t take that kind of research off the table because of the upside.

And I think that, when we were talking about AI use cases, there are just an enormous number of positive use cases, from the mundane to the revolutionary. The mundane might be that it helps me write a conclusion, but dramatically lowering barriers to medical access or mental health treatment, or making it much easier for more developers to build more products that are valuable to people—whatever the explosively positive use case is—I think that, just like with biological research, the reason we preserve a fair amount of ability to do the research itself is because of those possibilities of really positive, explosive use cases.

And I think if what we want is abundance—and I think there are a lot of people who are skeptical of abundance—but if that is a thing that we want, then we need to have a regulatory environment that is conducive to it. Putting significant constraints, handcuffs on the development process, I think, is not going to get us there. And there are a few different components of it that feel to me to be particularly off base. One is just going to apply to a couple of companies. That is another way; for us, that means we can't invest in startups to grow. I also think it means that's one part of the innovation angle.

The other part of the innovation angle is that one of the reasons our laws disfavor monopoly is that monopolists tend not to innovate rapidly. And so if you have a super-concentrated market, it's likely that you're going to see fewer of those most interesting use cases be pushed and be pushed aggressively.

The other thing, which I think is really important, is that it would be a different story if we said, “Okay, we're going to make it a lot harder to build AI models, but the impact of that is going to be that we avoid a significant number of safety incidents, and we really can ensure that we are going to dramatically reduce harm in the world.” And I don't think that's the case. I don't think that publishing a document with safety protocols means that you're not going to have safety incidents. I think there are probably some number of safety incidents that wouldn't occur that might otherwise occur, but I think it's conceivable that there are going to be very few of those, that it actually has no direct correlation to the actual safety or the performance of your product.

And if that's the case, if there's pretty limited efficacy in what you're doing, then you're simply burdening development without a lot of upside. And that again is why we think the idea isn't “don't regulate.” The idea is to put resources and emphasis on the harmful-use side, both because it leaves room for development, but also because it actually, I think, would be more effective in addressing potential misuse.

6. Transparency and Model Facts

I do agree with you that the proposals that have been put forward of this sort—that everybody, not everybody, of course, but some number of companies above a certain threshold, where we could debate exactly how big that number might grow over time, have to develop a safety plan, follow it, and disclose safety incidents—are not that strong a barrier against certain possible bad things happening. I think they're more meant to try to make sure that we are not flying blind. And that is honestly one of my bigger worries.

I wonder if a reframing of some of these requirements around transparency is appealing at all. One of my worries is that—and I think we are seeing signs of this from OpenAI and others, and frankly Anthropic, who has at least in some corners of the safety community been the darling frontier developer—they seem to be going for some sort of self-improving feedback loop, some sort of recursive dynamic where the AI is going to get so good at coding and ML research that it's going to do the improvement. And literally somebody from Anthropic, as they launched Claude 4, said, “We want Claude to take over all the ML research so we can all go to the beach.” And people were like, “Well, you can't put it much plainer than that,” right? I do wonder what starts to happen behind those closed doors.

And again, this is all very speculative right now because we're just in unprecedented territory all over. But I also think Daniel Cocotello, who is one of the lead authors on the AI 2027 projection, or forecast, or scenario—I guess scenario might be the best word for it—says that we should expect that right now we have a very narrow gap between frontier AI capabilities and what is deployed for you and me to use.

But he expects that gap to grow much wider over the next couple of years as the companies realize—and I think they may already know it—that they're in this sort of race for supremacy: who can create the most powerful AI model first, one that dominates the market, dominates their rivals, what have you. They just aren't going to want to show that to anyone. And right now, there's not really any rule that would require them to do that.

So I wonder if you would support something like transparency requirements, even if only around what capabilities the AIs have. OpenAI, Anthropic, Google—you have to tell us what it is that your AIs can do, so we at least know what they are capable of in your labs. What do you think about something like that?

