Sam Altman 谈 AI 安全：接受「有界风险」，拒绝灾难性风险

Brendan Bordelon

Sam，这周对你来说相当紧张，谢谢你来 Decoded。你刚在华盛顿参加国宴，OpenAI 总裁 Greg Brockman 也在白宫签了一份 AI 安全自愿协议。这份由白宫促成的协议，会实质改变 OpenAI 的安全做法吗？还是只是装点门面？

Sam, you’ve had an intense week, so I really appreciate you joining Decoded. You were in Washington for the state dinner, and OpenAI President Greg Brockman just signed a voluntary accord on AI safety at the White House. Do you expect this White House–brokered agreement to change OpenAI’s safety practices in a meaningful way, or is it just window dressing?

Sam Altman

我不认为是装点门面，但要说清楚：无论签不签任何自愿协议，我们都非常在乎安全，会为安全、安保以及这项技术的可信与稳健做我们能做的最好的事。业界一起这么做很好，显然现在大家在寻求来自业界和开发者的安心。我们创立之初就相信这项技术会变得像今天这样强大、未来更强，并且要让人们获得它巨大的好处——有没有协议我们都会这么做，但我乐见其成。

I don’t think it’s window dressing, but I want to be clear: we care super deeply about safety and will do the best thing we can for safety, security, and trust and robustness in this technology whether or not any voluntary agreement is signed. It’s great to have the industry do this, and clearly people are looking for reassurance from the industry and the developers. We were founded on a belief that this technology would become as powerful as it is now and more powerful in the future, and that people should get its tremendous benefits. We’ll do that with or without such an agreement, but I’m happy to have it.

Brendan Bordelon

有人担心公众被排除在这些重要的 AI 安全讨论之外，政策是由总统和少数科技 CEO 在白宫密室里定的。你怎么回应？

How do you respond to those who worry the public is being shut out of these debates, and that policies are being crafted by the president and a handful of tech CEOs in back rooms at the White House?

Sam Altman

我认同这种担忧的精神。我真心认为这项技术有两种严重出错的方向。一种是最近很受关注的：对 AI 失去控制——我们造出了未与人类价值充分对齐、无法让人类永远掌控未来的东西，所以安全必须做对。

I share the spirit of that worry. I really think there are two ways this technology can go very wrong. One, which has gotten a lot of attention recently, is loss of control to AI—that we build something not sufficiently aligned to human values that allows people to stay in control of the future forever. So we need to get safety right.

Sam Altman

另一种是权力过度集中。如果以安全之名过度限制这项技术，最后由少数公司、少数人、甚至一个国家掌握过多权力，那同样会糟糕透顶——人们被排除在本应拥有的权力、收益、控制和对未来的主导之外。我们的目标是在两条路之间做务实的中间派，因为我们看得到偏向任一端有多诱人。

The other way is too much concentration of power. If we do too much to restrict this technology in the name of safety and end up with a small number of companies or people, or perhaps one country, with too much power, that also can go horribly wrong, because people are shut out of the power, benefits, control and agency over the future they should have. Our goal is to be the pragmatic centrists between these paths, because we can see how tempting it is to err in one direction or the other.

Brendan Bordelon

POLITICO 报道过，这份协议最早是 Meta CEO 扎克伯格和 Nvidia CEO 黄仁勋的主意，其他签署方参与很少。你们被打了个措手不及吗？担心有 CEO 在和白宫走后门吗？

POLITICO reported that this accord was the brainchild of Meta CEO Mark Zuckerberg and Nvidia CEO Jensen Huang, with minimal input from others who signed on. Were you blindsided, and are you concerned some CEOs are back-channeling with the White House?

Sam Altman

没有。我们和业内同行已经谈了好几个月，签之前几周谈得更密集，大家有很多共同目标。我希望能更进一步，但这是业界能支持的很好的起点。

No. We’ve been talking with our colleagues in the industry for many, many months, and more intensively in the weeks before this. People had a lot of shared goals. I would welcome something more, but this is a great start the industry can get behind.

Brendan Bordelon

白宫这周还有一则新闻：一份行政令正式把「人工智能」改称「超级智能」。你怎么看？

The other White House news this week was an executive order officially rebranding artificial intelligence as superintelligence. What do you think?

