← The AI Policy Podcast
The AI Policy Podcast
What California's SB 53 Means for AI Safety
00:37:22
Greg and Sadie dive into California's new AI transparency law, SB53.
Loading summary
Transcript29 lines
- [00:00]
A
Foreign.
- [00:11]
B
Welcome Back to the AI Policy Podcast. Today, we're doing a deep dive into SB53 that enacts California's Transparency and Frontier AI Act. So I'm Sadie McCullough, and I'm joined, as always, by Greg Allen. So welcome back, Greg. I'm really excited for today's conversation.
- [00:27]
A
Me, too.
- [00:29]
B
All right, Greg, let's jump right in. So on September 29, California Governor Gavin Newsom signed into law SB 53, which enacts California's Transparency and Frontier AI Act. Exactly one year earlier, Newsom vetoed SB 1047, the predecessor of SB 53. So before we get into the details of the text, could you walk us through this brief history?
- [00:53]
A
Yeah. So I think it's worth taking us back to the spring of 2023, and that is when a group of very prominent AI researchers and leaders of AI companies signed an open letter saying that the risks, the existential risks of AI should be a national priority on a par with the things that we do to prevent risks of nuclear weapons or pandemics. So that was signed by Sam Altman. That was signed by Demis Hasabis, the CEO of Google DeepMind. A huge share of the leading lights of the AI industry were seriously talking about existential risk of AI, meaning the risk of human extinction. So from that letter, which I think I said at the time and I still believe had a transformative impact on Washington, D.C. and the conversation of AI. I mean, basically the conclusion of that letter was that serious people take existential risk of AI seriously. And so how can the government not take it seriously as well? These are a lot of things that people have been saying inside these companies. Now, they were saying them publicly, in the open, and in a way that was kind of impossible for the government to ignore. So out of that came the Biden administration's executive order. Out of that came Chuck Schumer's AI Insight forums when he was the Senate Majority leader trying to come up with comprehensive AI legislation to address, among other things, catastrophic risks. And one of the other things that came out after a lot of folks were disappointed that there was no movement in the US Congress was SB 1047, which was a landmark AI safety bill. And there's a bunch of reasons why this has the potential to be very important. You know, number one is, according to Governor Newsom, you know, of the top 50 AI companies in the world, 32 of them are headquartered in California. So California, if it regulates even just the companies that, you know, develop stuff in California, that can have big implications for what's going on in the entire AI industry as a whole. And the second thing is that California is a big market in the same way that Europe is a big market. So when Europe exercises its regulatory muscle, it's regulating, you know, what gets to be deployed to the European market. And sometimes that can have major implications that go far beyond just the European market itself, as people take those European standards and apply them globally. Such is the case in California, most notably with fuel efficiency standards. A lot of car manufacturers do not want to make one version of their car for the California market and another version of their car for the rest of America. And so California fuel efficiency standards have had an outsized impact. Impact in terms of driving U.S. automakers and companies that want to sell to the United States in increasing their fuel efficiency. SB 1047 was looking to extend that and do so in the name of AI safety, going after these potential catastrophic risks. And it had a bunch of pretty extensive provisions that at the time, I think to its proponents and to many others seemed appropriate given the way those risks looked. If we're dealing with human extinction, then mandatory third party audits don't seem that absurd. Right. We have mandatory third party audits for the people who run nuclear power plants, and they could only kill tens of millions of people. Right. And you're saying that AI could kill literally every human alive? Well, it seems like maybe you should have, you know, stricter regulatory standards. Well, that ultimately passed in California's legislature, but was vetoed, as you mentioned, Sadie, by California's governor, Gavin Newsom. He basically said that this bill was too onerous, given where we are in the technology's evolution. It wasn't clear that the things that they were trying to do would actually solve the problems they were trying to solve. And so he vetoed it. But he also set up a commission to write a report to make recommendations to sort of address this balance. And that included some of the leaders of AI thinking from multiple different segments, with a heavy, heavy emphasis on voices from California. They came out with the report, and it basically said, yes, something needs to be done. Here are principles for what reasonable legislation might look like, sort of given the state of the technology, the risks that it poses today and the risks that it could conceivably pose in the future, combined with the opportunities that it poses today and in the future. So basically, the. The group, the expert group that Newsom set up came back and said, yes, you do need to do something. California should do something, or. Or the federal government should do something. And if the Federal government fails to do something, California should do something. So the author of SB 1047, the lead author, I should say, Senator Scott Wiener, basically came back after that report was published and said, okay, now here is SB 1053, a slimmed down version with less onerous requirements that adheres to the principles that came out of that expert committee report and is now what California should pass. And remember, the legislature had already passed SB 1047, so they were ready to sign up for the very onerous requirements. It was only Newsom's veto that prevented that. So this again passed the legislature, this more modest set of requirements, requirements, and then Governor Newsom did indeed sign it. So now we are facing this as the new law with most of its provisions set to take effect on January 1, 2026.
