Inside AI Policy with FAI’s Chief Economist Sam Hammond

0:28Hey, everyone, and welcome to Generative Now. How do we make sure American innovation doesn't get lost in bureaucracy? We talk about where AI is going, how the governments are thinking about regulating this new technology, and where AI policy might be headed in the near future. Let's get into it. Sam, good to see you. Good to see you too. Thanks for doing this. I have a lot of questions for you. I want to get into all things AI and policy. Before we do that, I think it'd be great for the audience to know a little bit about you and your background in your own words. So tell us about Sam. I'm currently chief economist at the Foundation for American Innovation, FAI, 10-year-old organization founded by a couple of guys in the YC batch in 2014 that wanted to build a bridge between the cultures of Silicon Valley and DC.

1:17We do work now in policy at the intersection of national security, governance, and emerging technology. As our chief economist, I lead our AI policy work that includes working on national security issues like export controls, frontier mobile regulation. But then we also host a conservative AI policy fellowship to try to build some talent on the right for AI policy. I've heard you describe yourself as a technologist and a policy wonk. You told us sort of how FAI got started. What, I guess, drew you to the intersection of tech policy and all my FAI? Sure. I mean, I grew up in rural Nova Scotia, in a phishing community, and spent, you know, it was raised on the internet.

1:57So I'm very into philosophy, cognitive science, ended up fascinated by American politics, as many Canadians are. We're all kind of America watchers. I suppose my interest in AI kicked off with reading Ray Kurzweil back in, I think, 2004. I read his 99 book, Age of Spiritual Machines, and that's where he lays out the 2029 forecast for AI's passing a Turing test. And that has always sort of stuck in the back of my mind. and my academic interests steered more towards political science, political philosophy, how institutions evolved in light of the industrial revolution. And so I've brought that perspective in how I think about AI.

2:36If this is as transformative as the industrial revolution or the printing press or these other major technological paradigm shifts, then what does that mean for our institutions going forward? It seems to suggest that there's going to be much more radical, not just change on the technology level or the economic level or the output level, but like how, what institutions do we need to, to keep up with that pace of change? So you briefly mentioned some of the issues in the areas you spent a lot of time thinking about it in the beginning of the conversation. I guess like talk to us about some of the ones, especially now that you're most focused on in terms of AI issues or policy debates that you think are going to be the most important to AI.

3:16Sure. I mean, there are only a handful of like key inputs or ingredients that go into competing at the frontier. It's things like data, better algorithms, the talent, you know, having world-class software engineers, ML engineers, but then also energy. To your original question about like, what are the big policy things? It really cashes out to energy and chips. So on the energy front, you know, doing everything we can to build domestic energy capacity, especially electrical, you know, low, low intermittency energy, because when you're running these data centers, you made this sometimes billion dollar, multi-billion dollar investment, mostly going towards GPUs.

3:55To recoup that investment, you want those GPUs running 24-7 for training, for inference, for renting the value, whatever. And so that requires having persistent baseload, non-intermittent sources of energy, which the US has a lot of latent potential for, right? Out in the, you know, in the plains and in the Pacific Northwest, there's lots of potential geothermal. You're seeing more interest around rebooting the commissioned nuclear facilities or leaning into SMRs and these more cutting edge nuclear reactors, potentially even solar. But just the fact of the matter is, while we've added, you know, lots of renewables to the grid, we've taken off equal proportions of coal and fossil fuels and And so we have basically added nothing to the grid net to extrapolate to 2030 unless we add something.

4:44I think globally, data centers are adding about 100 gigawatts to global energy demand. The U.S. wants to be at least half of that. So we need to go full tilt on energy production, permitting, and all the other bottlenecks that stand in the way. And on the chip front, ensuring that we maintain our lead, whether that's through investing in R &D to make sure that we stay ahead of the curve in new kinds of silicon. Right. That makes a lot of sense. So I guess when you think about the ingredients and maybe with a special focus on the hardware, given your point that that's probably where our advantage lies today, ours being the U.S., I guess, like, what role do you think governments should play in sort of guiding both the development and really, like, I guess, the defensibility of these four ingredients here in the U.S.?

