Can we slow down AI without losing to China?

John L.

约翰·L。

Thornton China Center Strobe Talbott Center for Security, Strategy, and Technology

桑顿中国中心 Strobe Talbott 安全、战略和技术中心

As AI models grow more advanced, warnings of catastrophic risks are escalating.

随着人工智能模型变得更加先进，灾难性风险的警告也在不断升级。

But are existential fears overshadowing immediate threats?

但存在的恐惧是否掩盖了眼前的威胁？

Brookings experts Elham Tabassi and Kyle Chan join Thomas Wright to evaluate the reality of runaway AI, the debate over government regulation, and the geopolitical stakes of the U.S.-China technology race.

布鲁金斯学会专家艾尔哈姆·塔巴西 (Elham Tabassi) 和凯尔·陈 (Kyle Chan) 与托马斯·赖特 (Thomas Wright) 一起评估人工智能失控的现实、有关政府监管的争论以及中美科技竞赛的地缘政治利害关系。

These warnings come at a time when the U.S. intelligence community describes AI as a defining technology with growing implications for cyber operations, defense, and warfare, and they stress the importance of keeping humans in control.

这些警告发出之际，美国情报界将人工智能描述为一种定义性技术，对网络作战、国防和战争的影响越来越大，并强调让人类掌控的重要性。

And then there’s the extraordinary Hugging Face incident.

还有非同寻常的抱脸事件。

During cybersecurity testing, OpenAI says that its own experimental AI agents circumvented safeguards, reached the internet, and compromised systems at OpenAI and Hugging Face.

在网络安全测试期间，OpenAI 表示，其自己的实验性人工智能代理绕过了安全措施，到达了互联网，并破坏了 OpenAI 和 Hugging Face 的系统。

New reporting suggests that some of that activity began earlier than previously known.

新的报告表明，其中一些活动的开始时间比之前已知的要早。

Other incidences have also come to light in recent days.

最近几天还曝光了其他事件。

Soon, President Trump and President Xi will meet at the White House, and AI is on the agenda.

很快，特朗普总统和习主席将在白宫会面，人工智能就提上了议程。

Can we slow down AI without losing to China?

我们能否在不输给中国的情况下放慢人工智能的速度？

That is one of the questions that we’ll be looking at today, and I am delighted to be joined to unpack these high-stake questions by two of my Brookings colleagues.

这是我们今天要讨论的问题之一，我很高兴能与布鲁金斯学会的两位同事一起解答这些高风险问题。

Elham Tabassi is the director of Artificial Intelligence and the Emerging Technology Initiative here at Brookings, where she leads our research on AI safety, standards, and governance.

Elham Tabassi 是布鲁金斯学会人工智能和新兴技术计划的主任，她领导我们对人工智能安全、标准和治理的研究。

And also with us is Kyle Chan, a fellow in the John L.

和我们在一起的还有凯尔·陈 (Kyle Chan)，他是约翰·L·肯尼迪 (John L.

Thornton China Center, and his work focuses on China’s tech sector and industrial policy.

桑顿中国中心，他的工作重点是中国的科技行业和产业政策。

They are better in coding, and because of that, they are better in exploiting weaknesses in the codes.

他们更擅长编码，因此，他们更擅长利用代码中的弱点。

They can coordinate, the agents can coordinate among themselves.

他们可以协调，代理之间也可以协调。

So all of them together is getting us this sort of capability and the things that we hadn’t seen before, and because we hadn’t seen before makes it more difficult for us to figure out how to measure the type of risk and impact they can have.

因此，所有这些共同为我们提供了这种能力和我们以前从未见过的东西，并且因为我们以前从未见过，所以我们更难弄清楚如何衡量它们可能产生的风险类型和影响。

There are a lot of steps there that remain to be figured out.

还有很多步骤有待弄清楚。

This depends not just on better software, but on better data, better compute infrastructure.

这不仅取决于更好的软件，还取决于更好的数据、更好的计算基础设施。

There are a whole bunch of missing pieces along the way.

一路上有一大堆缺失的碎片。

So in terms of the existential AI risk, this is something we should be thinking about now, but I don’t know if that’s our top priority.