- [06:40]
B
Sure. So I was just going to follow up and ask you gave a brief overview of SB 1047. What are these key differences between that bill and SB 53 that we're looking at now?
- [06:53]
A
So at the highest level, I think we're dealing with the difference between an AI safety law and an AI safet transparency.
- [07:03]
B
Okay?
- [07:04]
A
That's how Senator Weiner describes the differences. And I do think that's a fair reflection of what is going on here. So whereas previously the companies, you know, were forced to submit to external audits, they faced penalties, you know, if they didn't do specific things in preventing the safety of it. This version, SB53, it says that you have to have your own internal safety framework, and you have to publish that conspicuously on your website as the legislation describes it. So, like, what is your safety framework? Are you adhering to national standards? Are you adhering to international standards? You know, what are you doing internally to ensure the safety of your models and preventing that from poisoning, you know, presenting catastrophic risks to Californians and therefore, of course, the wider world. So the point here is it's like it's not saying you shall use this safety framework, it's just saying you have to have a safety framework and you have to tell the world what your safety framework is, and you have to tell the Californian government whether or not you follow your safety framework. And the penalties here are quite modest. They are civil penalties, not criminal penalties, penalties. So nobody's going to do jail time for this. And they also max out at $1 million fines per violation. So let's just like walk through an extreme example here. I often think in extremes because it sort of highlights the boundary conditions of the situation. But let's imagine a hypothetical AI company called Evil AI, you know, not open AI. This is evil AI. You know, they're not about openness, they're about evil. And they publish a safety framework and it says, we're not going to do anything serious other than like have a staff meeting where we say safety's great and we should do it. And they publish that as like their safety framework. Not necessarily anything bad happens to them. Right? Like, they could have a terrible plan for ensuring safety. But the thing is, now it's going to go out on the Internet. And so the Internet is going to see, okay, their safety plan is obviously a joke. And so a lot of the sanction is reputational or putting themselves at risk of additional legislation. You know, if basically the California legislature sees that evil AIs, this hypothetical company's safety framework is we talk about it at one meeting per year and do literally nothing else, well, then the California legislature can say to itself, okay, we tried to be nice to you guys, but clearly you have to be forced, forced to take safety seriously. So that's like one version of the sanction. The second version is imagine that evil AI has a very robust safety plan, right? It's like we're gonna, you know, check everything a million times and at the slightest whiff of risk, we're gonna shut everything down. Then they publish that online. If they don't abide by that, well, then the California state attorney general can bring suits against them and fine them up to a million dollars. So every time they're violating their internal safety protocols, they can be fined a million dollars.
- [10:37]
B
So if they're violating their own safety protocols that they wrote.
- [10:41]
A
Yes. Which is kind of interesting. Right? So. So it's trying to give self regulation a little bit of teeth in this regard. Now, the teeth aren't that sharp, though you could argue, because we're talking about a million dollar fine. And OpenAI and Anthropic and XAI are all, you know, they're all signing deals for data center construction in the tens or hundreds of billions of dollars. So like $1 million fines, that is budget dust to them. So the monetary consequences of the fine are maybe not that big. But then again, it comes back to those reputational downsides, right? No company wants to see headlines called, you know, Company X repeatedly, extensively violates its own internal safety protocols, you know, when it comes to catastrophic AI risk, et cetera, et cetera. And again, you know, those violations, they have to be disclosed in an anonymized form where you try and not say, you know, which company experienced which safety incidents, etcetera, Etc. In an anonymous form is going to be published in an annual report. But, you know, the specific details of any incident can be referred by the Office of Emergency Services, which is sort of vaguely analogous to the California state equivalent of the Department of Homeland Security. It has other functions as well. But that's a useful analogy. So the Office of Emergency Services, which has some of the enforcement role in this legislation, and the California Attorney General, which has a big enforcement role in this legislation, they can now refer details of specific incidents to the California state legislature, right? So the public is going to get this anonymized report, but the California legislature might be getting the nitty gritty details of every single violation. And that means they can use that to inform their decisions about future legislation. Also, you know, as anybody in Washington D.C. knows, once Congress knows about it, the odds that it's going to leak to the media go way, way up. And so the same, of course, could be true at the California state level, right, where violations are disclosed to the various California regulators, then to the legislators, and then the legislators make it out to the media. So that then comes back to the reputational sanction and the risk of additional legislative action. So on the one hand, you could argue, you know, these teeth aren't very sharp. $1 million fines aren't the end of the world for most of these companies. On the other hand, you could sort of say, well, we're starting from a much lower baseline, right? We're starting from not quite nothing. So this is something, and it's something that you could imagine that the companies involved would care about.