5:31Yeah. So, you know, energy, as I already mentioned, a huge and probably, you know, over the medium to long term, the bottleneck. Yeah, investing in export control capacity for enforcement. So, you know, that is done out of the Bureau of Industry and Security, which is tasked with this, you know, equally, in my view, important mission, but something that requires a technical depth because the semiconductor supply chain is the most complex supply chain in the industry. And, you know, they have maybe 50 enforcement officers, 50 to 100 is in the budgets around 50 to 100 million. There were in 2024, three smuggling cases that the estimated profit that the smugglers were making off of smuggling chips was two and a half times BIS's entire enforcement budget.

6:13So I've been working on this bill called the Chip Security Act, which would require paying based location verification for exported chips. And there's also another bill called the Stop Stealing Our Chips Act, which would create a whistleblower program at BIS modeled on the SEC program for white collar crime, where if you're a guy in Malaysia and you catch wind that there's chips being smuggled, you could potentially win a portion of the resulting fine. So basically efforts to make these things scalable. Yeah. And then stepping back, the other potential sort of, I would say like tail risks to AI in the US are things like a bad copyright ruling.

6:55Explain that. Well, right now, imagine the meme of all of the economy standing, being balanced on fair use. So under US copyright law, fair use describes a doctrine that says you can use copyrighted materials if it's transformative in some way. And that can be a very subjective... So is someone on YouTube reacting to this podcast right now where they just have their little face in the corner? Is that transformative? Probably. And for the most part, that is. That's why there's so many React videos on YouTube. But if they just stood there in silence and didn't add any actual value, is that transformative?

7:31Then you get into the gray areas. So the big question for AI is training data. Right. You mentioned one of the ingredients, right? Data. Right. Exactly. And so the argument has been, and I think there's actually a recent case involving Anthropic where they prevailed on this argument that actually, yes, training on copyrighted data is protected by fair use. What's not protected is like if you go to an image model and say, make a Italian plumber with a mustache and it spits out Mario. Right. The outputs, you know, how you use these models and these models do end up memorizing a lot of their training data.

8:05And so if you're the New York Times and you put a lot of effort into getting into regurgitate an article, you can succeed at that. In some sense, you have to know what the contents of the article is already before you can actually do that. But otherwise, you know, the way the courts have so far interpreted it, and I think the way the labs interpret this is that, you know, training on copyright data is not that different than like an artist listening to music and getting inspired. What matters is how that gets transformed in the context of a neural network to like interpolate and extrapolate between those different styles and things of that nature to produce something generally novel.

8:37Right. So you mentioned a bad copyright ruling, I think is how you phrased it, could impact, drive some of the policymaking and regulation around AI. Is that what you mean? Yeah. I mean, a lot of just policy in general in the United States is determined by the courts. Right. It might be a negative ruling on fair use. The OpenAI New York Times case, You know, there's one of the arguments that's been made is that OpenAI is obligated to preserve everyone's, all their users' chat logs that has its own, you know, weird implications. Like, you know, imagine, you know, the privacy concerns if you have to store all that data and there's open questions around like, okay, so as these models get implemented in the real economy, you know, maybe you're using it to cheat at your homework.

9:25And so the teacher should be able to like somehow see that. But maybe you're using it as a therapist and, you know, a real therapist has a kind of privileged relationship with their client. But like if this data can just be, you know, discovered in court or something like that, these are just giant open questions. So energy, training data, practices, fair use. What are some of the other areas where policy may play a big role? There's lots. I mean, this is what's coming up in Congress. It's just everyone realizing that everything is an AI role. Right. And so big open questions about what to do around workforce development and retraining.

10:02Right, of course. I don't think there's any good answers on that so far. Yeah. What to do for enabling regulation. We've seen many jurisdictions roll out Waymo and now Teslas in Austin, fully autonomous self-driving cars. But to really get the full vision of full autonomy, maybe the government transportation needs to have some kind of framework, an enabling framework. And you can imagine similar sort of enabling frameworks for drones, for anything that interacts in the real world or that crosses state lines, that sort of thing. AI for science. So, you know, there's different pieces to this. You can imagine, I think there's already been studies showing that frontier AIs write better peer reviews than actual scientists.

