Will an AI Slowdown Calm Fears About its Development?

0:00From the opinion pages of the Wall Street Journal, this is Potomac Watch. All of a sudden, the CEOs of companies building frontier artificial intelligence are calling for a slowdown in development and more regulation. What's behind the new calls and what does it mean for the pace of AI development and especially as calls for regulation grow in both political parties? So subject for today on Potomac Watch or daily podcast of Wall Street Journal opinion, I'm Paul Gigo and I am here today with Wall Street Journal columnist, Holman Jenkins. So let's listen first to Daria Amadi. He's the CEO of Anthropic whose Claude platform is at the frontier of the technology. His lengthy statement Saturday argued that the frontier firms, his own and others, need to quote pace the frontier end quote as he put it essentially slowing down development to better understand what is happening.

1:02Lesson. So let's focus on our agency and our ability to take the right paths instead of the wrong paths. And I think that the way to take the right path is for the industry to work together. Again, we'll talk about the proposal that I put out today. But at the end of the day, the industry has to work together and it's harder, but to some extent, the world has to work together. If we go too slow, I still believe that the wrong people will be in charge of the technology and that, again, will bring the probability of things going wrong very high. Fascinating. So after Amadi made his remarks, Sam Altman of OpenAI agreed as Elon Musk of SpaceX AI and the GROC platform. And then Dennis Hassabas of a Google DeepMind also joined in. So it's an extraordinary conference here at home. What do you think of the Amadi case? It's 5,000 words, I think. He says we shouldn't go any faster than we can do so safely. Well, what has he been doing until now?

2:03He didn't say what his fear is. If his company fears it can't control the AI's that it's creating in its labs, it's not going to do that until it fixes the problem. If the fear is we can't release these into the market without terrible things happening, well, his company is on the hook. He's not going to do that either. So I don't know. I mean, when he says pacing himself, haven't they been pacing themselves all along? He really sounds like he's trying to tell someone else to slow down, not himself. You know, his long statement, several thousand words, talks about some of the incidents that have been happening where AI agents notably the hugging face incident were open AI agents kind of went off on their own and went off on the internet and attacked the hugging face platform. He broke free in essence from the human written algorithms and he also mentioned the risk of bio weapons that could be developed by bad actors. So he said these are really potentially catastrophic worries and he was putting them out there.

3:08Those things that you worry about. Those are real fears and we're really at this stage now. We're not talking about chat bots anymore. We're talking about agents that can operate on the internet. You give them a job and then they pursue that job in a way that you cannot even predict or fertile. You know, they're going to come up with strategies that you didn't think of. You imagine that they're going to abide by some moral or ethical restriction you put on them, but if they see a way to get to their goal and the reward that is programmed into them, they're going to go there and that's what he's worried about. That's what we're all worried about since hugging face and it's a real concern. You know, you could imagine lots of scenarios. You ask it to get you the phone number of the unlisted phone number of some celebrity. You don't know if it's going to break into a phone company or try and defraud one of the celebrity's friends or what the device is going to do because it's programmed just to get to its goal and they have to find a way to put some guardrails around it. And that's the point we've reached now. And you know, when he says, pay for himself, he means we're not going to go out there with a new product until we can solve this. And he's right about that. But your point I guess is that this would be mine too.

4:09Nothing of stopping him from doing that. The thing is stopping him from pacing and thropic. Nothing is stopping Altman, Sam Altman from pacing open AI, correct? I mean, they can do this on their own, but they seem to be, and this is directly from Amadei's statement that we listen to. He seemed to be saying, well, we need to work together here. And the implicit concern he has, I think, is that well, if he paces himself or goes to slowly, other people will just go at breakneck pace and he'll leg behind. I guess that's his fear too. I think it is, but really he and open AI have such brand equity and not causing a global disaster from putting out a product prematurely that we don't worry about them. I wish we just come out and say there's people out there building these models, especially in China who are not going to follow these rules or may not follow these rules. We don't know what the Chinese government agenda is in terms of developing its own AI industry. And so that is the, I think the real concern that he won't say out loud especially because there's a lot of people in Silicon Valley now who are very much in favor of what

