The Little Tech Agenda for AI

25:05I experienced it as feeling like the gold posts always move. And one of the things that I said, that I said, that I started asking people when I was really trying to settle into this regulate use, not development, policy position is, what do we miss? Like, if we regulate use, primarily using existing law, what are the things that we miss? And I haven't gotten very many clear answers to that. Like, you can't do illegal things in the universe, and you also can't use AI to do illegal things. And typically, when people list out the set of things that they're most concerned about with AI, they're typically things that are covered by existing law.

25:39Not probably not exclusively, but primarily. And so that at least seems like a good starting point. Some of the other issues that I think are like understandably ones that we should be concerned about have a range of different considerations associated with them. Like if you're concerned about misinformation or like speech that you think might not be true or might be problematic, there's significant constraints on the government's ability to regulate that. The first amendment imposes pretty stringent restrictions. And I think for very good reason, because you don't want the government to dictate the speech preferences, policies of private speech platforms for the most part.

26:13And so those issues might be concerns, but they're not necessarily areas I think where you want the government to step in and take strong action. And so there are, I think there are things that we should probably do as a society to try to address those issues, but government regulation maybe isn't the primary one. And again, in most of the things that people are most concerned about, like real use of the technology for clear, cognizable, real world harm existing law typically covers it. I have a theory on this. So I think everything that magic said is spot on, but then you're kind of sitting around, you're kind of scratching your head, it's like, okay, well, if you use covers it and there hasn't been a very incredibly fair rebuttal on to why use is not enough in terms of focus on the policy regulatory side.

27:00what's the answer? I think we're experiencing sort of this, I don't know if it's phenomena, but we're experiencing this pattern on the crypto side too, which is we're having a very, very spirited debate on the crypto side of things, on how to regulate sort of these tokens, and how do you launch a token in the United States? Is security is a commodity, and this is sort of this age old debate that's plagued securities, traditional securities laws for years, but also certainly with crypto industry. But what we have found is there are a number of people who have entered this debate, who are actually trying to get at the underlying securities laws.

27:38They want to perform securities laws. They don't want to reform crypto laws that involve securities. And this is their only venue by which they can enter that conversation because we're not having, there's no will from the Congress or from policymakers to go and overhaul the securities laws right now. You know, it's just not there. But what is moving is crypto. So people, you know, there are all these people that are now trying to end of this debate and like, oh, we should relook at this and like, well, this doesn't have anything to do with it. We should be entering this conversation. Yeah, they're still pushing, right?

28:11And that's kind of muddy the water. I think a very similar thing is actually happening on AI sign, which is, you know, there are a number of members of Congress that feel like, well, we missed it on the 96 telecom act. Like that wasn't, we didn't do good enough around then. So we need to rewrite the wrongs through the venue of an AI policy conversation, right? Because if you think about it, right, assuming that use doesn't go far enough for someone, right? And this is the same conversation that we're having in California right now, or in Colorado right now, if uses does not go far enough, okay, well then it would be really, really simple if you could have a privacy conversation around this.

28:51if you could have an online content moderation conversation or algorithmic bias conversation around it, you could do all of that, wedge it through AI, and then assuming AI is actually going to be the thing that we all think it's going to be, now you've put basically a regular regulatory funnel on the other side. Like you've put a mesh screen where everything has to run through AI, and therefore it runs through this regulatory proposal you put together. Yeah. The thing that I've really been wrestling with in the last few weeks is whether those kinds of regimes are actually helpful in addressing the harm that they purport to want to address.

29:24Colorado is a really good example. So there are all these bills that have been introduced at the state level. Colorado is the only one that's passed so far. It's set up this regime where you basically have to decide, are you doing a high -risk use of AI or low -risk use of AI? This would be for startups that don't have a general counsel, don't have a head of policy, can't hire an outside law firm to figure it out. High -risk, low -risk, and then if you're high -risk, you have to do a bunch of stuff, usually impact assessments, sometimes audit your technology, to try to anticipate is there going to be bias in your model in some form, which maybe an impact assessment helps you figure that out a little bit, but it's probably not going to eliminate bias entirely.