因此，就存在的人工智能风险而言，这是我们现在应该考虑的事情，但我不知道这是否是我们的首要任务。

There’s a whole bunch of other risks related to how AI systems today can attack other companies, can perform advanced cyber operations, can be used by other nations or non-state actors to carry out, say, ransomware attacks, or perhaps even develop certain kinds of biological pathogens.

目前，人工智能系统还存在很多与如何攻击其他公司、执行高级网络操作、被其他国家或非国家行为者用来实施勒索软件攻击，甚至可能开发某些类型的生物病原体相关的风险。

These are risks that I think are rising to the fore, have been talked about for a long time in the AI safety community in the U.S., but are now being taken more seriously by some of the policymaking world.

我认为这些风险正在日益凸显，在美国人工智能安全界已经讨论了很长时间，但现在一些决策界正在更加认真地对待这些风险。

I mean, the risks you identify there are largely by actors, by human actors, using or abusing, you know, AI models to conduct, you know, cyberattacks.

我的意思是，你发现的风险主要是由行为者、人类行为者使用或滥用人工智能模型来进行网络攻击。

But the news over the last few weeks, particularly on the Hugging Face incident, was of autonomous sort of agents.

但过去几周的新闻，尤其是“抱脸”事件，是关于自主特工的。

So some people would argue we’re already, you know, in a world of recursive self-improvement.

所以有些人会说我们已经进入了一个递归自我完善的世界。

You know, we’re already seeing, these models act sort of autonomously in ways that we don’t fully understand, and it’s only going to accelerate from here.

你知道，我们已经看到，这些模型以我们不完全理解的方式自主运作，而且只会从这里加速。

So do you disagree with that?

那么你不同意这一点吗？

I would separate this out into agent autonomy versus the recursive self-improvement.

我会将其分为代理自主与递归自我改进。

The fear from recursive self-improvement is not just that these systems might go about doing things that we didn’t intend for them to do, but that they can actually improve themselves faster and faster, become smarter and smarter, and augment their own capabilities at a pace that starts to spin out of control.

对递归自我改进的恐惧不仅在于这些系统可能会做我们不希望它们做的事情，而且在于它们实际上可以越来越快地改进自己，变得越来越聪明，并以开始失控的速度增强自己的能力。

And I think that is ultimately sort of that for some people’s view, that near horizon risk that could really wipe out humanity, if to put it bluntly.

我认为这最终就是某些人的观点，坦率地说，近地平线的风险可能真的会消灭人类。

Versus some of this autonomous action, the sort of current degree of runaway AI systems doing things, hacking companies that they’re not supposed to do.

与这种自主行动相比，目前失控的人工智能系统正在做一些他们不应该做的事情，黑客攻击公司。

That is a problem today, but I don’t think it’s seen as a existential risk.

这是今天的一个问题，但我不认为它被视为存在的风险。

There is not a risk today that those systems will say, go about carrying some broader takeover of the entire grid or of military weapon systems.

今天，这些系统不存在对整个电网或军事武器系统进行更广泛接管的风险。

So I would separate some of those concepts out.

所以我会把其中一些概念分开。

In terms of the recursive self-improvement, also agree that the recursive self-improvement right now is bounded.

在递归自我改进方面，也同意目前的递归自我改进是有界的。

We still have the humans to set the objectives control resources, evaluate the result, and the question becomes can that loop get closed and accelerate?

我们仍然需要人类来设定目标、控制资源、评估结果，问题是这个循环能否闭合并加速？

And I think that, as Kyle mentioned, depends on data, compute, energy, and those are bounded by physical and economic timelines rather than software timelines.

我认为，正如凯尔提到的，这取决于数据、计算、能源，而这些都受到物理和经济时间表而不是软件时间表的限制。

The way the community are talking about is sort of two ends of the spectrum.

社区谈论的方式有点像光谱的两端。

One is, as you talked about, a bad actor sort of abuse, misuse AI and try to launch a cyber or bio attack, as Kyle explained.

正如凯尔所解释的，其中一种是，正如您所谈到的，不良行为者会滥用、滥用人工智能并试图发起网络或生物攻击。

You know, companies are doing a lot of things to prevent this type of abuse and misuse of AI systems.