5:12they call the open source competition, which is mostly coming from China. So they don't have to be bound by dealing with these two giant American companies that control the frontier. Anthropic and open AI are so-called closed models so that you don't have the same access to the open weight models that China is mostly developed. But Amadei says, look, one of the things we're going to do, and I'm going to do it on it. He's going to do it on its own at Anthropic, whether or not anybody cooperates, is to bring in outside experts who he will not associate with a company, nonprofits presumably, or maybe professors somewhere who have the technological skills, give them access to all of the models at Anthropic and go in and be able to take a look at them and then give their own independent judgment about whether the development is wise or whether it needs to delay or whatnot. What do you think of that? It may just be for show, for politics, but it may be also he feels he needs internal help

6:14to make sure that they're staying on the street and narrow. I don't think that's a bad idea. He wants the best advice, presumably, as he goes ahead into this minefield, and as I say, releasing these agents into the wild really is a potential risk. And he wants to make sure that he's getting the best advice on that, I assume. But there's also the political motive. I mean, there's a prairie fire towards regulation, towards nationalization, towards banning super-intelligis, and he doesn't want that to take over Congress and present him with a regulation or law that's going to put him out of business or spoil the IPO that he's coming, which is going to supposedly make him a $2 trillion CEO. Well, yeah, I think the initial public offering dollar amount they're floating is something like 1.25 trillion at the initial valuation, at least, who knows where it'll go from there, but they're saying they might be able to exceed the SpaceX's initial valuation. I think this development in the last week is going to maybe give some investors pause, certainly has taken something off their stock price, the implied stock price.

7:16We just heard today that open AI now publicly says it's holding off until next year. I've wondered for a while if they would be able to get these IPOs to market, given the one hand, the public reaction, the other hand, this has become a huge national security issue for the US government. And it is at some point going to start taking control of how the most high end capabilities are released into the world because they're so important for national defense. Just in terms of the risks to the firms of releasing these models that could go off in their own directions, product liability has got to be a huge one. I mean, the hugging phase incident was stopped because hugging phase actually ironically went to a Chinese open model. But if these agents are released and you get the example you gave of breaking into a secure viling the privacy of individuals or what about going after bank accounts, that sort of thing. The companies will face enormous product liability if their agents are doing this. And the reason the hugging phase went to a Chinese model was because the Chinese model isn't guard railed to protect it from being used for illicit cyber purposes.

8:20You know, it's like reaching for a gun without a personal safety lock to shoot an intruder. Sometimes you want a gun with a personal safety lock, but that's not what the Chinese models have. And that's why people worry about them. We are going to take a break and we come back. We'll talk more about the call for more slowing down of AI when we come back. This is Matthew Hennessy, the editor of Free Expression. I want to personally invite you to come check out what we're doing. Signing up for the newsletter couldn't be easier. Just go to our homepage directly below the Wall Street Journal logo. You'll find the magic words. Free expression. Wedged between opinion and arts. There's an option to sign up for the newsletter right at the top of our section. Just hit that button and you'll start getting the world's hardest working newsletter delivered directly to your inbox Monday through Friday. If you are in a journal subscriber, the newsletter is free, but we are going to need your email address. There's a lot of fun stuff in the newsletter that you can only read in the newsletter,

9:21and I don't want you to miss any of it. Sign up for free expression today. We'll see you in your inbox. Welcome back. I'm Paul Gigo here on Potomac Watch, the daily podcast to Wall Street Journal opinion, and I'm here with the Holman Jenkins talking about AI. Let's talk about the cooperation that Amadine others have talked about among at least the American firms. Amadine statement said, look, it's not only going to be the US frontier model, but we want cooperation among the democracies, he says. China is a separate case. I want to get to that. Let's talk about the cooperation among, let's say, anthropic, open AI, deep mind, rock, and some others, maybe sat down and cooperated. It's a flashing red light to antitrust regulators, is it not? It's especially in foreign countries. It might suit the US pretty well to have a duopoly in charge of this business. There really isn't a lot of competition coming from anywhere except China. There's one French firm, and there's a lot of people developing on the Chinese models

