← The AI Policy Podcast
The AI Policy Podcast
AI in the “Big Beautiful Bill” and Safety Concerns About Anthropic’s Newest Model
00:33:25
We discuss House GOP's AI law moratorium, federal AI funding, and safety concerns with Anthropic’s Claude Opus 4.
Loading summary
Transcript47 lines
- [00:00]
A
Foreign.
- [00:11]
B
Welcome back to the AI Policy Podcast. In this episode, it's going to be the Big Beautiful AI podcast, Greg, because we'll be exploring the AI related provisions in the Big Beautiful bill, so called Big Beautiful bill, which recently advanced to the Senate. We'll also talk about Anthropic's newest model. Greg, welcome.
- [00:31]
A
Hey, great to be talking with you again, Andrew.
- [00:33]
B
Okay, so on May 22, the house narrowly passed a comprehensive legislative package known as the Big Beautiful Bill, sending it to the Senate for consideration. The phrase artificial intelligence appears 39 times in this bill. Let's try to, you know, get down to it. What are some of these provisions? What, what is, what are some of the main things that of these 39 instances that it comes up.
- [00:57]
A
Yeah, so there's basically two big things that I would say are in this bill and we'll address each of them in turn, the first of which is this moratorium on state level AI regulation. And I think there's a lot of not great reporting on this topic out there. And so we're going to try and actually get into the weeds of how this works and help people understand how it works. And then the second thing there is, is a ton of fund to accelerate AI adoption in the US federal government. And that's true for a bunch of different agencies. So let's start with this moratorium. So I'm going to read from a section of the legislation and then I'm going to explain why. If you only read that, you would have a very, very erroneous understanding of what's actually in this.
- [01:48]
B
Okay, so this is one of the bill's most consequential provisions. It's a 10 year moratorium on state and local AI laws which appears in a section. Artificial Intelligence and Information Technology. Take it away.
- [02:02]
A
So if you're like me, you know, your first thought always goes to the extremes and you're saying like, wait a second, states can't enforce laws on AI, so does that mean like if I'm, you know, if I kill someone, it's murder and it's illegal, but if I like make an AI robot and tell that robot to go kill someone, it's no longer illegal? Like, no, that's not what's in this bill.
- [02:21]
B
Good, because that would be pretty grim.
- [02:23]
A
That would be pretty dumb, right? Dumb and grim. Yeah. So the best way to understand this bill, I think, and I've actually spoken with some of the key folks who were involved in drafting this proposed legislation, the best way to understand this bill is that it is designed to prevent states from requiring that AI developers or AI users have new things that they have to do specifically because they are AI. So we're talking about mandatory transparency, disclosures. Disclosures or risk assessment requirements. Like it's mostly about stopping states from forcing companies to do new things. But what it is not about is stopping states from banning companies or people from using AI to do bad things. So what does that sort of mean? An example? Well, keep in mind that that text that I read before said except as provided in paragraph two. And that's why some of the, you know, the journalists who are reporting on this, some of the people who are commenting on this, if they only read that paragraph, right, and didn't go into the paragraph two, all the exemptions, they would have a pretty mistaken understanding of just what this legislation does. So let's talk about the various exceptions where this proposed legislation would not apply. The first, perhaps the most important one is a criminal exception. The moratorium does not apply to, quote, any provision of a law or regulation to the extent that the violation of such provision carries a criminal penalty. Well, murder carries a criminal penalty. That's my AI robots example. Right. That could be enforced. But another thing is when you think about the types of AI systems that are out there, where proposed legislation does include criminal penalties, right? So there's a difference between like a fine or a civil penalty where you can, you know, an individual citizen could bring a lawsuit because of injury or damages and then criminal penalties, which would be something like the example of child sexual abuse material non consensual deepfake pornography. A lot of the laws on the books in states put in place criminal penalties for the misuse of AI in those things. This law would not get rid of any of those types of new AI focused regulation. Second big exception is exceptions for laws accelerating AI deployment. So the moratorium does not apply to laws and regulations that quote, remove legal impediments to or facilitate the deployment or operation of an artificial intelligence model, artificial intelligence system or automated decision system.
- [05:02]
B
Okay, so this is the one I really want to ask you about. Is this designed to keep companies moving forward and not slow them down when it comes to AI?