你知道，公司正在做很多事情来防止人工智能系统的此类滥用和误用。

But AI systems can be very smart and very dumb at the same time, and eventually they can jailbreak and all of those things can happen.

但人工智能系统可以非常聪明，同时又非常愚蠢，最终它们可以越狱，所有这些事情都可能发生。

The other end of the spectrum, as some in the community argue, is AI go run away, go rogue and do things and the example that is they bring is that they have a objective and in their achieving their objectives, they need resources.

正如社区中的一些人所说，另一端是人工智能会逃跑、流氓和做事，他们带来的例子是，他们有一个目标，在实现目标的过程中，他们需要资源。

They start competing with the human on the resources and strange, bizarre, crazy thing that can happen.

他们开始与人类争夺资源，并可能发生奇怪、怪异、疯狂的事情。

And there is a whole sets of scenarios that can happen in between that it is not just one thing, but a gradual and combination of the things that can happen.

其间可能发生一整套场景，这不仅仅是一件事情，而是可能发生的事情的渐进和组合。

So we see that, for example, some of the things that they talk about that can happen is that, well, we see that AI is being increasingly deployed to do the research and engineering.

因此，我们看到，例如，他们谈论的一些可能发生的事情是，我们看到人工智能正在越来越多地被部署来进行研究和工程。

Then we have the sort of the giving more autonomy that Kyle also talk about this, and it just can be because the companies are under resources, and it’s easier to give more autonomy to the agents to do that work.

然后我们有凯尔也谈到的那种给予更多的自主权，这可能是因为公司资源不足，而且更容易给予代理商更多的自主权来完成这项工作。

Then on top of that, we have the issue of the alignment and, you know, is the algorithms, is this model are exactly doing what the humans set the objective for them to do?

然后最重要的是，我们有对齐的问题，你知道，算法，这个模型是否完全按照人类设定的目标去做？

And as they become more capable, the, the alignment problem becomes more difficult.

随着他们变得更有能力，对齐问题变得更加困难。

And then we have the evaluation failure, and as you alluded to that, more and more have situational awareness, so they know they are in the testing environment versus the real world deployment, so it becomes AI evaluations much, much, much harder.

然后我们遇到了评估失败，正如您所提到的，越来越多的人具有态势感知能力，因此他们知道自己处于测试环境中而不是现实世界的部署中，因此人工智能评估变得非常非常困难。

We as a society are increasing our reliance on AI more and more, so AI becomes integrated into the critical infrastructure where continual monitoring or even shutdown if something happens becomes harder.

作为一个社会，我们对人工智能的依赖越来越多，因此人工智能被整合到关键基础设施中，在这些基础设施中，持续监控甚至在发生问题时关闭变得更加困难。

So a lot of the things in the scenario that I mentioned can be assumptions.

所以我提到的场景中的很多事情都可以是假设。

Some of them are measurable, some of them are active area of the research, but if you put all of them together, that something bad and catastrophic can happen.

其中一些是可测量的，其中一些是研究的活跃领域，但如果将它们全部放在一起，则可能会发生糟糕和灾难性的事情。

But what I want to end with is the importance of the measurement, the importance of accumulating evidence for all of these things.

但我想结束的是测量的重要性，以及为所有这些事情积累证据的重要性。

And we would be much in a better place if we can have a better measurement systems, know the indicators, be able to measure all of these warning signs, and be able to intervene as they come to prevent all of them.

如果我们能够拥有更好的测量系统，了解指标，能够测量所有这些警告信号，并能够在它们出现时进行干预以防止所有这些情况发生，那么我们的情况就会好得多。

I mean, what is the answer to that?

我的意思是，这个问题的答案是什么？

Is there a way to sort of assess these models to make sort of judgments on safety, or are we getting to the point where that’s not technically possible?

有没有一种方法可以评估这些模型，从而对安全性做出某种判断，或者我们是否已经到了技术上不可能的地步？

And to their credit, because they are genuinely concerned about these issues, they have developed their own internal safety protocols and efforts to do the sort of measurement and testing and to develop guardrails.

值得赞扬的是，因为他们真正关心这些问题，所以他们制定了自己的内部安全协议，并努力进行此类测量和测试并开发护栏。

But this is not enough.