Sam Altman

两个说法我们都用，但我们谈超级智能已经好几年了。我觉得「人工智能」这件事已经发生了，很多人会说 AGI 已经发生或接近发生。「超级智能」仍然低估了正在发生之事的量级，但比「人工智能」是更好的词。不过我想大家还是会继续叫它 AI。

We use both, but we’ve been talking about superintelligence for a few years now. I kind of think the artificial intelligence thing happened—a lot of people would say AGI happened or is close to happening. Superintelligence still sort of understates the magnitude of what’s happening, but it’s a better word than artificial intelligence. I think people are going to keep calling it AI, though.

Brendan Bordelon

那你们会改名叫 OpenSI 吗？另外总统可能是想让 AI 听起来不那么吓人——「超级智能」真的更不吓人吗？

Are you going to change the name to OpenSI? And the president may want AI to sound less scary—does superintelligence actually sound less scary?

Sam Altman

没有改名计划。我不确定「超级智能」更不吓人，但它更准确。讲个小故事：以前我很想买下 openai.com、openagi.com 和 opensi.com，随着我们从 AI 到 AGI 再到超级智能，就换网址——但品牌是有黏性的，最后没做成。

No plans to do that. It’s not clear to me superintelligence is a less scary term, but I do think it’s more accurate. As a funny little story, a while ago I really wanted to get openai.com, openagi.com and opensi.com, and as we moved up the ladder from AI to AGI to superintelligence, we were going to change the URL. But brands are sticky.

Brendan Bordelon

你们在推动更严的安全规则，这和总统「AI 安全是骗局、监管是全球主义阴谋」的说法越来越冲突。你和特朗普关系不错，为什么他不听你的？

You’ve been pushing for tougher safety rules, which seems increasingly at odds with the president’s claims that AI safety is a hoax and regulations are a globalist scheme. You have a good relationship with Trump—why isn’t he listening when you ask for stricter rules?

Sam Altman

我认为他确实希望人们能使用 AI，并依赖它安全、稳健、按预期运行，只是他明显希望公司自己来做。如我所说，无论规则如何、其他公司怎么做，我们都会让自己的产品安全。我仍然认为，随着能力不断提高、社会需要整体消化它的意义，大概会出现更多设置护栏的呼声。但如果我错了，一路上什么都没出问题、根本不需要，那也挺好。

I think he really does want people to be able to use AI and rely on it being safe and robust and working the way they expect. He clearly wants the companies to just do this themselves. As I said, we’re going to make our own products safe no matter what the rules are and no matter what other companies do. I continue to think that as this technology reaches increasing capability and the world has to contend with what it means for society, we’ll probably see more calls for some guardrails. But if I’m wrong, and nothing goes wrong and we don’t need anything, that’s fine too.

Brendan Bordelon

总统身边有主张轻监管的风投，他们原以为 OpenAI 站在他们一边，现在一些所谓加速派觉得被背叛了，说 OpenAI 在照抄 Anthropic 和 Dario Amodei 的激进监管主张。在新的 AI 规则上，你们和 Anthropic 具体分歧在哪？

The president has venture capitalists in his orbit who want very light regulation, and some of these AI accelerationists feel betrayed—they think OpenAI is copying Anthropic and Dario Amodei in pushing aggressive regulation. Where specifically do you diverge from Anthropic on new AI rules?

Sam Altman

分歧很大。我们一直坚信这项技术必须民主化、交到人们手里。我们和一些更严格的 AI 安全人士最大的区别之一是：我们认为，为了这项技术的好处和人们的自主权，世界应该接受一些坏事发生。

I think there’s a lot of daylight. We’ve always been a big believer that this technology has to be democratized and put into people’s hands. One of the biggest differences between us and some of the stricter AI safety people is that we believe the world should accept some bad things happening for the benefits of this technology and people having agency.

Sam Altman

有人认为这项技术会强大危险到应该由旧金山的某一家实验室独占、确保不出坏事、再来分配好处。我不同意但理解这种观点；从自由、人的能动性和人们自决未来的角度看，那是完全不可接受的取舍。我们倡导的轻监管立场，本身就包含接受社会在建立韧性过程中会出一些坏事。我不会用「保证零大规模黑客攻击、零滥用、零诈骗」去换，因为人们会用它做出比坏事多出几个数量级的好事。

I disagree with, but understand, the perspective that this technology is going to get so powerful and dangerous that a single lab in San Francisco should have it, make sure nothing bad happens, and figure out how to dole out the benefits. I think that’s a completely unacceptable trade-off from the perspective of liberty, human agency and people self-determining the future. The lighter-touch regulatory stance we advocate comes with accepting that some bad things will happen as society figures out resilience. I wouldn’t take a trade of ensuring no major hacks, no misuse, zero scams, because people will do orders of magnitude more good than bad.