10:48You could say, well, maybe the process should be that your journal article goes through the AI peer review first, and then the peer reviewers check it. There's all kinds of ways you could see the process changing. The question is, will our institutions adapt and co-evolve and embrace the technology and potentially use it to accelerate the pace of scientific grant making and that sort of thing? That's the institutional side. And there's on the AI for science per se side where, you know, there's proposals and people working on ideas around like self-driving labs. Could you have an AI scientist that is continuously generating hypotheses, running experiments and actually connected to, you know, a laboratory where it has a petri dish and so on?

11:30What could that potentially unlock? And is there a role for the national labs, the Department of Energy to enable some of that maybe by, you know, through their facilities or potentially through unlocking federal data sets? Right. You know, DOE probably has the world's most comprehensive data set on like chemistry, material science. Right, right. A lot of that data has strange data sharing agreements and permissions. And so, you know, there's all kinds of work that needs to be done. I want to better understand from you, like how sort of policymaking and lawmaking actually will like in practice impact all this stuff.

12:03But maybe like first zooming out, just like what do you feel is at stake here? There's sort of what you could think of as the more pessimistic world for AI. Not that AI goes bad, but the technology sort of hits some walls or plateaus. It's a really important technology, but it's maybe like the internet or something like that. And so the upshot of maintaining a lead means that when we cross these sort of thresholds where we can have large chunks of the economy become driven by AI, fully automated factories, advanced R &D, that this will produce a growth dividend, right? Even under pessimistic scenarios, a major increase to productivity growth.

12:41Then there's like the more, you could say, more aggressive scenario where like we are about to cross over thresholds of AI capability where, you know, it's not just AI giving people an uplift in their job or giving a productivity dividend. But we have like AI agents that can work for eight hours or a week or a month at a time uninterrupted, potentially able to run thousands or millions of those in parallel, including applied to the AI R &D enterprise itself, potentially unlocking a recursive self-improvement loop where we pull ahead dramatically. Or even if we have a six month lead over our next competitor, that six months is sort of telescoping in progress and compressing the pace of progress for that six months makes all the difference.

13:26So six months makes all the difference or it could make all the difference. Like then I guess let's let's talk about sort of mapping this notion of like time being an important factor to lawmaking and regulation. I guess like talk to us about that tension. I mean, with the technology this important, it feels like there's going to be regulation. How do you view the importance of that regulation or do you feel like there shouldn't be? And how do you square that with the race to maintain this competitive advantage? Yeah. So, I mean, there's different elements to this. On the permitting side, for instance, permitting just to get the environmental reviews complete for a high voltage DC transmission line is seven years.

14:06And there are some transmission line projects that take 30 years. Right. And so when we're talking about adding, you know, tens to hundreds of gigawatts to the grid by 2030, we need the power lines to come up a lot faster than that. And so this time pressure is really coming to bear in a way that it hasn't in the past. And that's leaning towards, you know, taking more emergency measures. Can we like leverage national security provisions in various environmental laws to expedite process, things of that nature? Can we lean into behind the meter energy production where you don't even need to connect to the grid?

14:39but there are certain advantages on the permitting side. Then there's this broader perspective, which is like, you know, AI is going to affect everything. Right. And so, you know, when AI comes for drug discovery, you know, we're going to need, we can't just give FDA officials like AI co-pilots. We're going to like rethink the drug approval process. And same with like transportation and all these other areas. And so most of the stuff is in statutes. Right. There's limited things you can do. You can, you know, maybe do a pilot program at FDA, but like to really change the process, you're going to need Congress to step up and act and have a sort of burst of legislative productivity that we haven't seen since like the New Deal era.

15:15So what actually happens? Like, what do you do? Again, when time, you just mentioned, like six months could make the difference. Yeah. So this may be the ultimate bottleneck is, you know, I lived in DC during COVID and it was striking how early a lot of technologists were to the pandemic, right? You know, I remember like Bellagio Svendalvassin on Twitter with his famous going viral thread in January, where he said, you know, this could be the one, right? And you had to talk to normal people, normies, you know, they'd say, well, you know, it's like 17 people in Italy, you know, I'm not that worried, but they just didn't understand that exponential.