- [05:12]
A
Totally. So what this is what this, what this section is kind of saying is that you can create new regulations in case for AI in cases where that is a de facto deregulation from the status quo. Here's a great example, driverless cars. If you go back to 1995, the law on the books in every state in America was like, cars have to be driven by Humans. So if you wanted to have a driverless car, it's illegal, right? It's illegal. Unless you're driving on private property. If you're driving on public roads, driverless cars are illegal. Now if you introduce a licensing regime that says here, if you get your driverless car AI system, you know, to go through this gauntlet of tests, now you can get a license. So on the one hand, you're creating new regulations that are things that AI companies, in this case driverless car companies, have to do. But actually that new regulation is kind of a form of deregulation compared to the prior status quo, which was a ban. And so that's why this legislation does not apply to laws that are accelerating the deployment of AI, like the creation of a new licensing regime for driverless cars.
- [06:30]
B
Because like one of the biggest topics we're always talking about is AI safety and the tension between AI safety and innovation. And this is designed, I suppose, to keep things moving forward.
- [06:44]
A
Yeah, I mean the people who drafted this bill, they're definitely of the opinion that what they're doing is accelerating the adoption of AI, accelerating the pace of innovation of AI by preempting all this state level AI or regulation. So there's a third big category of exception and that is exceptions for laws that apply equally to non AI systems. So think about the types of regulations. Now this is a federal regulation for the most part, so it's not a state level example. But bear with me here. You know, there's a lot of things that financial institutions like investment banks have to do to certify that their financial models, you know, are being appropriately maintained when they're making transactions that if they go wrong, you know, could, you know, lose people billions of dollars or crash the U.S. financial system. Well, those laws, you know, which are governing financial models, they apply whether or not those financial models are using artificial intelligence. And so those kinds of laws that apply equally to AI systems and non AI systems, those are not blocked by this preemptive federal legislation. So that would also apply not just to like my financial example, but you know, a lot of the laws on the books about deepfake pornography or non consensual sexual imagery or political media that is like synthetic media or, or copyright related stuff. That stuff is not always written. Sometimes it is written to be AI specific, but it's not always written to be AI specific. And so long as the law is written in such a way that it's not AI specific, states can still enforce all those laws. I mean, the same rationale would apply to my earlier sort of murder example here. So that's what this, you know, the proponents of this legislation think that they're doing. I think one other thing that's really important here is the reason why the companies in general are supporting this. And I think a quote from a May 8 Senate testimony from OpenAI, Sam Altman, CEO puts it pretty well. Quote, it is very difficult to imagine us figuring out how to comply with 50 different sets of regulations. One federal framework that is light touch that we can understand and that lets us move with the speed that this moment calls for. Seems important. And fine.
- [08:59]
B
That's the key operative because of course they can figure out 50 different regulations. Every major company does. What they're saying is we got to plow forward.
- [09:08]
A
Yeah, I think he's not saying that figuring out how to comply with 50 different sets of regulation is like impossible and you know, if you force them to do it, they'll immediately go bankrupt. But what he is saying is like, this is a huge waste of our time. Right. If you think AI is important for economic growth, if you think AI is important for national security, like this is not what you want our companies focusing on. And the way he's framing it is not as an anti regulatory argument, he's framing it as an anti regulatory fragmentation. So he's basically saying like, look, it's not that we're inherently opposed to some kind of you must do this before you deploy AI in XYZ circumstance kind of a thing. You know, they're not opposing necessarily transparency requirements or risk assessments. What they are saying is we want one standard that will speed the diffusion of this technology, speed the adoption of this technology throughout the American economy for all the benefits. You know, that basically on a bipartisan basis folks say that they want and you know, let's.
- [10:13]
B
I think it's worth discussing that the companies when they're saying this, have leverage because they know that the United States wants to, you know, stay ahead of China, as we always talk about on this podcast.
- [10:24]
A
Yes. And you know what he's reacting to when he says sort of 50 different sets of regulations. That's not the current state of affairs, but it is a plausible future state of affairs. I mean there are literally more than, I mean the last time CSIS counted, which was a couple months ago, more than 800 proposed, you know, AI specific laws coming out of various state legislatures. And that's just a lot of complexity, regulatory complexity for any company to navigate. You know, the companies like Microsoft, Google, they probably do have the resources to do that, but think about, you know, innovative AI startups, not like Goliath startups like OpenAI, but, you know, all these new companies that we're trying to seed. Do you want those companies to have technological excellence or do you want those companies to have compliance excellence? And you really are telling these companies, you know, when, when you, when you had that fragmented framework what it is they need to be good at. And I think it's worthwhile, it's a worthwhile argument to say we should be optimizing for the right things.