- [13:38]
B
So we've talked a lot about, like the penalties for not complying and who's enforcing this bill, but who does this actually apply to and under what conditions does it apply to them?
- [13:47]
A
Yeah, this is a really important question. So there are carve outs that are designed to make this legislation apply to, number one, the types of organizations that can afford to comply with it, and then also to people who are actually in a position to cause great harm. So the first thing is it only applies to the word in California law is persons. But persons under California law means much more than just people. A company can be a person, person, a nonprofit can be a person, et cetera. But it's persons who had more than $500 million in annual gross revenues the previous year. And it's focusing on frontier developers. So if you think about, by contrast, the EU AI act. The EU AI act has regulatory hooks that apply to the people who create models. It also has regulatory hooks that apply to the people who use models. This legislation is overwhelmingly focused on the developers of the models, so called frontier model developers. And because of this $500 million revenue position, it's not applying to small startups, it's only applying to those who are doing something at a relevant scale.
- [15:06]
B
So does this include AI labs as well?
- [15:09]
A
Yes, definitely. I mean, OpenAI is going to be hit, Anthropic is going to be hit, Google's going to be hitting. Interestingly, there are provisions in this dealing with open source and open weight models, but none of them limit the applicability of this regulation. So Meta will also be affected whether or not they are trying to sell their model. Even Deepseek. Deepseek, the Chinese AI model developer, because they're making it available to Californians, because they're deploying it in California, this would apply to them. And what's interesting is that that $500 million revenue hook, it applies to the person or its affiliates. And so that means your corporate owners or your corporate subsidiaries. So in Deepseek's case, I think Deepseek has very little revenue and certainly had very little revenue last year. But Deepseek is a subsidiary of High Flyer Capital Management, a quantitative hedge fund in China that definitely makes a boatload of money and probably $500 million. That's why they could afford to start something like Deep Seq and not really need attract external financing in a significant degree. So that to me says, you know, theoretically, Deep Seq is going to have to publish, assuming they want to continue making their models available in California. They're going to be legally required to disclose, you know, what is their safety plan, how are they going about it, and they're going to have to have the same publish that openly on the web in a conspicuous manner, et cetera, et cetera. So that is some really interesting implications for who this applies to.
- [16:47]
B
Okay, Greg, so you've already talked about frontier model developers who meet the financial threshold. Can you talk about other frontier developers that maybe don't meet this threshold and how it applies to them?
- [17:00]
A
Yeah, so technically, all frontier developers, again, you know, making their products available in California, deploying in California, do have some obligations. I mean, maybe, maybe not. A frontier model developer that's only doing it for internal use, it's not clear to me that that would be covered. But if they're making it available for public use, they do some requirements, and those are transparency requirements, but they don't go to the same level of you have to publish a safety framework, you have to tell US of instances in which you violated your safety framework, et cetera, et cetera. So that is more about inventorying everybody in California who's doing this stuff. Like, you have to tell us if you are indeed a frontier model developer versus telling us all the myriad ways in which you're working to ensure that you are safe. That sort of those safety transparency provisions, I think, are more targeted at the large frontier developers.
- [17:55]
B
All right, okay, so what exactly is a frontier model?