15:49I think we're in a sort of similar space with AI where, you know, it started to change even in the last six months, but the degree of sort of common knowledge of the urgency among members of Congress and policymakers more generally is still in that January 2020 phase where there are people who are early and see where it's going and are trying to raise alarm bells. But a lot of people are like, in my own experience during that January 2020, I talk to people like, are we going to do anything about this pandemic? And they say, the NDA reauthorization is up. We don't have any room in the calendar.

16:26We're never going to get to it. And I remember March 12th, 14th, whatever it was, when the NBA season was canceled. Tom Hanks got COVID. The WHO declared a global pandemic. Two weeks later, we passed the CARES Act, which was this$2.2 trillion pandemic response bill. So when Congress - They can move faster. They can mobilize. When there's a crisis, Congress can move quickly. Yeah. The question is, will they do it smartly? What do you foresee? What's a potential crisis with respect to AI that could mobilize Congress in that sort of timeframe? We're talking about inching our way towards artificial general intelligence.

17:07And the point of AGI is the generality. And that makes it very difficult to talk about because when you have a very intelligent general system, it affects everything. And I think you could argue that we already have sort of baby AGI's with the frontier models. What they lack is autonomy. Yeah. Right. And autonomy is going to be this massive unlock that we're just starting to see with reasoning models. And now you have the meter research showing that the task horizon length that these models can perform at is doubling roughly every four to seven months. So, you know, if today they can do a 30 minute tasks, you know, 30 minute long task and in four to seven months, they'll do an hour long task and two hours and four hours and eight hours.

17:45And so pretty quickly, we're looking at systems that will be able to perform continuous, reliable work for weeks, potentially months at a time. And that, I think, will be a wake-up call because that autonomy will manifest in various different ways. It will manifest in potential cyber vulnerabilities. It's one thing to say that hackers in Moldova are using Cursor to give them a boost. it's another thing if like anyone anywhere with some GPUs can launch thousands or tens of thousands of phishing attacks of you know agents that are patiently waiting in the contributing to the repo waiting for the right moment to try it.

18:25But in theory like wouldn't something like that like that type of disruption actually mobilize lawmakers to probably restrict AI development like what like it almost feels like if you compare it to the COVID example it would probably have like almost like a deceleration effect. Yeah, I think another commonality of COVID is it's sort of like once it's out of the lab, it's out of the lab. And there is this giving the declining cost of compute, given algorithmic and hardware improvements, models that today cost a billion dollars to train and a few years time will cost$10 million to train. And so there is going to be this diffusion of capabilities, including through open source, that mean there is no real slowing down path.

19:07And the extent that regulation does try to slow down the frontier is going to be slowing down the good guys, the labs that actually have oversight that are policing misuse through their APIs and that sort of thing. And so - What are the bad guys? The bad guys are, well, there's misuse, right? But also just like, you can imagine once we have very powerful AI agents that can do computer use and long-term planning and that sort of thing, that you have a disgruntled ex-boyfriend who wants to create the, you know, malware bot that once it's released and self-replicating is like permanently terrorizing his ex-girlfriend.

19:39Right. Sending her Domino's pizza. You know, she moves, she like posts on Instagram that uses its, you know, geoguessing abilities to guess her new location. Yeah. And I think that's going to be a kind of genie that releases out of the bottle that like we will likely adapt to, but that adaptation will generate, you know, So we won't be able to adapt without also deploying AI tools on the other side to detect whether the user on the social media site isn't a real person or not. And so it's a dual use nature of AI means that we can't afford to slow down because we need even better AI to defeat the bad guys.

20:22What are some other examples of where you where you foresee sort of like the ground truth of what's happening with AI and development and maybe some of the industries that are affected by it compared to, you know, where lawmaking may not may not keep up? Like where where is there going to be like another collision or mismatch? Yeah, well, I don't think it's going to keep up with any of these places. This is one of my worries is during COVID when FDA was still walking the vaccines, there was lots of calls to do right to try at the state level. Some points, the states have started asking, does the FDA have authority over us?

21:02Or can we just approve this drug immediately so we can get it out of our homes? And so if there's this rapid progress that our institutions are failing to keep up with, I think there will be a tendency for things to sort of find the weakest link, find the water will flow, you know, to jurisdictions that don't have similar bottlenecks through, you know, just people sort of nullifying what the rules say. So, you know, when it comes to Congress, I do worry to your point that the immediate response will be kind of reactionary in the sense of... This is like a disruption. So it's the idea and the position, I guess, then about the AI, it's like no regulation, completely hands off, like let this technology accelerate.