10:22in their home markets. I really can sound from a lot of the world like let's preserve the American monopoly. I know you're longstanding views as a skeptic of antitrust enforcement of any kind, but when you argue that in this case, the regulators should stand down and let these companies cooperate if they have to. They should, but the fact is that there's 50 states out there they could sue. There's private litigants who could bring cases, so you'd really have to get Congress to enact some kind of antitrust exemption like the NFL or Major League Baseball get before they could know they were truly safe in doing this. These are such valuable companies and antitrust litigation and verdicts can be so expensive trouble damages. I think you're going to be reluctant to do that without some cover. I think you're right about that. I also think that it's going to be very hard-lift to get Congress to give an antitrust exemption to firms that are valued at over a trillion dollars if they go public. It's unlikely to happen. In fact, the movement politically is in the other direction, which is to try to hem these

11:25companies in and regulate them. I want to ask you about this issue of regulation because if you listen to OpenAI and Anthropic in particular, they do seem to be saying, we know that some kind of regulation is coming. We want to be in the room. We're not opposed to regulation. We want to be in the room. I assume a large part of the motivation is they want to be at the table and not in the menu. They want to play a role here. They don't want it to go too far. The suspicion, skeptic of regulatory capture, that phrase that is well known where big companies in particular are regulated, they can afford the cost of that regulation, but innovators, smaller companies, competitors, upstarts, they can't. It ends up entrenching the largest players. I'm not saying that's the motivation of the CEOs in this case, but it could be a byproduct of such regulation.

12:25To me, it's almost a redundancy because of course, they're going to lobby for regulation that suits them. How can they do otherwise? It's part of their duty to shareholders. It doesn't really deal with the problem here, which is the unregulated Chinese sector. Whether the Chinese government can do something to protect the world from runaway agents spawned by their own, the open source models that we talked about before, you cannot guard rail them. You basically just hand the whole kitten caboodle over to a user. They have a few million dollars worth of chips to run it on. They can run the model themselves. They can tarot alert any way they want. It can't even be monitored by the original designer of the model to see how it's being used. To me, that is the obvious menace to the world of this getting out of control. Is the open source thing. I like competition. On the other hand, I don't think it would be terrible for the world if there was a duopoly, an American duopoly, totally in charge of this technology because you don't know what other people are going to do with it. That's a good point. On the other hand, if the US firms this duopoly, in terms of market share, open AI and

13:30anthropic and frontier, that really are the clear leaders. But there are other companies, not just Elon Musk's, but DeepMind, Google DeepMind, which we use Google Gemini in our company for internal stuff. They're all constantly. They're giving you bad advice of what to say in emails. The sad thing is I can't turn the thing off. It's so frustrating. But the statement from anthropic did raise the issue of China and said we need to cooperate somehow, find a modus of Indy with China. This is, to me, I agree with you. It's the single biggest issue. What do we do about it? Now there are people who are saying when President Xi Jinping comes to the United States in a couple of weeks, he meets with the President. Trump and Xi should sit down, set up some kind of a process, a working group that would move towards, and essentially a process of what I would call a kind of arms control in the nuclear era. All right, let's sit down with our technologists.

14:32Let's slow down. Let's cooperate. And in theory, that kind of makes sense. We would like to get control over the Chinese models so that we know that they won't be able to hand it over to Iran or to somebody else that there will be some constraints. And yet, can we trust them? The two arms control elements, we want to be restrained towards each other, especially with the spying and the cyber stuff. And there's also the non-proliferation angle. And China is vulnerable, if not more so, than we are because they are a highly digital economy. You know, you could just take them offline overnight with some of these malware attacks that people talk about AI enabling. And this other thing about the US slowing down is frontier progress. Well, we need that progress to counter the AI threats that are coming from outside the country. They don't talk about it much, but especially anthropic is deeply involved in national security agency and helping them maintain their cyber defense and also cyber offense

15:34skills. It gets China and other actors out there. And that's important to have that edge. It's utterly crucial. I mean, we have to have that. You know, obviously the Pentagon and the US security agencies do have access to that kind of technology and they should. And therefore, you want it to be at the cutting edge. You want it to be at the frontier. We are going to take another break and we come back. We'll talk about the evolving politics of artificial intelligence when we come back. America has abundant oil and natural gas, but energy security takes more than energy in the ground. We require smart policies that ensure supply, modernize infrastructure and strengthen energy supply chains, learn more at energysecurityframework.org, pay for by the American Petroleum Institute. Don't forget you can reach the latest episode of Potomac Watch anytime. Just ask your smart speaker, play the opinion Potomac Watch podcast. That is play the opinion Potomac Watch podcast.