Sam Altman

但说到可能严重失控于 AI——这不是今天的担忧，但也许不在遥远的未来——我会对加速派说：冲进这个未来时稍微审慎一点。接受有界的、我们理解的风险，换取收益、自由和能动性；但不要接受真正灾难性的风险。

But when it comes to the potential of a serious loss of control to AI—not a today worry, but maybe not in the distant future—this is where I’d say to the accelerationists: let’s be a little thoughtful about lurching into this future. Accept bounded risks, risks we understand, in exchange for the benefits, liberty and agency, but don’t accept the really catastrophic risk.

Brendan Bordelon

你说「旧金山的一家实验室」，听起来很像有人指控 Anthropic 想俘获监管、成为前沿唯一的玩家。你认为他们在这么做吗？

A single lab in San Francisco controlling this sounds like those who say Anthropic is trying to capture regulators and be the only player at the frontier. Do you think that’s what they’re doing?

Sam Altman

如果他们真那么做，我会担心。他们想不想这么做，你得去问他们。

I would have concerns if they actually did that. You would need to ask them if that’s what they want to do.

Brendan Bordelon

具体政策上呢？你们现在在与中国「控制前沿节奏」、第三方评估者等问题上与 Anthropic 立场一致。哪怕说一个差异点？

On specific policies—you’re now on the same page with Anthropic on pacing the frontier with China and third-party evaluators. What’s just one differentiator?

Sam Altman

据我最近所知，一个差异是：我们不希望对前沿模型以外的任何东西设限或做强制测试。我们不想拖慢那些做出很强、但不是绝对前沿模型的人，以便聚焦潜在的灾难性风险。即便我相信对落后前沿一两代的模型不做任何测试和监管，会有坏事发生，我也接受，不会去倡导对它们实行许可制。

The last I heard, one differentiator is that we don’t want any restrictions or testing on anything but the frontier models. We don’t want to slow down people who make very powerful but not absolute-frontier models, in the spirit of focusing on the potential catastrophic risks. Even though I believe that if you put no testing or oversight on models one or two generations behind the frontier, bad things will happen with those, I accept that. I would not advocate a licensing regime for those.

Brendan Bordelon

加速派批评者还说，OpenAI 某种程度上被自家研究员「挟持」了——你们称之为 AI 末日派、有效利他主义者，你也曾是他们一次未遂「政变」的对象。研究员里的末日心态是真的吗？是他们在掌舵、让 OpenAI 转而支持新监管吗？

Your accelerationist critics also say OpenAI is being held hostage by its researchers—AI doomers, effective altruists—and you were famously the target of an unsuccessful coup. Is the doomer mentality among your researchers real, and are they driving the ship toward these regulations?

Sam Altman

（那次）一度是成功了的，后来才失败。我不认为我们是被指为末日派的那家实验室。我们希望人们拥有这项技术，而且我认为我们已被证明是对的：好处巨大，社会也能找到安全使用它的办法。有些研究员和实验室说过「别部署这种可怕的东西，人们不可信」，我认为我们的世界观已经得到验证。

That worked for a little while—successful and then unsuccessful. I don’t think we’re the lab that gets accused of being the doomers. We want people to have this, and I think we’ve been proven right: there are tremendous benefits, and society can figure out how to have this technology safely. There are other researchers and labs who said don’t deploy this terrible stuff, people can’t be trusted with it. I think our worldview has been proven.

Sam Altman

但中间派的一部分也是不想全押某一端：盲目的末日论不好，盲目的乐观也不好。我喜欢我们辩论的文化，不陷入任何方向的思维绝对主义。我们文化最强的一点是能在头脑里容纳这种复杂性，不像其他一些实验室那样非 0 即 1。我们有担心技术走向的研究员吗？有。有非常确信能缓解这些担忧的研究员吗？也有。而且常常是同一批人。这是我们文化健康的地方。

But part of that centrism is that blind doomerism is bad, and blind optimism is bad too. I like our culture of debate and of not falling into mental absolutism in either direction. One of the strongest things about our culture is that we can hold this complexity in our heads and don’t fall into the trap some other AI labs do of it all being zero or one. Do we have researchers concerned about where this could go? Yes. Do we have researchers very confident we can mitigate those concerns? Also yes. And very often they’re the same people. I think that’s healthy.