30:00Certainly isn't going to like end racism in our society. There was a Colorado is now their governor, their attorney general, have put pressure on the legislature to roll back this law, because they think it's going to be problematic for AI in Colorado. And so there was just a special session there to consider various different alternatives. One of the alternatives that was introduced proposed codifying that the use of AI to violate Colorado's anti -discrimination statute is illegal. That's consistent with the regulate harmful use framing that we talked about. And instead of having this amorphous process where maybe you address bias in some form, maybe you don't, this goes straight at it.

30:41It's not a bank shot. Go straight at it where if someone uses AI in a way that violates anti -discrimination law, that that could be prosecuted, the Attorney General could enforce. And I still don't understand why that approach is not, is somehow less compelling than this complex administrative paperwork approach. I think it's kind of the reason that Colin's describing, which is like people want another, a different bite at the apple of bias, I suppose. But it's not clear to me that it is actually the best way to effectuate the outcomes that you want, as opposed to just criminalizing or creating civil penalties for the harm that you can see clearly.

31:15It's also, I mean, in policy making and go writing, it's really, really easy to come up with bad ideas. It's easy, right? Because they're not well thought through. The first thing comes your head, someone publishes a paper on something. Here we go. It takes real hard work to get something that actually works. And then it's even harder to actually go through a political and policy negotiation with a diverse set of stakeholders and actually land the plan on something. Yeah, I think that's part of the reason that people think that we are anti -governance because when we I mean Colin again, he lived this history.

31:48I'm coming in late to it But like as we were ramping up our policy apparatus These were the ideas in the ecosystem licensing nuclear style regulation Yeah, like flops threshold based disclosures really complicated transparency regimes impact assessments audits Which are a bunch of ideas that we think are not going to help protect people and we're gonna make it really hard for low resource startups. And so we've been trying to say, no, no, don't do that. And so that sounds like deregulate. But for whatever reason it's been hard so far to shift toward like here's another set of ideas that we think would be compelling for in actually protecting people and creating stronger AMR markets.

32:26Right now we don't see terrorists or criminals being aided, 1 ,000 X with AI and in performing terrorism or crime. Like when I ask people like, what are you truly scared about? like give me a concrete scenario people, you know, they'll be like, what about like bioterrorism or something or what about, you know, cyber security, you know, after something, we seem very far away from that. Is there any amount of development at, you know, in the next few years, any amount of breakthroughs where you, where you might say, oh, you know, maybe use isn't enough or, or do we think that that will always be a?

32:58I think it's conceivable. I mean, and I think we've been open about that. Like we, we, we think existing law is a good place to start. It's probably not where we end. So, Martín Casado, one of our general partners, wrote a great piece on marginal risk in AI, basically saying like when there's incremental additional risk that we should look for policy to address that risk. And so the situation you're describing, I think might be that. I think what you're getting at is a really important question about just potential significant harms that we don't yet contemplate. We get asked often about our regulate use, not regulate development framework.

33:31or are you just saying that we should address issues after they occur? And I understand why that's a concern. Like there might be future harms. And wouldn't it be nice if we could prevent them in advance? But that is how our legal system is designed. And typically when you talk to people about ways that you could try to address potential criminal activity or other legal violations ex -anti before they occur, that's really scary to people. Like Eric, what if we just learned a lot of information about you and then predicted the likelihood that you might do something on lawful in the future. And if we think it's exceeded a certain threshold, then we're gonna go and try and take action against you before you've done it, so that we can prevent future crime.

34:09That you're laughing, because it's laughable, we don't want a kind of ex -anti -surveillance, both because it feels invasive, but also because it often is ineffective. Like, you might, we might run some tests that shows that maybe you're likely to be predisposed to some kind of criminal activity, but we don't know until you've done it, that you're going to do that you've done it. And so I think that kind of approach, again, I think it's motivated by a really valid concern and a valid desire to prevent harm. What if we could prevent harm before it's occurred? The challenge is the regulatory framework, I think probably won't do that.

34:47Probably won't have the effect of preventing harm. And there are all these costs associated with it, mainly from our perspective, inhibiting startup activity. Yeah. Mark once told me in a podcast, he told me the joke, which is, man goes to the government, you know, go to the government because I have this big problem. Now I get a lot of regulation. Now I have two problems.