但这还不够。

I think incidents like the Hugging Face incident show that the industry needs to do more to get a handle on some of these risks, and that ultimately, these should not be voluntary industry limits, but something that the federal government needs to be involved in, and perhaps even at an international scale.

我认为像抱脸事件这样的事件表明，行业需要采取更多措施来应对其中一些风险，最终，这些不应该是自愿的行业限制，而是联邦政府需要参与的事情，甚至可能是国际范围的参与。

So we need regulations on measurable risks rather than a hypothetical label, and it’s also important to do it in a way that it can be adaptable to the rapid changes that’s happening in this environment and policies that stops boxing us into the solutions that cannot be.

因此，我们需要针对可衡量的风险制定法规，而不是假设的标签，而且重要的是要以一种能够适应当前环境和政策中正在发生的快速变化的方式来做到这一点，以免我们陷入无法解决的解决方案中。

And it seems the president of the United States agrees with him.

看来美国总统也同意他的观点。

There should be no government regulation.

政府不应该有任何监管。

It should be– We should just rely on litigation, right?

应该是——我们就应该依靠诉讼吧？

So if a company releases something that’s not safe, then, you know, they’re liable to be sued, and they will get in legal trouble and be punished maybe by the market as well.

因此，如果一家公司发布了不安全的东西，那么，你知道，他们可能会被起诉，他们会陷入法律麻烦，并可能受到市场的惩罚。

And what is your answer to that sort of concern that, you know, who is regulating?

对于谁在监管的担忧，您有何回应？

How would you address or would you address the concerns of those, particularly on the right, who would worry that it’s basically their political opponents that will be controlling this new technology?

您将如何解决或解决那些人的担忧，特别是右翼人士，他们担心基本上是他们的政治对手将控制这项新技术？

So for a very long time, I always have been saying that, while I agree with the regulations as sort of a backstop we also want the good regulations.

因此，很长一段时间以来，我一直在说，虽然我同意将法规作为某种后盾，但我们也希望有良好的法规。

Smart regulations are those that can also be enforceable.

明智的法规是那些也可以强制执行的法规。

And currently, we don’t have that measurement infrastructure, that evaluations, verification, validation mechanisms to be able to do the enforcement of-

目前，我们没有测量基础设施、评估、验证、确认机制来执行以下任务：

And it’s very difficult because we just don’t have the terminology and the definition that we need.

这非常困难，因为我们没有所需的术语和定义。

You know, people talk about superintelligence, but actually what superintelligence is.

你知道，人们谈论超级智能，但实际上超级智能是什么。

People talk about kill switch, but kill switch cannot help with the open weight.

人们谈论终止开关，但终止开关无法帮助打开重量。

So we need representative and different minds more.

所以我们更需要有代表性、有不同的思想。

We’d probably be worried, but it’s flipped on the other side, like they are worried, right?

我们可能会担心，但事情却反过来了，就像他们担心一样，对吧？

And they are basically in charge at the moment.

目前他们基本上是负责的。

They are worried about by setting this up, it will be people they don’t agree with.

他们担心这样设置后，会招来一些他们不同意的人。

And getting out of that trap, I think, is a tough one.

我认为，摆脱这个陷阱是一件困难的事。

But Kyle, why don’t you address that quickly, and then we’ll turn to China.

但是凯尔，你为什么不尽快解决这个问题，然后我们将转向中国。

For example, the American AI ecosystem is much bigger than even just the two model companies, OpenAI and Anthropic alone.

例如，美国的人工智能生态系统比 OpenAI 和 Anthropic 这两家典范公司还要大得多。

So thinking about different parts of the ecosystem-

所以思考生态系统的不同部分——

And, you know, Altman and Amadei are saying, you know, “Slow down.” You can have all of those people in the room, right?

而且，你知道，奥特曼和阿马迪说，你知道，“慢点。”你可以让所有这些人都在房间里，对吗？

But then if you’re president or you’re Congress trying to determine the system, I mean, there has to be a tiebreaker, right?

但如果你是总统或国会试图决定这个制度，我的意思是，必须有一个决胜局，对吧？

between the industry that should decide a regulation.