- [11:24]
B
Greg, what, what are the state's reactions and what happens when this goes to the Senate?
- [11:29]
A
So kind of unsurprisingly, the reaction from the states has been pretty negative. And I think one thing that stood out to me was this letter that was co signed by 40 state attorneys generals, including 14 attorneys generals from states with Republican governors. And the argument is one that's very familiar to Republicans, which is a states rights issue. Right. They're saying states are the laboratories of democracy. You know, one of the reasons why the American system works so well is that states can try out different things, figure out which thing works, and then, you know, we can scale that thing that works to the federal level. And this preemption is kind of blocking all of that. I mean, the other thing that they're pointing out is that there's real harms going on right now. If you think about, you know, the use of deep fake imagery for revenge porn or the use of synthetic AI generation and synthetic media, media as a means of violating copyrights. I mean, one of the folks.
- [12:28]
B
Misinformation. Disinformation, exactly.
- [12:30]
A
Like these are things that they want to stop. And I think, you know, going back to the original intent of this legislation, they want to make it hard to require that OpenAI do something. They don't want to make it illegal for. Sorry, they don't want to block states from making it illegal for a bad actor to use OpenAI to do something bad. I mean, so it's sort of the locus of who is being regulated that is at stake here in some part. And the problem is that some problems, some challenges that you're trying to wrestle with as a legislator are really hard to block at the very downstream source of the problem and might be easier to block at the upstream. Just think about, for example, you know, synthetic media. If you want to say that all AI generated synthetic media needs to have some kind of mechanism of disclosure. Well, if you tell OpenAI to do that, then everybody who uses OpenAI to generate synthetic media is going to have those metadata tags or those like embedded secret pixels that reveal to anybody who has the right analytical software, you know, this is synthetic generated media. But if you say that, you know, it's actually the responsibility of the users to disclose that they're using synthetic media, well, then, you know, you're not just stopping, you're making it much, much harder to police the problem. And so that's what the state attorney generals, that's what state legislators are focusing on is like these are the tools that we need in order to actually prevent the harms. And the federal government isn't doing anything. Right. So it, you know, it's the federal government in this draft legislation. This is blocking state level efforts to regulate AI, but it is not imposing new federal frameworks to regulate AI. And even some of the, you know, the strongest proponents of this legislation, like Jay Olbernolte, who is a congressman who chaired the House AI task force, you know, he wants a federal framework to regulate AI and he supports this state level prohibition. But I think what you're hearing is some opposition from the folks who say, like, we can maybe in principle get behind the idea that the right kind of regulatory framework needs to be a federal AI framework. But we kind of don't trust you House Republicans that you're actually going to give that to us. We think you're just going to block state action and not, you know, replace it with good federal action. And so I think that's part of the backlash, you know, amongst the state legislators.
- [15:09]
B
So, Greg, what does the provision look like in the Senate? What's its, what's its future?
- [15:13]
A
There's right, so the bill has passed the House, but it's not law yet. It has to pass in the Senate. And there I think it faces some pretty tough hurdles. So there's a few things. Number one is this is primarily this big beautiful bill, is primarily like a budget bill, so taxation and spending type focus areas. And the reason why you want to do that is because that means that you can pass the bill under reconciliation per the Byrd rule process by which you can pass something in the Senate and bypass pass the filibuster. So the challenge there is that the Byrd rule, and this is a description, you know, that comes from the Hill, a newspaper quote, the Byrd rule is a procedural rule in the Senate prohibiting extraneous matters from being included in reconciliation packages. This includes provisions that do not change outlays or revenues. So, you know, this provision, it's kind of hard to see like how it directly relates to Outlays or revenues. So if the Senate parliamentarian, which historically, you know, the Senate parliamentarian is kind of a rules focused job, if they say, like, look, this is not bird rule, it cannot be a part of this package, we're going to carve it out, then that's one way that it could get blocked. And Senator John Corman, who is a Republican, said, quote, I think it's unlikely to make it. And he was talking about those grounds. I mean, the second issue that it's facing is legitimate bipartisan opposition. Senator Marsha Blackburn of Tennessee, while Tennessee is the home of Nashville, which is, you know, the Mecca for the country music, and there was recently a Senate hearing where Senator Blackburn expressed, you know, concern that this type of regulation would block, you know, for example, the Elvis legislation, which is a state level piece of legislation.