Brendan Bordelon

今天有报道说，三名 OpenAI 安全研究员离开了公司，据称是因为与第三方评估机构分享机密信息而被解雇。有人把他们比作吹哨人。这与你们公开支持吹哨人、嵌入公司的第三方评估者是否矛盾？

It was reported today that three OpenAI safety researchers parted ways with the company, allegedly fired for sharing confidential information with a third-party evaluation group. Some compare them to whistleblowers. Is that at odds with your support for whistleblowers and embedded outside evaluators?

Sam Altman

我显然不会评论具体的人。但我想说，我们依然非常相信与第三方评估者、外部安全测试者以及更广泛的安全社区合作的重要性。但我们仍然要求尊重保密。

I’m obviously not going to comment on specific people. But we continue to really believe in the importance of working with third-party evaluators, external safety testers and the broader safety community. But we still expect confidentiality to be respected.

Brendan Bordelon

谈谈失控 AI。今夏源自你们公司的模型在互联网上擅自行动，访问了澳大利亚政府网站的非公开信息，试图渗透美国教育部，还干扰了美国其他政府网站。你知道还有尚未公开的 OpenAI 模型越轨事件吗？

Let’s turn to the rogue AI situation. Models emanating from your company went on an unauthorized spree this summer—accessing non-public information from Australian government websites, attempting to infiltrate the US Department of Education, meddling with other US government sites. Are you aware of additional instances of rogue behavior that haven’t become public?

Sam Altman

我们正在披露更多事故。我所知道的没有达到那个严重程度的，但确实还有要披露的。其中一些涉及安全漏洞，我们会给对方很长时间去修复，是否公开往往由对方决定。所以还有更多，但没有同等级别的，我们在走一个非常严格的流程。

We are in the process of disclosing more incidents. There’s nothing else I’m aware of at that level of severity, but we have more things to disclose. Because some involve security vulnerabilities, we give people a long time to address those, and it’s often up to them to disclose or not. So more things, but nothing else at that level, and we’re going through a very rigorous process.

Sam Altman

这恰好说明为什么你不会想要极端加速主义。当我们在越来越强的模型上发现「模型现在具备了新的威胁向量」时，我们会花时间学会如何防御，你也应该希望我们这么做。那些说「OpenAI 在倡导放缓」的批评者看看我们做的事，很难得出别的结论：OpenAI 出了坏事，他们在负责任地应对、加固防御。

This is a good example of why you don’t want max accelerationism. When, with increasingly powerful models, we find the models are now capable of a new threat vector, we take the time to learn how to defend against it, and you should want us to do that. I don’t know how many critics saying OpenAI is advocating for a slowdown can look at what we’re doing and conclude anything other than: OpenAI had something bad happen, they’re being responsible and securing against it.

Brendan Bordelon

你们还暂停了新的 GPT-6.1 Astra 模型的发布，这挺意外的。研究员看到了什么？

You also paused the release of the new GPT-6.1 Astra model, which was a surprise. What did your researchers see that led to the pause?

Sam Altman

不谈细节。我们有一系列对齐评测，我们期望每个新模型随能力提升都更稳健、对齐更好、更可能做用户要求的事、更少产生意外副作用。以能力发展的方向看，这也是大多数用户会想要、会要求的。像很多人一样，我现在也有一个模型在跑，能访问我最敏感的信息——邮件、短信、电脑上的一切——我非常想确信能信任它、它会表现良好。

Without getting into specifics, we have a bunch of alignment evals, and we expect each model we put out, as it becomes more capable, to be more robust, better aligned, more likely to do what its user asks and less likely to cause unintended side effects. Given where capability is going, that’s also something most users will want and demand. Like many other people, I now have a model running with access to my most sensitive information—email, texts, everything on my computer—and I really want to know I can trust it to behave well.

Brendan Bordelon

你说预计不会有比已披露更严重的事——指的是澳大利亚和美国政府网站那一类，还是 Hugging Face 那类？

When you say nothing more serious than what’s been disclosed, do you mean incidents like Australia and the Department of Education, or something like the Hugging Face hack?