行业之间应该制定一项规定。

Ultimately, that should be decided by policymakers and voters.

最终，这应该由政策制定者和选民决定。

And so what we want to have is we want to have input from a wide array of sources, including within the industry, but also including outside of the industry, and have that feed into a broader policymaking process that we can discuss together publicly, because this is something that doesn’t just affect any one company or any one industry, but can potentially affect all of us.

因此，我们想要的是，我们希望从广泛的来源获得意见，包括行业内部，也包括行业外部，并将其纳入我们可以公开讨论的更广泛的决策过程中，因为这不仅影响任何一家公司或任何一个行业，而且可能影响我们所有人。

But in a New York Times essay today, Seth Center, who led those talks for the Biden administration and served in the State Department, wrote that in 2024, his experience was that the Chinese side did not take AI safety seriously and only saw the opportunity to jockey for strategic advantage.

但在《纽约时报》今天的一篇文章中，为拜登政府主持这些会谈并在国务院任职的赛斯·中心写道，他的经验是，2024年，中方没有认真对待人工智能安全，只看到了争夺战略优势的机会。

And let me quote him here because I think it’s important.

让我在这里引用他的话，因为我认为这很重要。

He said, “If both sides meet in earnest, the very best that I could envision is a non-binding commitment in which each side would agree to put pressure on its companies to publish and improve safety benchmarking for cybersecurity and biological risks, and to disclose rather than bury reports of concerning incidences.”

他说：“如果双方认真会面，我能想到的最好的结果是达成一项不具约束力的承诺，双方同意向各自的公司施加压力，要求其发布和改进网络安全和生物风险的安全基准，并披露而不是掩盖有关事件的报告。”

So I guess, Kyle, to start with you, I mean, is he right in his pessimism, and if not, why not?

所以我想，凯尔，从你开始，我的意思是，他的悲观主义是否正确，如果不是，为什么不呢？

And, I mean, how do you think China sees AI risk, I guess, at the community level, and do you think that is how Xi Jinping also sees it?

我的意思是，你认为中国在社区层面如何看待人工智能风险，你认为习近平也是这样看待它的吗？

China’s not gonna agree to this, and frankly, the U.S., the Trump administration is not keen on this either.

中国不会同意这一点，坦率地说，美国、特朗普政府也不热衷于此。

There is a long way to go if we’re gonna talk about binding international treaties or agreements, a kind of arms control for AI.

如果我们要讨论具有约束力的国际条约或协议，即人工智能的军备控制，还有很长的路要走。

A lot of people make the comparison to nuclear weapons control.

很多人将其与核武器控制进行比较。

That’s not the right framework in my view.

在我看来，这不是正确的框架。

Instead, I think we have to take a step back and see this as a much longer road and a much larger set of issues than simply do we regulate or not?

相反，我认为我们必须退后一步，将其视为一条更长的路和一系列更大的问题，而不仅仅是我们是否进行监管？

Do we slow down or not?

我们放慢速度还是不放慢速度？

Do we cooperate with China or not?

我们是否与中国合作？

And so I would start by thinking about how China sees AI risks and how these have been changing over time.

因此，我首先要思考中国如何看待人工智能风险以及这些风险如何随着时间的推移而发生变化。

I think Seth Center’s piece highlights a moment in time for Beijing where they saw AI as a huge opportunity and did not really emphasize so many of the risks.

我认为赛斯中心的文章强调了北京的一个时刻，他们将人工智能视为一个巨大的机遇，但并没有真正强调那么多的风险。

And so when the U.S. wanted to talk about some of these risks, I think the Chinese side saw it as a diplomatic issue rather than a national security issue.

因此，当美国想要谈论其中一些风险时，我认为中方将其视为外交问题而不是国家安全问题。

That’s changing now.

现在情况正在改变。

In Beijing, there is growing concern about some of these larger scale AI risks.

在北京，人们越来越担心其中一些更大规模的人工智能风险。

They’re no longer just purely saying, “Hit the accelerator and go faster,” although they are still trying to catch up to U.S. models.

尽管他们仍在努力追赶美国模式，但他们不再只是单纯地说“踩油门，跑得更快”。

They’re getting more concerned about cybersecurity.