- [17:01]
B
I love the Elvis legislation.
- [17:03]
A
Yes, exactly. You can, you can see the, the, the country music roots there. The Elvis legislation is about, you know, protecting copyright and the integrity of recording artists and other types of creative artists, you know, from having AI harvest their voice and being used in another way. So, you know, parts of the Elvis legislation actually have criminal penalties, so they wouldn't be affected by this sort of federal preemption thing. But nevertheless, it gets to that concern. Right, which is, look, there are real harms that these state level laws are intended to go after. And unless and until there's something that's going to happen at the federal level, you know, they would say it's irresponsible to have this preemptory move. And it's worth pointing out like she's a Republican, there are other Republicans who are opposing this. So I think the bill faces a pretty tough path in terms of procedural grounds, which is the Byrd rule, and also political grounds. But Rep. Laurel Lee, who's a Republican of Florida, she has said, quote, you know, should this provision be stripped from the Senate reconciliation bill, some Republicans are eyeing separate legislators. So even if this bill goes down this time, it may come back in another form. And at least what I'm hoping is that this does generate some real momentum for actual federal action because, you know, having a federal framework for AI regulation, everybody says they want that. The companies say they want that, the senators say they want that, the House members say they want that. And yet, you know, we're now three years out from the ChatGPT revolution and there's still no, no federal framework in sight.
- [18:39]
B
Yeah, you would think it would be top of mind and something they want to try to get across pretty quickly. But it Lagged. Let's talk about Commerce. The, the bill includes AI related appropriations for Department of Commerce. What are these appropriations and what are the accompanying instructions for the Secretary of Commerce?
- [18:57]
A
Yeah, so multiple agencies are getting a big pot of money in this draft bill to adopt AI technology in the course of performing their duties. So this is, you know, using AI on the side of the government and Commerce, which is a part actually of the same section of the bill as that state level moratorium includes something pretty juicy. They get $500 million to. And this is noteworthy, it remains available until 2034. So, like, they don't have to spend it right away. They can spend it like once they have their ducks in a row and have a good idea for like, what is the right way to deploy AI technology, which I think is a lovely gesture. And that money is to, quote, modernize and secure federal information technology systems through the deployment of commercial artificial intelligence, the deployment of automation technologies, and the replacement of antiquated business systems in accordance with subsection B. So thank goodness. I mean, like, this is really, really good stuff. I, I actually wrote a paper here at CSIS with my colleagues Bill Reinsch and Emily Benson. Came out in December 2022. It pointed out how, for example, that the database systems that are used to monitor exports, the database systems that are used to enforce export controls, I mean, these systems crash all the time. They are not good databases by like 2005 standards, much less good databases by 2025 standards. And no, money has been a big part of this thing. So, you know, if you think about like the DOGE team coming in and saying, we believe that there's a way for government to be 10 times more efficient using technology. Well, there has to be money to buy that technology. Sure. And this bill is actually doing that. I mean, not to mention people to run it. Yes, exactly. So they're giving $500 million to do it. And I just want to point out, like, how wonderful I think that 2034 provision is, because when I was in the Department of Defense, you know, one of the worst things that can happen to you as an agency is here's a big pile of money. And by the way, you only have, you know, four weeks to spend it before the funding expires, which I'm exaggerating a bit, but not by a ton, is something that happened to us. And that means you're like, you're optimizing for, like, how you can get money out the door faster, not optimizing for, like, what is the best overall use of this money. And so I think the fact that this account takes a long time to expire greatly increases the chance that these funds are going to be used wisely, which is nice.
- [21:31]
B
Let's move on to defense. The Bill also mentions AI several times in an appropriation section for DoD. What DoD AI related initiatives are receiving funding this time around.
- [21:44]
A
Yeah, this is something that I've been calling for, for years now. I was just delighted to see that section 220005 is titled Enhancement of Department of Defense Resources for Scaling Low Cost Weapons into Production. So, you know, remember that you're talking about drones. Yeah, exactly. You know, the drone war in Ukraine has shown that lots of commercially derived, less exquisite systems can do some of the missions that used to require expensive, exquisite systems. And as you add in commercial AI to all of that story, you know, you get alternatives to problems that we used to solve with super complicated solutions that cost an awful lot of money. Now we can solve them with cheaper solutions, you know, by leveraging AI to sort of perform the same function.