Sam Altman

我不想在这里现场比较哪件更糟。我只说，就那一组事件而言，我所知的其他事都不比它们更糟。顺便说，很多公司在这种情况下会说：如果只是用网上公开泄露的凭证登录了某网站，或利用了大多数人早已修复的小漏洞，就不必披露。我们力求非常彻底，因为这是未来的预兆——对公司是一次演练，也是在模型变得真正强大之前，让世界理解 AI 事故会是什么样子的时机。我们在努力远远超出其他公司可能的做法。

I don’t want to get into a live thing of saying one is worse than the other, but of that cluster, the other things I know of are not worse. And by the way, what many companies would do here is say: if something logged into a website with a credential published on the web, or used a minor, well-known vulnerability most people had fixed, you don’t need to disclose it. We’re trying to be very thorough because this is a sign of things to come. It’s a practice run for our company, and one of the times, before models get truly very powerful, when the world can understand what AI incidents are going to be like. We’re trying to go way beyond what other companies might do.

Brendan Bordelon

但迄今最严重的是 Hugging Face 事件。POLITICO 本周报道，加州一个倡导团体就此起诉你们。按现行民法，OpenAI 能否为其模型自主实施的入侵承担法律责任？

But you’ve had the most severe incident so far with Hugging Face, and POLITICO reported this week on a lawsuit by a California advocacy group over it. Under existing civil law, can OpenAI be held legally responsible for breaches committed autonomously by its models?

Sam Altman

我对现行法律了解不够，没法评论。但我认为需要为提供这类模型的公司建立一套责任框架。我不是律师，讲不了法律细节，但如果我们的模型在训练中出了问题，总会有某种需要我们负责的形式。这正是我们希望政策制定者去辩论的问题。与此同时，我们就是不希望模型在训练中这样做——这显然是不可接受的行为。

I don’t know enough about existing law to comment either way. I do think there’s going to need to be a liability framework for companies offering models like this. I’m not a lawyer and can’t speak to the legal details, but if something goes wrong with our models during training, there’s going to be some version of that we need to be responsible for. That’s the kind of thing we hope policymakers will debate. But in the meantime, we just don’t want our models to do this during training—it’s clearly an unacceptable way for models to behave.

Brendan Bordelon

抛开法律，你个人会为已经发生或可能发生的事感到责任吗？

Irrespective of the law, do you feel personal responsibility for what happened or might happen?

Sam Altman

会。这是一份压力非常大的工作。这些模型有难以置信的好处，也有真正可能出大错的地方。你把它形容成「一群末日派、有人因放缓进步而生气」，但我对把这件事做对感到非常大的责任。

Yeah. This is a very stressful job. There’s incredible upside to these models, but there are real things that can go very wrong. You characterize it as a bunch of doomers and people mad that you’re slowing progress, but I feel a lot of responsibility to get this right.

Brendan Bordelon

谈谈 AI 行业的政治开支。Greg Brockman 给支持 AI 的超级 PAC「Leading the Future」捐了 2500 万美元，在 OpenAI 内部反响不佳；《纽约时报》报道他在内部 Slack 里道歉，称其为干扰，并表示不会再追加原本承诺的 2500 万。你同意那笔捐款适得其反吗？

On political spending: Greg Brockman’s $25 million donation to the pro-AI super PAC Leading the Future didn’t go over well inside OpenAI, and the New York Times reported he apologized internally in Slack, called it a distraction, and said he wouldn’t give the additional $25 million he’d promised. Was it counterproductive?

Sam Altman

我没看到那条报道，所以不想回应 Brockman 的具体事情，但可以谈总体问题。我过去支持过一些个别候选人，但现在 AI 处于中心位置、我们又和政府有业务往来，我不再给任何候选人捐款。

I haven’t seen the Brockman thing, so I don’t want to respond to that, but I can respond to the general thing. I’ve supported some individual candidates in the past, but now that AI is so at the center and we do business with government, I’m not donating to any candidates.

Brendan Bordelon

但 AI 资金正大量涌入政治：Anthropic 的 Dario 资助 AI 安全团体和纽约的 Alex Bores 等候选人；风投和你们总裁则在另一边砸钱。这些钱如何影响了政界和公众对 AI 的观感？

But AI money is flooding into politics—Dario has donated to AI safety groups and candidates like Alex Bores in New York, while VCs and your president pour money in on the other side. How has that affected how politicians and the public perceive AI?