35:11Okay, let's talk about the state of AI policy today. There's a lot that's happened last few months with the the moratorium, the action plan. What are some of the things that we're excited about right now? What are some of the things we're less excited about right now? Why don't we give a breakdown of where we're at right now? So I think given what Collins described about where things were a couple years ago, it's great to see the federal government, certainly the executive branch, but not just the executive branch. I think it's in Congress across both Isles being supportive of frameworks that we think are much better for a little tech.

35:39So trying to identify areas where regulatory burden outweighs value and where we can write size regulation to make it easier for AI startups. as Colin said, support for open source. We were in a really different place on that a couple years ago. Now it seems like there's much more consensus. And again, it actually was across the end of the last administration and the current administration around the value of open source for competition and innovation. The National AI Action Plan also had great stuff in it about thinking through the balance between the federal government and state governments, which is something that we've done a lot of thinking about.

36:14There's an important role for each, but that we think the federal government should really lead regulation of development of AI, states should police harmful conduct within their borders, and I think there's stuff in the action plan that would try to ensure those respective roles. There's also a lot of stuff in the action plan that wasn't really talked about much, it wasn't sort of the headline grabbing stuff that I thought was incredibly compelling in terms of, again, trying to create a future for AI that just works better for more people. And a really good example is the stuff on worker retraining that focused on different programs that could help workers if they're displaced as a result of AI, as well as monitoring AI markets and labor markets to make sure that we understand when there are significant labor disruptions.

36:57So I think it sort of gets at a point that you were alluding to a couple minutes ago about like what happens when there's something really disruptive in the future? Can you predict with certainty that there won't be this crazy disruptive thing? And no, we can't. There might be significant labor disruption. Others at the firm have talked extensively about how typically there's always, there are worries about labor disruptions when there's new technology introduced. Typically there are increases in productivity, then it ended up being good for labor overall. We think that's the direction of travel, but you never know, we can't predict it with certainty.

37:27And so I think it's a really strong step to try to just monitor labor markets to see what the disruption might look like so that we're set up to take strong policy action in the future. Can I just say one thing about the action plan? And I don't want to juxtapose this to what we saw under the Biden administration, which is incredible amount of activity in the Biden administration, an incredible amount of activity under the Trump administration. But look, I kind of view these executive orders and these plans that come out from administration are very, very important. And some of them have true policy.

38:01They direct the agencies to do things, to come out with rewards, and then take under rule -making and things like that. But from an AI action plan perspective, for me, it was so significant because I think it turned the conversation on its head. Before it was, we have to only focus on safety with a splash of innovation. And now it is, we understand how important this is from an security perspective. We understand how important this is from an economic perspective. we need to make sure that we win while keeping people safe. And that dynamic and that shift of rhetoric is incredibly important because what that does is it signals to the rest of the world it signals to other governments that this is the position of the United States and will be the position for the next three and a half years.

38:51And this is the position of the United States to the Congress. So when the Congress is looking at potentially taking up pieces of legislation or taking actions or even committee hearings, which, you know, for the broad base of what we're talking about are fairly insignificant. All of that is sort of kept in mind. So now the conversation has shifted significantly and that is really, really important. Speaking of winning Colin, I'm curious for our thoughts on AI policy vis -a -vis China, whether it's export controls or any other issues we care about. Yeah, I mean, well, I mean, look first and foremost, we've talked about already.

39:27I mean, we have to win, right? And I think that that is at the main thrust of a lot of what we're doing here and a lot of the way that we think about this from a firm perspective. You know, I think first is making sure that the founders and the builders can build appropriately with appropriate safeguards and an appropriate regulatory structure. The second is how do we win and make sure that America is the place where AI is is probably the most functional and foundational in all of these of each China. I think that there has been a long conversation, the diffusion rule that came out from the Biden administration, specifically on export controls.

40:09Many, I think, panned that proposal. I think that that was a lot of people suggested it was probably too restrictive. It wasn't the right way to think about things. I think we have spent most of our time Matt leading this effort has spent most of his time, our time, specifically focused on how are we regulating the underlying models and how are we regulating, hopefully the use of these models versus specifically sort of on the export control piece. What I will say though is very concerning sort of some of the proposals that came out from the Biden administration, some of the proposals that we've seen in state level and some of the proposals that we've seen at the congressional level of federal standpoint.