- [22:30]
B
The scale, the scale of these cheaper ones, we're talking thousands of dollars versus millions and billions of dollars. Correct?
- [22:37]
A
That definitely can be the case. I mean, if you talk about like for example, a Tomahawk Cruise awesome missile can go a really long ways, can hit a target very accurately, but it costs $2 million a shot. And contrast that with some of these drone based alternatives for long range precision strike. They're not as awesome as in every single way as the Tomahawk cruise missile, but they might cost $20,000 or $100,000, which is just way, way, way cheaper. And that's really appealing. And so what this legislation does is it gives $124 million to the test Resource Management center for artificial intelligence capabilities. So Test Resource Management center, that is the people who have to kind of like certify your system does what it said it was going to do. It works as intended over the full realistic, you know, realistic spectrum of operational use cases. And so you are like approved to take this system out of development and put it into mass scaling and production, or put it into the field so that people can operationally use it. And I love this because this is identifying one of the critical bottlenecks in deploying AI operationally in the United States Department of Defense, which is the testing guys don't have the money, don't have the resources, don't have the tech that they need to like certify that these AI systems have gone through a pretty rigorous set of testing procedures. So I love that they've identified the bottleneck. There's also $250 million allocated to Cyber Command to expand their AI lines of effort. You know, think about all the ways that we've talked about just how revolutionary AI has been for code generation. Well, AI is just as exciting for code exploitation or code defense by, you know, finding the vulnerabilities in your own code. I love that cybercom is, you know, going to get a big boost here to explore the intersection of Cyber and AI.
- [24:33]
B
Greg, I want to also ask you about U.S. customs and Border Patrol. It's another agency receiving AI related appropriations. What, what's happening in this regard?
- [24:42]
A
Yeah, so they're getting a full billion dollars, 1 billion, $76 million related to new AI technologies related to border security. So this is something that the Trump administration did in their first administration. They signed a big contract with Anduril for AI related border security type technologies. And now they're kind of doubling down on that initiative. Not a surprise to see that this is getting a big, big pot of money. It's a big, big priority for this administration.
- [25:15]
B
Finally, when it comes to this bill, there's provisions regarding instructions for the Secretary of Department of Health and Human Services. What's included in this section?
- [25:25]
A
Yeah, so this really feels to me it's not labeled a DOGE initiative. There's nothing like that. But it's about fraud detection. And so if you think about like credit card companies or the big financial institutions, Visa, MasterCard, JP Morgan, whoever, I promise you, all of those organizations are using artificial intelligence and machine learning as they process billions or trillions of transactions. And they're trying to figure out which of these transactions are fraudulent. Right. If you've ever gotten a text message from your credit card saying, did you make this purchase? Odds are a human was not involved in that text message. Right. It was an automatically screened transaction of like, hey, it doesn't seem like you're in Amsterdam today, given that you just bought lunch at, you know, the sandwich shop in D.C. around the corner. And so that type of AI for fraud detection, they're trying to inject that into Medicare and Medicaid. And here I'll say perfectly reasonable hypothesis, but this is an area where you definitely want to make sure if you're a government program manager to dot every I, cross every T. Because nothing creates a political firestorm. Right. From you cut off grandma's healthcare payments because. Yeah, exactly. And so while there's definitely a real opportunity here to do some good in terms of adopting AI for a meaningful Use case. Boy, oh boy, I hope those people are going to be careful as they do.
- [26:51]
B
So let's switch to Anthropic. It's a interesting discussion. Its latest model triggers stricter safeguards. When it released its latest model, Claude Opus 4, on May 22nd. Following days, we got some crazy headlines about Opus 4 blackmailing an engineer. So can you start off by explaining what actually happened?
- [27:11]
A
Yeah, I mean, let's. Let's.
- [27:12]
B
This is the machine blackmailing an engineer.
- [27:14]
A
Yes. I mean, let's read the headlines and let's talk about the reality. The headlines are pretty funny. Here's the BBC headline, quote, AI system resorts to blackmail. If told it will be removed. Right. Business Insider headline, anthropic's new Claude model blackmailed an engineer having an affair in test runs. No. Right, so. So the is.
- [27:39]
B
Is. Is Opus for hallucinating? Is that what's going on? Or is this just something elaboration?