- [18:00]
A
Yeah, so, like, the regulation applies to frontier developers, AKA developers of frontier models, logically implies, like, what is a frontier model? And what's so interesting here is that they have defined it using a familiar definition. Quote, a foundation model that was trained using a quantity of computing power greater than 10 to the 26th integer or floating point operations. Now, that 10 to the 26th number, I think, will be familiar to many listeners of this podcast because that is the same number that was used in the Biden administration's executive order. So what they're basically saying is we think that safety risk is correlated with intelligence. Right? The smarter something is the theoret radical, you know, greater degree that something bad could happen if that intelligence was put to bad use. So they don't really have a perfect way of measuring the smartness of a model. And so they're using this imperfect proxy, which is how much computing power was used during the training phase. And just to be clear, that really corresponds not to the amount of computing power that was used to train the original ChatGPT. It's really like one level beyond what was available at the time, the Biden administration executive order. So what they're sort of saying in the executive order, they were saying, we don't want these stipulations to apply to the state of the art now. But whatever the state of the art is going to be a year from now, it's going to apply to you. Well, now we're past that year, right? So this, this applies to everything that is the state of the art right now is using more than 10 to the 26th. And it's interesting because it contrasts with the EU AI act, which had a lot of their hooks applying at the 10 to the 25th threshold, because they did explicitly want to say, hey, this applies to the stuff that's out right now. So I think it's also worth noting that when we hosted Michael Kratzios, the director of the White House Office of Science Technology Policy, one of his criticisms of the Biden administration's executive order was that use of the 10 to the 26th computing flops threshold. He basically said, look, this is a, a obviously arbitrary number. Like, how does that help us decide, you know, who poses a safety risk and who doesn't? And I heard from Biden administration officials at the time, and they're like, we absolutely concede that this is a highly imperfect number. If anybody has anything better, you're welcome to tell us what it is. We think this is good enough for right now for what we're trying to cover. And I've heard similar things from some of the companies involved in creating these things that, that 10 to the 26th is a highly imperfect threshold, but a threshold that they understood and could live with. So here we are. Now that 10 to the 26th is being incorporated into, you know, one of the most important pieces of state AI regulatory legislation that we've seen.
- [21:01]
B
So why are people like Michael Kratios and the Trump administration having such problems with 10 to the 26 flops being the benchmark, especially if the EU AI act is using 10 to the 25th?
- [21:17]
A
So I think, you know, Kratzios, I have the transcript of the conversation here in front of me and his line, he has a few things. Number one is, quote, it set very weird arbitrary limits around pre deployment testing of AI models. And it created this sort of fear in the AI community that the government was going to come down and overly regulate this test technology. So I think there is. He has two big criticisms. Number one is that the Biden administration was overly focused on fear. In fact, he said, quote, fear was what led almost all of their policy decisions, end quote. And then the second thing is that the rules and the limits around which they dealt with that were like, arbitrary. Which is to say, you know, you, you could draw the line at 10 to the 26 flops. Why don't you draw it at 10 to the 27th flops? Why don't you draw it at ten to the 25th flop? And so his point was like, look, that's just an arbitrary threshold. And I think the Biden administration would concede, yes, it's an arbitrary threshold. All we're really sort of saying is we're trying to capture that we're interested in state of the art models. 10 to the 26th. At the time the Biden administration was drafting, that EO represented what was expected to be the next generation state of the art, so not the current generation state of the art. And I think it's also worth noting that in this California legislation, they specifically say, you know, that this threshold should be evaluated again. And potentially updated on an annual basis. So it's the job of the California government's executive branch to sort of say, hey, like, even my children's calculator now was trained on 10 to the 26 flops. Are we really sure this, you know, merits these kind of fancy evaluations? Maybe we should raise the threshold to something more significant. They can do that, but it does require the legislation being updated. So the executive branch can't do it by themselves. The legislature will have to update the threshold, but the executive branch is instructed to make recommendations on that on an annual basis.
- [23:20]
B
Okay, thanks for explaining that. It's really helpful to understand. I kind of want to shift gears a little bit and ask you about what other interesting policies sb53 introduces and what else you found interesting about these transparency acts that you want to highlight.