40:52point that dealt with specifically export controls on models themselves. And we're still kind of having this conversation. There's a policy set that has been kicked around for a while. It's called the Outbound Investment Policy, which is basically how much US money from the private sector is flowing into Chinese companies. And it's very noble, laudable, super supportive of that concept. We are a very primary American, first sort of organization here. We're investing primarily in American companies and American founders. So we're very supportive of it. But when you edge into the idea that we might inadvertently ban US open source models from being able to be exported across the country, by definition of open source, There are no walls around these types of things.

41:51So that's one of the areas that we've been very, very focused on. And I think obviously very important to make sure that we don't have these very powerful technologies, US main technologies in the hands of our Chinese counterparts in the PLA and the CCP using this against us. But I also think that we need to make sure that we're not extending too far and limiting the power of open source technologies to be able to kind of be the platform around the world. You have a final point that I'd make here is, we do ultimately and fundamentally have a decision to make as the US, which is do we want people using US products across the world, which helps for a whole bunch of different reasons, but certainly on soft power from that security perspective, or do we want people to use Chinese products?

42:38The more that we lock down obviously American products, the more Chinese, the Chinese will enter those markets and sort of take a land grab in that space. When you get into more, what happened with the more time and the fallout then, soon? I think this one is a bit complicated. There was a perception about the moratorium when it came out that it would have prohibited all state law from existing for a 10 -year window. Obviously, that's a long period of time. I'm not sure we would necessarily completely agree with that policy stance. That, from our point of view, is a misinterpretation for a whole bunch of different reasons of actually what the language said.

43:13But sometimes in DC, a lot of times in DC perception is reality. And that kind of took hold. But I also think that there are also strong competing forces like we've discussed, right? From the, I think the Doomer crowd or the Safety crowd that were very, very anti that had used all of their tentacles that they've spread out over the last decade to try and move and try and kill this. I think they also were successful in leveraging some other industries to try and come in and also move forward to try and kill this thing. And look, by virtue of the vehicle, the underlying procedural vehicle, this reconciliation package that it was moving in, it was a partisan exercise.

43:59It was going to be Republicans on Democrats, and that was that, right? And there was nothing even a prominent AI policy that was going to be dropped in a reconciliation packages that was ever going to drag Democrat votes over it because it was such a big sort of Christmas tree style thing that had all kinds of tax reform positions, et cetera. And if you were in one of those situations, the margins on the votes become very, very, very small. So all it took was, you know, one or two Republican senators hitching their wagon to some of these ideas that were out there to tank this thing, right? And look, I think that's going to be a situation that you're going to fight in any sort of political, policy, legislative outcome or any sort of any any sort of issues that you're going to be running within the Congress, right?

44:52But I think more so than anything. And we heard this repeatedly from a whole bunch of different people and this is what we've also experienced. The industry was just not organized well enough, right? And that's not just the industry. It's also the people who care about this thing that aren't actually industry stakeholders. The stakeholders who were pro some level of moratorium or some level of preemption were just not organized. And I think that that was both an eye -opening moment but also an important moment. Because I think what we have done in the proceeding three, four months since this thing has gone down is we've taken a long hard look at what we need to do collectively from a coalition and to be able to be in a better position next summer there.

45:37And so what does that look like, right? I mean, first and foremost, it comes with writing, doing podcasts, talking about these things, talking about the details of what's actually in these proposals and what it actually means for states and the federal government to make sure that we're fighting through the FUD that's coming through, because it's always going to be there. There's misrepresentation all over the field. The second piece is, let's all get on the same page, which I think we've worked very hard to do. And where we can find alignment, I think we've found that alignment between big, medium, and little.

46:11And then I think the third and probably the most important is, what are we doing on sort of the political advocacy side to make sure that we have the appropriate tools to be able to push forward in a way that ensures that America can do the lead and that we don't lose out on this race to China. And that's part of the reason that we have recently announced our donation to leading the future pack, which will have several different entities underneath it, which I think is designed to be that political center of gravity in the space. And that will fight at the federal level and the state and local level.