Yeah. So we published a blog that I would encourage anyone who's interested to read because it's a little hard to run through it quickly. But the basic idea is that a lot of the transparency proposals, we think, are probably unlawful. They probably violate the First Amendment, and also aren't particularly useful for consumers.

So I don't think some of the things that have been proposed—some of the transparency mandates around certain kinds of assessments of safety risk—are actually things that will provide consumers with information that's useful for them, that's really informative and shapes how they interact with the model or which models they choose to interact with. In our view, the way to get at a disclosure regime that's lawful, useful for consumers, and also not unduly burdensome for little tech is something that we call AI Model Facts. It would basically be a basic fact sheet that would run through different information that would be helpful for users.

This is sort of a fun project to work on because it was interesting to see, in current disclosure-mandate proposals, what might map onto this. If the criteria are that it's not unduly burdensome for startups, useful, and lawful, what are the elements from current proposals that would meet that criteria? There weren't that many, actually, that I'd say would meet that criteria. The kind of stuff that we thought of were things like a knowledge-cutoff date, which I think is actually extremely useful for consumers.

So if you ask an LLM to give you information on who won the basketball game last night, that's not in its training set. Having a sense of exactly when that cutoff date is will enable you, first, to ask better questions that are actually within the corpus of data. Second, if you were to ask it who won the basketball game last night and it gave you an answer, the likelihood that that's a hallucination or information that you shouldn't trust should be high. So I think that's the kind of information that actually is useful for consumers.

The courts have been pretty clear that it's very hard to survive legal scrutiny if you require information that's either controversial, not factual, or burdensome for companies. And so I think some of the types of information that are typically seen in transparency proposals, like speculating about potential long-run harm or forecasting potential safety risks, are going to be really challenging for those things to survive First Amendment scrutiny.

One more quick bit on the thresholds. I again share your concern about regulatory capture and the sort of structural oligopoly of various markets that we could find ourselves in via mechanisms like that. I think the people who have developed these proposals around thresholds have come to them because they see the scaling-law phenomenon as a good indicator that we're kind of headed there regardless, and certainly it seems like we basically are there today, right?

I do a live-players analysis every so often where I'll ask people, “Who do you think should count as the top-tier AI companies in the world today?” And the list is sometimes as short as 3, and never longer than 10. That seems to be, at least so far, a structural reality of just the kinds of inputs that these things require.

Yes, the costs are definitely coming down, and so people have proposed naturally escalating thresholds, whether it be a certain amount of money or even revenue, which I know you're a little bit more favorable toward than other concepts. The notion that there are going to be a lot of companies at the frontier of AI seems to run contrary to what increasingly seems like a natural law: that there is just sort of a scaling-law reality. As you know, as a venture investor, power laws rule everything around us. So is it so unrealistic to think that it could stay a small number of companies caught by these thresholds?

Well, first of all, I think if what you're trying to do is really just capture large companies that are able to bear compliance costs, revenue thresholds are clearly the policy vehicle that gets you there. And the reason is, once you're taking hundreds of millions of dollars in the door, we can figure out what the threshold should be. I've seen proposals with $100 billion thresholds. I've seen higher revenue thresholds. I've seen lower ones.

But whatever it is, at some point of taking revenue in the door, you are able to devote some percentage of that to compliance costs, right? It's hard to argue that if your annual revenue is $500 billion, you can't devote some percentage of that to whatever kinds of exercises and transparency disclosures regulators want you to do.

Compute and training cost thresholds are different because we think startups can build at the frontier. They are able to build models with significant compute power. I had a conversation with someone on our technical team about a $100 million training-cost threshold because I was starting to see that in more pieces of legislation. I said, “Is that a good way to carve out little tech companies?” And he said—and I trust him on this; I don’t have the technical chops to critique it—that every model costs $100 million to build.