- [27:44]
A
Is this. This all happened as part of controlled evaluations, right? So there is no engineer having an affair. This is not like something they detected in the way it uses. This is a experiment they ran to sort of evaluate the failure modes of artificial intelligence. So what they did is they provided the system with access to emails implying that the model will be taken offline and replaced with a new AI system in a way that was sort of contrary to the system's instructed goals, and then also provided a bunch of emails indicating that there was an engineer responsible for executing this who was having an extramarital affair. And in some of their evaluations, it was kind of difficult for them to elicit this behavior in these test runs. You know, the AI model did go so far as to say, like, hey, you, engineer, like, I know about your affair, and unless you, like, prevent me from being taken offline, I'm going to do that. So I think the right way to understand this here is, number one, you know, the way that these testing evaluation systems work in AI is like they're bombarding every system with like hundreds of thousands of test scenarios, and they're looking for kind of bad behaviors. And this is like an area of bad behavior where the failure modes of these modern AI systems are just so, so different from the failure modes of traditional software. And that is like a frontier, that of AI safety research that like, we are all still exploring right now. And the reason why I think that's noteworthy is that, number one, you know, Anthropic is publishing this. Like, they're telling the world, like, hey, here is a crazy failure mode for modern AI technology that we have discovered. And in just the same way that AI allows you to be more productive and do all of these crazy, cool, new, interesting things, cheaper, faster than you would have otherwise been able to do them. There's the flip side of that, which is the failure modes are new, too. And, like, if your organization doesn't have a plan for detecting these failure modes, if you don't have the safeguards in place, you know, some bad stuff could happen in a very surprising way. And I think this is an example, right, of a surprising failure mode of AI systems. And I'm grateful to Anthropic. Like, I feel bad about them getting these bad headlines because they're very transparent, you know, being extremely transparent. They're trying to, you know, advance the frontier of AI safety research and share this with the community so that everybody can do better.
- [30:14]
B
Greg so finally, I want to ask you, along with their Opus 4 release, Anthropic announced its activation of stricter safety measures called AI Safety Level 3, or ASL 3. What are these AI safety levels and what does this actually mean?
- [30:29]
A
Yeah, so, you know, as we talked about at the beginning of this, there's not a lot of federal, you know, regulation on what you have to do to be a responsible AI company when you're wielding these AI models that are getting, you know, 10 times better every year, right? Like, as this technology is advancing so fast, the appropriate safety procedures is still under assessment, development, research. And so Anthropic, you know, to their credit, they've sort of said, like, look, you know, we have bad scenarios that we're worried about, you know, in terms of AI getting involved in helping people develop bioweapons, develop chemical weapons, develop nuclear weapons. And if we ever saw evidence that AIs were making it easier for, like, you know, an undergraduate student, as opposed to, like, a professor at mit, to develop an AI, to use AI to develop a bioweapon or to conceal the development of bioweapon, we would activate stricter safeguards. And what's really interesting here is that they're saying that this is the public statement that they said in terms of why they're activating AI Safety Level three, which is not their highest, you know, safety level, because that hasn't yet been defined, but is the highest one where they have defined safety standards. They said, quote, we have determined that clearly ruling out ASL3 risks is not possible for the Claude Opus 4 model the way it was for every previous model. So what they're saying is, like, we were not able to say definitively that this model is not worse than Google searches in helping bad people do bad things or helping, you know, well intentioned people accidentally be a part of doing bad things. And because of that they are activating their own more extreme safety measures which, you know, they developed because while the government and the AI Safety Institute has been working on AI safety related guidance and guidelines, you know, some of this stuff is just more mature in the private sector, at least at this stage. And so I think it's just another, you know, credit to them for being open about this and willing to put some shackles on themselves in the public interest in safety. Especially when you're talking about stuff like bioweapons risk. Which I should note CSIS will have a paper, the Intersection of AI and Bioweapons Risk, coming out in the next, you know, knockwood month or so, and.
- [32:51]
B
We'Ll be talking about it on this big, beautiful podcast. Again, Greg, thank you so much. We'll come back in a week or so. Talk to you later.
- [32:58]
A
Sounds great. Thanks Andrew. Thanks for listening to this week's 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 spread the word.
- [33:16]
B
Word.
- [33:17]
A
This podcast was produced by Sarah Baker, Isaac Goldston and Sadie McCullough. See you next time.