46:48So we're happy to be a part of it and we expect there will be others that join this sort of common cause fight on the AI side. If we could wave a wand, what would we like to be done at the state level? What would we like to be? There's the federal level versus how should we think about that in our play? That could be better to where we're at now. Yeah, so I think the helpful answer here comes from the Constitution. Constitution actually lays out a role for the federal government and a role for state governments. Federal government takes the lead in interstate commerce. So governing a national AI market and governing AI development, we think is primarily Congress's role.

47:24Sometimes when people say that, I think what other people hear for some reason is, states should do nothing. And we have tried very hard to be very deliberate in not saying that and making clear that states have an incredibly important role to play in policing harmful conduct within their jurisdictions. So criminal law is a perfect example. There is some criminal law at the federal level, but the bulk of criminal laws at the state level. Like when you think about routine crimes, if you are going to prosecute a perpetrator, it's likely that that would occur under state law. And so to the extent we want to take account of local activity that would where there's criminal conduct involved.

48:05And we want to make sure that the laws are robust enough to protect people from that activity. That's going to be primarily state state law. Oddly enough, I mean, as Colin is describing, like we, this isn't the delineation that that we've started out with, there are a lot of state laws that have sort of taken the approach of sometimes explicitly, Congress hasn't acted, so we have a responsibility to act. And that's true to some extent, like you can act within, states can act within their constitutional lane, some of what states have done have gone outside that lane. And so we actually just this week released a post on potential dormant commerce clause concerns associated with state laws.

48:43And the basic idea there is that there's a constitutional test that says that states cannot excessively burden out of state commerce, when that greatly exceeds the in -state local benefits. And so courts actually weigh that. There's a balancing test. Are the harms cost to out -of -state activity to those significantly outweigh the benefits on the local side? And we think that at least for some of the proposals that have been introduced, it's likely that they won't, that the benefits are somewhat diminished relative to what the proponents think they are, and that the costs are significant, like the cost of a developer in Washington state for complying with a law that's in California or a law that's in New York is gonna be significant.

49:25And so our hope I think is not that the dormant commerce clause ends up serving as a function that makes it hard for states to enact laws, but actually just get serves as a guidepost for states around the kinds of laws that they might actually introduce. And I think it pushes in the direction that's consistent with our agenda, which is to take an active role in legislating and enforcing laws that are focused on harmful use. Looking in the next six months to a year, what are the issues that we're most focused on or that we're thinking about are going to be playing a role in the conversation? Yeah, I think it's first and foremost some level of federal pram shouldn.

50:02And I want to be very specific about this, again, to Matt's point. We're not talking about pramting all state law. We're talking about making sure that we have a federal framework specifically for this model regulation and hopefully how the models can be used, right? I think that's going to be so, so critical because we can't just like any other technology, no technology can live under a 50 state patchwork. And that's been the biggest issue that we've been fighting over the last year and a half or so. So I think that there are some other sort of policy sets that I think will be handled beyond that.

50:47That I think can kick into sort of workforce training. I think there's some literacy things that should be coming up. Obviously, there's a huge robust conversation around data centers and energy. And I think it will be really, really important. But above all, I think most of our time and energy will be focused on trying to have some level of federal standard here. to try and drive the dividing line between the federal and state government, which I think Matt has already done a ton of great work on. Yeah, I think this is just a super exciting policy moment for AI. There's the last couple of years where I think there are a bunch of ideas that have been proposed and for the reasons that we've discussed.

51:23We think those ideas fall short, both in terms of protecting consumers and in terms of ensuring that there's a robust startup ecosystem. most of those laws, I think, have actually not succeeded in passing. So, like, there were a number of laws introduced at the state level in this past years legislative sessions that we thought had a strong likelihood of passing. And I think to date, none of them have passed. Colin has also been building out the expertise and skill set and capacity on his team. We just hired Kevin McKinley to lead our work in state policy. And he, I think, will help us to take a real affirmative position in the legislative sessions ahead on what might actually be AI policy that's good for start -up.