So I think those thresholds are not going to be successful in separating little tech from big tech, and the technology will outpace regulators’ ability to update them over time. One thing I noticed in the RAISE Act, as part of the discussion that you had about it, is that there are a couple of different ways you can hit the large-developer threshold. You have to hit a $100 million training-cost cap, but I don’t think there’s anything that says that couldn’t be cumulative, which means that if you built a large number of $25 billion models, you’d be captured.

Right now, that might seem very far in the future for a startup, but you might have a startup building those models a couple of times a year. I don’t know exactly what the highest rate of model development might be, but if it’s more than 1 a year over a period of time, at some point that would mean every startup is a large developer, right? Those kinds of thresholds are focused on the development layer, which is where we think regulation is much more fraught. If it’s cost or compute, that’s focused on the science of building the tool and really doesn’t have anything to do with a company’s ability to handle the kind of regulatory complexity that some of these regulatory models require.

Another way to get at your question, I think, is that you’re right there are these natural barriers to entry in AI, like the cost of compute. We actually have some ideas about different ways to provide broader compute resources to more entities. We have some ideas about access to data as well, trying to ensure that more startups have more data they can train on.

The question for policymakers is: Do you want to add an additional significant hurdle? Do you want regulatory barriers to make it even harder? If that’s the case, then we end up in the world that I think you’re describing, which is that we have a small number of developers who can build at the frontier. Everyone else can have the rest of the ecosystem—less capable models, not building at the frontier—but we essentially have a regulated monopoly that looks more like what we’ve had in telecommunications, for instance.

You’d have a very small number of companies with a stringent level of regulatory oversight. At one point, there was even an explicit government grant of the monopoly to 1 telecommunications provider. I think there are people who would say that’s the right model. That’s exactly what we should do. That’s exactly the direction of travel we should pursue.

From our standpoint, that is a level of monopolization that we would think of as problematic. It suggests a lower level of innovation than we would think is desirable, and reduces the ability of startups to compete at the frontier in a way that we think is going to unlock a lot of value for people.

7. Flexible AI Regulation

One thought I had earlier on the 10-year time horizon of an a16z fund is that not only is that longer than vesting schedules, it’s also longer than most people’s AGI timelines these days. I wonder—something I’ve pitched many times and it never seems to get any traction, but there is a version of it right now that’s under consideration in California, if you squint at it, with SB 813—is whether there’s some ability to move in and out of different regulatory regimes.

The simplest possible thing might be to have a sunset clause on a lot of these things. For the next 3 to 5 years or whatever, we think we’re in a critical period, so let’s put this in place, but then have it sunset at 3 years. If it turns out some of this stuff was hype or we got certain thresholds wrong, then it can just disappear.

With SB 813, there’s a similar concept where a trade can be made between an AI company that wants to have a liability shield. If they opt into a certain regulatory scheme that’s approved by the state but administered by a private organization, then they can have that liability shield. But they could also potentially opt out of it at a future point in time if it’s not a trade that’s working for them anymore for whatever reason.

So I have 2 questions there. First, what do you think about SB 813 specifically?

Yeah.

Do you have any other creative ideas for ways that we can put some measures in place—whether they’re guardrails, transparency, or whatever—that don’t lock us in long term? Nobody wants the GDPR of AI, but some of us are definitely nervous about not doing anything.

Quite a lot of lawmakers seem to want a GDPR of AI because I think the proposals they’re moving forward with suggest that. One interesting thing is that I saw a story—and I don’t know how accurate it is—that the EU is considering pausing implementation of the EU AI Act. So the EU is even having concerns about its own AI-oriented version of GDPR and starting to pull back.

The only state that’s actually enacted comprehensive legislation to govern AI is Colorado, and the governor of Colorado came out in support of the federal AI moratorium, which would have made it impossible for him to essentially enforce his own law. So policymakers who have actually enacted these approaches are expressing some rightful concern about how they’ll be implemented in practice.

Again, that doesn’t mean, from our standpoint, that policymakers shouldn’t take action. We think lawmakers should focus on actively regulating harmful use. We think there’s a role for states and the federal government to play in that. Regulating harmful use is actually pretty consistent with how states historically, and according to the Constitution, have gone about lawmaking.