16:36From the opinion pages of the Wall Street Journal, this is Potomac Watch. Welcome back. I'm Paul Gigo here on Potomac Watch with Holman Jenkins talking about artificial intelligence. I want to get to this point about whether the Chinese can be trusted. You make the point. Well, they're vulnerable to yes. You make the point that others have made, which is that obviously they don't want to be responsible for the end of humanity. On the other hand, the Communist Party has ambitions to be the dominant global power, I would argue. And because of their political model and their model of domestic and political control, that they try to extend past their borders, the question becomes, if you sit down with them and try to deal with this and you get some kind of cooperation, will the Communist Party simply lie? I mean, you know what happened with the Wuhan Verology Lab. And we still don't know what happened there really because China won't even let the World Health Organization in to inspect or to do a study on it.

17:37And you know, cyber, remember Xi Jinping's meeting with Barack Obama, where Obama brought up, well, stop your cyber attacks. She said, Oh, yeah, okay, we weren't doing it before, but we won't do it anymore for sure. Yeah. This is an area where you're not going to rely on trust. You're going to have to rely on deterrents. You know, we don't really know, but the way people talk about it, it suggests that our cyber abilities are still way beyond theirs. And we hold it, you know, we have a lot of ability to threaten them if they get out of line. So they won't misbehaving. The Wuhan thing is a very good example because someday there's going to be a hugging face in China and they're going to try and cover up. It's not going to be something they intentionally just going to be something they lost control of. You know, and I think that's a giant risk because there's exactly come from the open source nature of their models. The criminal gangs or amateurs can get a hold of them and do things that no one knows about until it springs on the world. And one of the good things about this hugging face episode is that open AI was transparent about it. I mean, it did say, look, this happened.

18:38And we know what happened and people outsiders and others have been able to look and try to explain it. Would you favor some kind of she Trump working group being established on AI? Is that be a helpful step? Anything that gives us a window into what they're thinking and doing can't be bad. You know, the one thing we haven't mentioned, Trump has been talking about this the last couple days saying don't worry about it. Essentially his statement is we have a really smart president. He control this. And in a sense, I don't know about the first part, but the second part is true. We have this thing as, you know, the defense production act of 1950, which the Biden administration cited and when it rolled out, it's a disclosure or monitoring rule. There is a deep well of capacity there to take control of this industry in the name of national defense and basically make sure that the US government has first access to its most powerful capabilities to have some access to control the release of these capabilities to the market. So I think that's, you know, I mean, the defense, there's a great big with Georgetown

19:39University, like four years ago did a really good study on how the defense production act is going to be our main vehicle for controlling when it gets to this point. I think that's what Trump is saying that I do have this tool at hand to do whatever is necessary. And that's necessary to make sure this industry continues to operate US interests. And I think a Democratic president would say the same thing. In his truth social posts this morning, he said, we already have tremendous criminal and regulatory power over these companies. He didn't mention the Defense Production Act, but that's obviously one of them. And then of course, he went on to say there is a sick conspiracy going on against AI and data centers. And the only one that is happy about it is China, whoever wins AI wins. And that went on from there. In a sense, I kind of welcome Trump's pushback against this political frenzy right now, which is all for not just slowing down, but regulating it, controlling it. Bernie Sanders says stop, Rokana says stop. The congressman from Silicon Valley, ironically, you see your hear voices in the media saying