- [23:38]
A
Yeah, so I think there's a. There's like, so many things that it's just sort of occurring to me that we're going to know in. In not too long. You know, one thing is that it instructs everybody to make clear in their published safety frameworks whether and how they are drawing upon national standards and international standards. So there are some really big standards that loom large in industry, and that industry has been a part of creating. And that. That, to me, you know, one that stands out, of course, is the NIST AI Risk Management Framework that nist, you know, used an industry consortium and an academic consortium took a long time with a big working group trying to come up with that. It was roundly praised by many, many companies. And now these transparency requirements should give us, you know, a novel degree of insight into what actually implementing that looks like in practice. The other thing is it's trying to guide standardization of this transparency reporting. So, you know, again, you could have a company like Evil AI Co, which says, like, here's our published safety framework, and all it says is, we are really safe. You know, like, you know, it's basically a marketing document as opposed to a substantive process implementation document. And I think the hope here is that there's going to be sort of a growth in industry consensus either around these kinds of standards or industry best practices for that. So we can sort of say, this is what good looks like. This is, you know, not what looks like. We're also going to learn about implementation of things that are international. So a lot of companies have signed up to various parts of the EU AI act code of practice. What's interesting is, you know, even though this is a California piece of legislation, the companies are required to disclose how they're going about implementing that. So we're going to learn what it looks like to a greater degree than I think we probably would have otherwise, what it looks like to implement the EU AI Act Code of Practice, which is a big, you know, important regulatory piece. I talked about how the teeth of this law are comparatively dull because it only has $1 million fines. Well, the EU AI act, you know, some provisions of it can fine you up to 7% of annual turnover, which is, like, vaguely analogous to revenue. So that is big, very, very, very sharp teeth. And what this means is that California government officials are going to have a remarkable amount of information about what's going on inside of these companies in a way that, like, Europe's government is going to get a lot of information around what's going on inside these companies, but the US Federal government not. And so I think that's one of the interesting outcomes of transparency provisions like this is I think we're just in a position to see under the hood in a way that we couldn't. We'll talk about this more, I'm sure, but I also think it's worth noting that like Anthropic, which is a very prominent frontier AI model developer, very publicly and loudly supported this legislation. They almost supported SB 1047, although they said, you know, it needed amendments. But this is a company that was founded in part by a group of people who quit OpenAI over debates about issues of AI safety and whether or not OpenAI was going far enough to be safe. Anthropic CEO Dario Amadai has sort of said that he wanted to create a race to the top in AI safety. And it's interesting, you can sort of see their fingerprints on this legislation, because in one mind, an analogy that comes to mind for me is Amazon. Amazon, for a long time was not supportive of increasing the federal minimum wage. And then Amazon was having a lot a big problem recruiting enough people to staff its fulfillment centers to handle, you know, like the Christmas holiday rush. And so one day, Amazon increased their corporate minimum wage to $15 an hour, which at the time was more than twice the federal minimum wage. And what's so interesting is that as soon as they did that, they said, and also we support increasing the federal minimum wage to $15 an hour. Right? So they're like, once we're incurring the costs already, we want all of our competitors to also be incurring the cost. And I think one way of thinking about this piece of legislation is at least from. From the perspective of a company like Anthropic is we are bearing a lot of costs trying to be safe to and there are reasons why we do that that are aligned with our corporate best interests. Right. Which is basically reliability and safety have very similar technical requirements. And customers who want reliability are going to probably like a lot of the features that Anthropic has deployed. I think would be their sort of self interested argument for why they care a lot about safety. But there's also the societal part of AI safety, like the risk of accidentally helping bad people create bioweapons or really powerful cyber weapons, weapons, et cetera, et cetera. And their Anthropic doesn't want to be alone in incurring those costs of safety. They want their competitors to also be incurring those costs and taking safety seriously. And so that's how legislation such as this doesn't force every company to be way, way safer, but it does help force, I think it's fair to say, companies be honest about the extent to which their safety practices are less robust than hypothetically somebody like Anthropic.
- [29:34]
B
So I'm seeing a lot of comparisons and similarities between this policy and the EU AI act, which is obviously highly debated in the US right now. So how have AI companies and other stakeholders reacted to SB53 so far?