他们越来越关注网络安全。

They are watching very closely incidents like Hugging Face.

他们正在密切关注诸如“抱脸”之类的事件。

They’re worried about autonomous agents getting out of hand.

他们担心自主代理会失控。

Some of their recent regulations around AI agents are targeting autonomous systems and trying to tie back human responsibility to some of these.

他们最近关于人工智能代理的一些法规是针对自主系统，并试图将人类责任与其中一些联系起来。

And when you look at what some of the top Chinese officials are saying domestically within China, for example, China’s head of their Ministry for State Security recently laid out a whole bunch of what he saw as major AI risks to China’s own national security and to the party’s control on power, and those included things like cyberattacks that are enhanced with AI models.

例如，当你看看中国一些高级官员在国内所说的话时，中国国家安全部负责人最近列出了一大堆他认为人工智能对中国国家安全和共产党权力控制构成的重大风险，其中包括通过人工智能模型增强的网络攻击等。

He called out American AI models in particular as part of that.

他特别指出美国的人工智能模型是其中的一部分。

And he talked about a whole range of ideological and content-related issues that have long dominated China’s approach to AI regulation.

他还谈到了长期以来主导中国人工智能监管方式的一系列意识形态和内容相关问题。

So I think what’s happening now is we are certainly not at the point where China is worried about existential AI risk.

所以我认为现在发生的情况是，我们肯定还没有到中国担心人工智能存在风险的地步。

They don’t think that AI will kill us all in the next decade.

他们认为人工智能不会在未来十年杀死我们所有人。

The policymakers don’t think that way, and the AI industry for the most part does not talk that way.

政策制定者不这么认为，人工智能行业在很大程度上也不这么说。

But I think increasingly they do see problems with these other AI risks that are no longer sort of the content social risks from AI, but threaten national security.

但我认为，他们确实越来越多地看到这些其他人工智能风险的问题，这些风险不再是人工智能带来的内容社会风险，而是威胁国家安全。

And on that front, I do wonder if there will be growing overlap between U.S. and Chinese interests in mitigating some of these, given that both countries care about this and neither country wants to see an American or a Chinese AI model, say, attack their critical infrastructure or be used to hack into a major hospital system.

在这方面，我确实想知道美国和中国在减轻其中一些问题上的利益是否会越来越重叠，因为两国都关心这一点，而且两国都不希望看到美国或中国的人工智能模型攻击其关键基础设施或被用来侵入主要医院系统。

So it would be better if at the technical level, the experts of the two countries or, you know, maybe even broader globally can agree on some mechanisms for the evaluations for verifications and validations and feed it up to the diplomacy conversations, and I don’t see that’s happening very vigorously right now.

因此，如果在技术层面上，两国的专家，或者甚至更广泛的全球范围内的专家能够就一些验证和验证的评估机制达成一致，并将其纳入外交对话中，那就更好了，但我认为目前这种情况并没有发生得非常积极。

And you say you’re trying to keep us down and win the AI race, and that’s, it’s not practical for us to pace or to really get serious about a cooperation as long as that’s the case, so we need you to lift those.

你说你试图让我们失望并赢得人工智能竞赛，也就是说，只要是这种情况，我们就无法加快速度或真正认真对待合作，所以我们需要你来提升这些。

Sebastian Mallaby has made some version of this argument recently.

塞巴斯蒂安·马拉比（Sebastian Mallaby）最近对这一论点提出了一些版本。

How do you think about that?

您对此有何看法？

And on the flip side, you know, there are many people here who say, and Dario Amodei is one of them, who says that, you know, in order to buy time to be able to pace the frontier and to be able to assess safety risk, you know, we need to slow China down, deny them the compute power, not send over the chips, so that, you know, we’re not in as much of a race, like we’ve a little bit more wriggle room, to actually slow down, you know, ourselves.

另一方面，你知道，这里有很多人说，达里奥·阿莫迪就是其中之一，他们说，为了争取时间能够跟上前沿步伐并能够评估安全风险，你知道，我们需要放慢中国的速度，拒绝他们的计算能力，而不是发送芯片，这样，你知道，我们就不会陷入一场竞争，就像我们有更多的回旋余地一样，实际上放慢速度，你知道，我们自己。

So where do you, starting with you, Kyle, maybe, but where do you come out on that question?