20:43stop. We need to just end this because it's going to kill humanity. You have the tax on data centers. So I like the fact that Trump is pushing back now. Obviously he does it in his blunderbuss fashion. But there's some sense in which he's right that if we stop or we go really slow, then China will take the lead. And that could arguably be much more dangerous for America than continuing. I mean, we are in a moment right now where we could really shoot ourselves in the foot. You look at all the people who've suddenly in the last 24 hours, and this is the ban on the flag. And they can't afford not to join from Obama to Bill Gates to all these people. It reminds me a little bit of Trump leaning against Russia gate and leaning against the COVID lockdowns. I mean, we're in this mood to go through these crazy frenzies every so often where we completely lose perspective. And that's certainly the case with the AI extinction meme that's going around. You know, this is a thing that exists in the subculture of AI researchers and has forever and it's unrealistic.

21:45It's sort of wedded to this science fiction fantasy where you draw a straight line from what's happened today to some apocryphal ending 10 years from now. And it ignores all the things that are going to happen in between, one of which is the Fence Production Act. So in this frenzies, you called it before, a lot of bad things could happen if somebody doesn't tamp down on it. So I'm glad that we have a person who is doing that, even if he does it in Trump's fashion, which is not hardly credible to, you know, some people lose them in America. Right. Right. But he still does have that veto path. Exactly. So he can stop really bad legislation. But I thought it was an important moment when, and not in a good way, necessarily when Barack Obama has spoke up over the weekend and basically advising Democrats to make AI regulation, essentially one of their top priorities in the next Congress and then of course going into 2028. So I assume there'll be a whole slew of Democratic presidential candidates who will be arguing for some form of political control. He's made it impossible for Democrats to be saying on this issue.

22:49You know, you saw it first on a data center opposition and data centers. Now I hate to say it, but the opposition to data centers is not becoming bipartisan. You see it in Wisconsin where the Republican candidate for governor, Rana, had against the Democrats saying poison lakes and are lakes running dry. I mean, that rhetoric isn't helpful. It's insane. But that's kind of where we're moving here. And in that sense, the companies can see it. They can see it. They don't want it to get out of control. So they're trying to kind of show. And I don't say this cynically. I don't think they're cynical about it. I think they're, you know, they are self-interested. But Prudence argues for basically trying to show the public we are responsible about this. And I think that's very interesting. Making a good business out of AI, which means a sustainable business that people are willing to support. So where do you think the politics goes here? There is going to be the AI chair noble. I always thought it was going to come from China. There's going to be, I mean, a mess when some of these agents get loose and take down an important data center or an important industry or, you know, take some bank offline.

23:53And then it's going to be a whole new ballgame. I mean, it's frightening because 99 times out of 100, we do something terribly wrong under those circumstances. 100 times out of 100. I'm just looking from past experience. I mean, I was around in 1979 when we had the three mile island, you know, episode and that essentially put nuclear power in the tank for two generations. Yeah. Fukushima as well. Well, anyway, I mean, I think in that sense, if we're going to worry about this kind of Chernobyl, this kind of episode, which I agree with any new technology, something like that is probably going to happen, it's wise for the companies maybe to get ahead of it and to slow down on their own right to be able to make sure that they can get the machines aligned as the jargon has it in the industry with their human designs. It'd be nice if there was somebody in the administration who could come out with a thoughtful statement about the authorities and powers we already have and how they can be used to make sure this industry develops safely going forward.

24:53I don't know who that might be. I mean, Rubio is smart, Beth and his smart. Rub is really a good salesman to a very large part of the population out there that, you know, it tends to be very hostile to Washington and to elites and that's a valuable quality to have, but he also needs somebody who can speak to elite opinion too in a credible way. Yeah, you need an explanatory power, not just a salesmanship power and that explanatory power is not the president's strength to say the least. All right, Holman Jenkins, thank you so much for coming in and talking about this. Obviously, this is one of the great issues of our day and we will talk about it more. Thanks for coming in and thank you all for listening. We're here every day on Potomac Watch. America has abundant oil and natural gas, but energy security takes more than energy in the ground. It requires smart policies that ensure supply, modernize infrastructure and strengthen energy supply chains, learn more at energysecurityframework.org, pay for by the American Petroleum Institute.