- [29:54]
A
Well, I think there's been a diversity of opinions, but all of them are best understood in, in comparison with SB 1047. So there are people who opposed SB 1053, people in organizations who opposed it. So for example, the venture capital firm A16Z, which is headed by Mark Andreasen, they've had people come out opposing SB53 saying it's still too onerous, it's still too early to have this kind of regulation and outright opposing it. But what's interesting is that there was a lot of people who opposed SB 1047 who were silent on the SB 53 debate. So I think you saw companies like Google, which explicitly opposed SB 1047, saying it, you know, goes too far, it's too much, and then were silent in the SB53 debate, which I think is noteworthy. And then you also have folks like Anthropic who are loudly endorsing it, and they were close to supporting SB 1047, but did not, and they have this time. And I think perhaps one of the most interesting endorsements in the entire conversation comes from Dean Ball, who was previously a senior advisor in the Trump White House. And he, he had opposed SB 1047. He had been a part of debates in the US government and especially the Congress about state preemption of AI, or sorry, federal preemption of state level AI regulation. And he had some pretty kind things to say about SB53. So here's one quote that he posted on X. It is rare that a state law introduces a genuinely novel legal mechanism, but the latest version of California's frontier AI safety bill, SB 53, does just that. SB 53 outlines a mechanism whereby the state government can designate a federal law, regulation or guidance, even if the federal thing doesn't preempt as meeting the standards set forth by state law and thus allow companies to opt in to complying with state law via a federal alternative. Really interesting. Almost as though even California is saying, please, federal government make some federal standards for AI. And I think that line is so interesting because some of the criticism of SB53 is like, hey, AI is best regulated on a federal level. And what's so interesting is like the companies who are saying AI is best regulated on a federal level. California, which through this legislation has some of the most important state legislation, is saying, we also agree it would be best if this was regulated on a federal level. And we are creating hooks in our mechanism that can allow federal regulation to supersede this. And they almost encourage that from being the case. Now one final thing I think is worth saying from Dean Ball. It comes from a substack post where he basically said, I found myself opposed to California's frontier AI safety bill, SB 1047. And then he goes on to say, my confidence in the catastrophic risk threat model improved. And a year later I find myself supportive of California's SB 1047 SQL, SB 53. So he's straight up endorsing it, which I think sort of says, like, this is a pretty modest set of requirements going on companies. One other voice that I think is important in this conversation is Senator Ted Cruz, who had previously said, quote, do you really want Gavin Newsom and Karen Bass and Comrade Mandani in New York City setting the rules for AI and governing A across this country? I think that would be cataclysmic. Cataclysmic. And so he's talking about how, you know, we should not be rising to the most severe regulation that is coming out of states. We should have a federal standard, which I think most people in this story agree with. So that's all very interesting. One thing that we haven't talked about in terms of this legislation that I think is really important is the whistleblower provisions. So the idea here is that if There are catastrophic risks presented by an AI system. You want not just to say, hey, companies, you have to tell us when you have these incidents. You also want a mechanism that can deal with evil AI Co. Right, the kind of company who's not going to tell you about those incidents. And so it creates a framework whereby individuals in those companies who have privileged knowledge of safety incidents can violate their non disclosure agreements, which almost all of their company, all these companies make their employees sign and tell the world, hey, my company is doing something that is very, very unsafe. And they can do that in a way that that information gets sent to the California state government that their identity is protected and that their legal obligations are covered. Covered. And that is something that I think a lot of people had supported in SB 1047. And so I was not surprised that it made it through to the SB53 as well.
- [35:11]
B
Okay, so we've heard a lot of great, a lot of great information about this bill. So what should we keep an eye out for as California moves to implement this policy?
- [35:20]
A
Well, I think there's this bill which is going to go into effect quite soon, January 1, 2026. There's also two other related bills that have come out of California. One is around transparency related to training data, which I think will be a big part of the debate on copyright type issues and intellectual property protections. It has relevance to the safety debate of course as well, but that could mean we'll get an unprecedented degree of information around the training data going into these models. And then there's an additional California law that's dealing with deep fakes and the risk of AI generated misinformation. And so that is requiring model developers to have capabilities to sort, sort of disclose to the user when they are interacting with AI generated content. We're already dealing with a flood of new AI generated content, not least of which with the announcement from OpenAI of its latest version of Sora, which has really impressive capabilities, especially around preserving an individual's face across a really diverse range of potential situations. And so having the models sort of embed like, this is AI, this is AI into that I think is going to be really interesting. And California is one of those states that has been putting regulatory muscle behind that.
- [36:42]
B
Well, thank you, Greg, for breaking down this new policy. I appreciate you answering all my questions. I'm looking forward to talking more about this in the coming months. But thank you so much. Thanks everyone for listening.
- [36:55]
A
Great. Thanks, Sadie. Thanks for listening to this episode of the AI Policy Podcast. If you like what you heard, there's an easy way for you to help us. Please give us a five star review on your favorite podcast platform and subscribe and tell your friends. It really helps when you spread the word. This podcast was produced by Sarah Baker, Sadie McCullough and Matt Mann. See you next time.