52:02So instead of being in the position of saying no, because we're sort of starting late and kind of with one hand behind our back, I think we're in a position to really actually try to articulate in advance a proactive agenda in AI that's compelling. I think Colin hit the main parts of it, ensuring proper rules for the federal and state governments, focusing on regulating harmful use, not development, and there are specific things that you can do there in terms of increasing capacity and enforcement agencies making clear that AI is not a defense to claims brought on under existing criminal or civil law and technical training, I think, for government officials to make sure that they can identify and prosecute cases where AI is used in a harmful way.

52:41And then all this infrastructure and talent stuff that Collins describing, worker retraining AI literacy, we've also given some thought to the idea that has been articulated by a number of lawmakers and was in the National AI Action Plan of creating a central resource housed in the federal government. You could also do it in state governments as well that lower some of the barriers to entry for startups, you know, compute costs and data access. And we think that's really compelling in terms of ensuring that startups can compete. And that idea, like many of these, is bipartisan. It's been supported by the current administration.

53:16It was supported by leading Democrats over the last couple of years. So that's the kind of thing that we are hoping that when we have the room in position to really advocate for an affirmative agenda that will get some traction in policy circles. We are not always in a hundred percent alignment with other people in the industry. You know, and I think that that's, you know, big, medium, little, you know, across the board. There's other sort of like consumer -avocative groups that obviously feel differently about these things. I think for the most part, the industry is generally aligned on some level of a federal standard here.

53:52And understanding that the thing again, that won't work is a 50 state patchwork. And I think that that's super, super important because I think for the first time you actually have this sort of alignment there. And if you have that sort of alignment, that's kind of momentum that you can actually push things over the finish line and get something done. And I think look also the Trump administration to their credit has also been incredibly supportive of this idea too. There's an incredibly important point. One criticism, usually raised in an implicit criticism sort of way, is, hey, you're the little guys, but often you align with the big guys.

54:27So aren't you just in favor of a deregulatory agenda that works for Big Tech? And one of the things that I think is really extraordinary about the little tech agenda is it's really nonpartisan and it doesn't take a position on Big Little. It basically says, here's the agenda. And when you agree with us, we'll support you. And when you disagree with us, we'll oppose you. And that's not party line, it's not big little. And so I think what we saw over a certain, the phase that Colin was referring to, kind of initially in the recent set of AI policy, was a phase of divergence between big and little.

55:01Licensing regime, bigs were sort of pushing it. Little was concerned about it. Then there's a period of convergence. And I think actually if you look at like the National AI Action Plan, comments across a range of different providers, as Colin's saying, like a lot of them, they had some core similarities. So lots of large companies have advocated for federal preemption. We don't oppose that just because big companies are advocating for it. We think that that's good for startups. I think it's possible, I'm curious, I mean, this is really, you know, Colin really understands this in a way that I don't like how the political chips will fall.

55:33I think it's possible we're in a period of some divergence. And one thing that we hear repeatedly, which is sort of funny, is people will bring us stuff and they'll say, industry agrees with this, so we expect you to agree. You can't, the industries already agree. You can't disagree. And we say the big parts of the industry have agreed, but we sometimes we agree with them, but sometimes we have different views. And so when we disagree, it's not because we're trying to like blow up a policy process or make it difficult for lawmakers who are trying to move something forward, it's because when we're looking at it, we're looking at it through this particular lens.

56:04And I think I hope it's not the case, but I think there might be more fracturing than the months ahead. Yeah, I agree with you on that. And by people, he means lawmakers, just to be specific. Yes. That's a great place to wrap. Colin Matt, thanks so much for coming to the podcast. Thanks very much. Thanks for listening to the A16z podcast. If you enjoyed the episode, let us know by leaving a review at ratethispodcast .com slash A16z. We've got more great conversations coming your way. See you next time. As a reminder, the content here is for informational purposes only. Should not be taken as legal business, tax, or investment advice, or be used to evaluate any investment or security and is not directed at any investors or potential investors in any A16Z fund.

56:48Please note that A16Z and its affiliates may also maintain investments in the company's discussed in this podcast. For more details, including a link to our investments, please see A16Z .com forward slash Disclosures.