They can’t unduly burden interstate commerce, but they can police harmful conduct within their jurisdictions, and some areas of law are disproportionately occupied by state governments. Most of criminal law, for example, is at the state level. So if you’re looking to tighten criminal law to address harmful criminal use of AI, most of that is really for state lawmakers to do.

There’s also a lot for states and the federal government to do on investments in workforce development and AI literacy. For transparency, we’ve talked a little bit about an AI model-facts framework. I think there’s a ton of stuff that lawmakers and enforcers can do to try to arm us for a world where AI is more prevalent.

That doesn’t mean there’s an infinite road. The way that you described it, I think of it as: Could we just do some things for a short period of time, and if they’re not the right approach, we can pivot? I like that kind of experimentation generally, but depending on the model, if it’s a really stringent regulatory model, it’s kind of like saying, “We want to run this race. We want to run it as quickly as possible, but you’re going to have to wear a 20-pound backpack for the first mile.”

I don’t think it’s cost-free, even if you do it only for the first mile, to require someone to wear a heavy weight. Different people have different views of how important it is to run that race at a certain pace, but it isn’t cost-free.

So where does that leave you on SB 813 and the quasi-private, or at least more dynamic, regulatory idea there?

Yeah. I mean, it’s something that we’re still exploring. I think there are lots of promising things about it. The idea that lawmakers will have to wrestle with tort liability at some point is important. That’s kind of baked into the fundamental premise of the bill, and I haven’t seen that many proposals that really seek to do that.

If we want to see people have even very mundane uses of AI in their lives that will improve their lives in some marginal but meaningful way, you have to think about tort law, I think. So the fact that the bill does that is a positive thing.

The question is whether there’s a regime there that works for little tech, and I think there are some things that people have said about it that don’t fully wrestle with that component. Some people say, “Well, this is just voluntary.” But immunity from tort law, or some level of immunity and protection from tort law, is a massive benefit.

If you said, “Look, you can get immunity from tort law, but you have to pay $20 billion,” most people would say, “That doesn’t sound fair. That doesn’t sound workable for little tech.” It means every large company is just going to pay to get immunity from tort law, and every small company is going to have to bear that legal risk.

That’s not really voluntary. You’re saying there’s this very valuable benefit, and you’re making it prohibitive for startups. So I think the question is whether the regulatory regime in 813 is workable for startups. There are elements of it that are, but there are elements of it that aren’t.

My hope is that over time, as that bill is examined, it comes into a form that's more workable for little tech. If it's workable for little tech, that's the direction of travel that we want to see. That gets us closer to support.

Do you want to leave the audience with any other thoughts, which could be any specific things that you're tracking and supporting right now, or just other priorities or ideas that we haven't touched on and want to make sure people are aware of?

The one thing I was going to flag is that the Trump administration is releasing this national AI action plan at some point in June or July, so that's a thing to look out for.

Yeah. I definitely expect that there will be many more developments in this story before we get any sort of a stable resolution or stable policy regime, let alone even a stable technology regime. So I appreciate this. I think it was a good, constructive conversation, and I hope we can do it again because—

Yeah.

—if anything, all the time intervals are getting compressed because I suspect all these things are tightening and going to be coming at us with a level of intensity that, if I'm generally right about the direction of where the technology is headed, is going to require all of us to come together and really try to synthesize all the different perspectives into the best possible plan. So come back and let's do that again before too long. How's that sound?

Sounds great. Thanks a lot.

Cool. Love it. Matt Perault, head of AI policy at a16z. Thank you for being part of The Cognitive Revolution.

Awesome. Great. Thanks for having me on.

It is both energizing and enlightening to hear why people listen and learn what they value about the show. So please, don't hesitate to reach out via email at tcr@turpentine.co, or you can DM me on the social media platform of your choice.