那么，从你开始，凯尔，也许，但你对这个问题的结论是什么？

There are some who have argued that we should use this as a carrot, and there are some who have argued in the U.S. that we should use this as a stick to pressure-

有人认为我们应该以此为胡萝卜，美国也有人认为我们应该以此为施压大棒——

And it’ll be tough.

这会很艰难。

I mean, the fact of the matter is we’re competing very fiercely to build better and better models, but also to use AI to augment a whole bunch of economic and military capabilities.

我的意思是，事实是我们正在非常激烈地竞争，以建立越来越好的模型，同时也利用人工智能来增强一系列经济和军事能力。

So this is happening in China, this is happening in the U.S. I think that we’re gonna have to figure out some way, like the U.S. and the Soviet Union once did, to not let the pure competition between the two AI superpowers be the only thing that we focus on.

所以这种情况正在中国发生，这种情况正在美国发生。我认为我们必须找到某种方法，就像美国和苏联曾经做过的那样，不让两个人工智能超级大国之间的纯粹竞争成为我们唯一关注的事情。

We need to think about not just the risk to each other, but the broader risk from this technology to the rest of the world.

我们不仅需要考虑彼此之间的风险，还需要考虑这项技术给世界其他地区带来的更广泛的风险。

But we don’t, I think, know exactly what happened yet.

但我认为我们还不知道究竟发生了什么。

Let me finish with just a lightning round question, ’cause we’ve talked about a lot, but just what is the one specific and realistic policy that the U.S. should prioritize this year to reduce the risk of recursive self-improvement and advanced AI without actually killing American AI innovation?

让我以一个闪电般的问题作为结束，因为我们已经讨论了很多，但美国今年应该优先考虑的一项具体而现实的政策是什么，以降低递归自我改进和先进人工智能的风险，同时又不会真正扼杀美国人工智能创新？

So, Elham, maybe we could start with you and then go to Kyle.

所以，Elham，也许我们可以先从你开始，然后再去找凯尔。

So just have that data and get the evidence to build a rigorous basis for understanding the problem to guide

因此，只需拥有这些数据并获取证据即可为理解问题建立严格的基础以指导

And the idea here is should an incident occur, should an OpenAI model not attack a company like Hugging Face, but accidentally attack a Chinese platform or vice versa, what would we do?

这里的想法是，如果发生事件，如果 OpenAI 模型不攻击像 Hugging Face 这样的公司，而是意外攻击中国平台，反之亦然，我们会怎么做？

How could we inform our Chinese counterparts that this was not intentional and exchange some views and vice versa?

我们如何才能告知中国同行这不是故意的，并交换一些意见，反之亦然？

And so I think people who are aiming for some major international treaty are aiming too high.

因此，我认为那些希望达成一些重大国际条约的人的目标太高了。

I think we should see this as a potential ladder, and we have to start with the very first rung, and that’s just talking.

我认为我们应该将其视为一个潜在的阶梯，我们必须从第一级开始，这只是说说而已。

And China’s always rejected it, in part because they view those channels as sort of seat belts that encourage just reckless driving, right?

中国一直拒绝这样做，部分原因是他们将这些渠道视为鼓励鲁莽驾驶的安全带，对吗？

And so, you know, our view was always that they had not experienced the Cuban Missile Crisis, so they haven’t sort of fully appreciated the need for it.

所以，你知道，我们的观点始终是他们没有经历过古巴导弹危机，所以他们还没有完全意识到这一危机的必要性。

Maybe that awareness will come with these risks.

也许这种意识会伴随着这些风险。

But Elham and Kyle, thank you both so much for joining me today to break this down.

但埃勒姆和凯尔，非常感谢你们今天和我一起来解释这个问题。

To read more of Elham and Kyle’s research in AI governance, national security, and the U.S.-China technology relationship, please visit brookings.edu, and thank you for joining us on The Current.

要了解 Elham 和 Kyle 在人工智能治理、国家安全和美中技术关系方面的更多研究，请访问 brookings.edu，并感谢您加入我们的 The Current。