21:47I mean, we want to go faster. We want to build lots of energy. The acceleration-deceleration dichotomy misses the ways in which technology is not a linear process. There are different kind of branching paths we could go down. There's, I think, a strong case for a more defensive accelerationist point of view where We want to pull forward the AIs that can detect zero-day exploits in our codebases before we have the AIs that can exploit those codebases. And some of that may require differential access on the part of the labs. If OpenAI develops some really radically more capable model, let the national security actors know in advance so they can actually both have some foresight, but then also brace for impact, so to speak.

22:33And then I also support broadly like, you know, standard setting from the newly formed Casey, which is the formerly the AI Safety Institute, voluntary standards, transparency. I think there's some prerogative on labs to actually share their information, including with each other so that they sort of know where the state of the art is. What types of information, model weights or? No, not necessarily IP, but like, you know, incident reporting. Okay. if they stumble on some new alignment technique. I think also these data centers are going to become increasingly critical infrastructure and assets, both for national security purposes, but also imagine we roll the clock forward 10 years and a large cross-section of knowledge work is being performed in data centers.

23:21That's the single point of failure. If you do a cyber attack, it's not just turning off the lights. It's turning off a third of your economy. You want those things to be heavily fortified. And I think you could argue there are market failures right now where data center developers, they do some infrastructure security, but they're not putting up the fixed cost to be defended against nation state actors. What about maybe just a couple other specific examples I'd love to get your take on? What about fair use? You mentioned that this is a big topic for training data. How do you see that landscape playing out?

23:59And I guess, where does FAI stand on sort of policymaking in that area? Yeah, I should have said it. The FAI, we don't have a one voice policy. So we have broad sort of values and principles with the builder ethos and so on. But on copyright, I think we're mostly all aligned that so far the courts have held up fair use. But if that was ever in jeopardy, I would like to see an affirmative policy or for lawmakers to just put in a statute that training data is completely fair use. Because we see abroad in the EU with their copyright directive, but even countries like Canada or Australia just have modestly less permissive copyright.

24:39It is a wet blanket on developers. Well, I mean, what would happen, I guess, if there was a different type of ruling and maybe lawmakers were slow to respond to this? Yeah, then suddenly data collection halts. you could even see injunctions that force models offline. Oh, wow. Like voice open AI to shut down. I mean, in principle, if the models that they're surveying are trained on data that they don't have rightful access to, there will still be open models, right? But fines, massive fines, some of these cases are talking about every article that you trained on from this news service is$100 ,000 or something like that.

25:21And so those add up when you're talking about trillions of tokens. Yeah. But you're saying the current take is that, and you said, I think you said there have been some recent cases, the current determination from courts, I guess, not lawmakers, is that this is fair use, but that could change. What about, I mean, what about anti-induced labor disruptions? Like, how do you square this? As you mentioned, the cat's out of the bag, but there, you know, I think the belief is there will be labor disruptions is the idea that like you just let that happen or there should be some policy or regulation to sort of ease us into this new world.

26:00Like, I don't know. I'm curious how you think about that from a regulation standpoint. Yeah. So before I joined FAI, I used to be a director of social policy at a think tank called the Niskanen Center. I was there from 2016 up through 2023. And my prerogative there, even though I come from a tech policy background, was to think about what social insurance and other complementary institutions are needed to facilitate creative destruction so you don't get these perennial backlashes. And so we worked a lot on child tax credits, income support programs, unemployment insurance reform, workforce and retraining.

26:33And I got a crash course in just how backwards and convoluted these systems are. So going back to another COVID example, remember we went into lockdown and quarantine, business to shut down around the country. there were lineups around the block to claim unemployment insurance. Sure. Some of these unemployment insurance offices in different States, most of them run on like COBOL or, or for mainframes. Most of the people who can actually code that have retired. Yeah. Some of the, some of them had websites that were only active during business hours. And so, you know, that was a, a, a massive, massive, you know, cluster, you know what?