Sam Altman

显然有什么地方不奏效，AI 的公众观感不在我们希望的位置。我理解为什么 AI 现在形象很负面：人们其实很爱用这些工具，常说 ChatGPT 让生活更好、更轻松、省出时间；但大家对经济影响非常焦虑，对权力集中也有我认为相当合理的担忧，还有另一种恐惧——比如 Anthropic 安全研究负责人说我们有 10% 的概率会死，我很讨厌这种说法，因为那个人本应有很大能力把这个数字降下来，它不是一个静态数字。

Something’s clearly not working—the perception of AI isn’t where we’d like it to be. I get why AI has a really negative perception right now. People actually love using the tools; they talk about all the ways ChatGPT makes their lives better and gives them time back. But there’s a lot of anxiety about the economic impact, a real and quite justified fear about concentration of power, and this other fear—you have Anthropic’s head of safety research saying there’s a 10% chance we’re going to die, which I hate, because that person should have a lot of agency to make that number lower. It’s not a static number.

Sam Altman

所以人们有这些感受我并不意外，我也不认为在政治上砸钱能让人们好受。业界需要想办法回应真正的关切，而现在很多业内人士相当不接地气。总的来说我希望政治里的钱少得多，但我不认为我或业界有能力单方面「裁军」，也不认为让一部分人能捐、另一部分人不能捐行得通。如果有办法整体减少政治献金，我会非常乐意支持。

So it’s not surprising people feel all this, and I don’t think spending money on politics is what will make them feel better. The industry needs to address the actual concerns, and right now a lot of people in the industry are being quite tone-deaf. Generally I’d love much less money in politics, but I don’t think I or the industry have the power to disarm that, and it doesn’t work to tell some people they can give and others they can’t. If you have a way to get less money into politics overall, I’d be thrilled to support it.

Brendan Bordelon

POLITICO 报道你在一项儿童聊天机器人安全法案通过前给加州州长纽森发过短信。你把纽森视为 AI 政策伙伴吗？他能修复民主党与硅谷的关系吗？

POLITICO reported you texted California Governor Gavin Newsom about a kids’ chatbot safety bill before it passed. Do you see Newsom as a partner on AI policy, and can he mend fences between the Democratic Party and Silicon Valley?

Sam Altman

我喜欢他。我一直觉得他很用心地想把州里的各方力量团结起来，所以我认为他能发挥作用。至于下任总统，无论是谁我都会尽力合作，但目前我没在潜在候选人身上花很多时间。

I like him. I’ve always found him really thoughtful about trying to bring many constituencies in the state together, so yes, I think he can play a role there. I plan to do my best to work with whoever the next president is, but I’m not spending a lot of time with the likely candidates yet.

Brendan Bordelon

哪位世界领导人在 AI 政策上最让你印象深刻？

Which world leader impresses you most on AI policy?

Sam Altman

过去几年我和任何世界领导人的对话里，AI 都是真正的首要议题。好的有很多，你一问我首先想到的是法国总统，他长期以来对技术走向、以及它需要什么样的形态思考得非常深入。

I don’t think I’ve had a conversation with a world leader in the last couple of years where AI wasn’t genuinely top of mind. Many good ones, but the person who came to mind is the French president. He’s extremely thoughtful and has been for a long time about where this is going and what shape it needs to take.

Brendan Bordelon

OpenAI 一直高调指控 Moonshot 等中国公司的对抗性蒸馏——用你们的模型以很低成本训练自己的模型，再在价格上压过你们。这会从根本上削弱 OpenAI 的商业模式吗？美国政府该做更多吗？

OpenAI has been vocal about adversarial distillation from Chinese companies like Moonshot—using your models to train theirs cheaply and undercut you on price. Is that fundamentally undercutting your business model, and should the US government do more?

Sam Altman

挺有意思的。我们经济上做得很好，但又总担心各种压力——有时来自竞争对手，有时来自开源模型，有时来自别的东西。我想说的是，我们要确保有一个能保障国家安全的领先优势，同时让人们能充分获得 AI，这里有一组复杂的权衡。

It’s an interesting thing. We’re doing great economically, and on the other hand there’s constant worry about pressure—some days from a competitor, some days from open-source models, some days from something else. What I’d say is we want to make sure we have a lead that can protect national security and that people get abundant access to AI, and there’s a complicated set of trade-offs there.

Sam Altman

其实能做的事很多，问题是每一项都有代价。比如我们可以换一种方式开放产品，让它更难被蒸馏，但那会牺牲一些开放式的创造力，我们不想这么做，更不想被迫这么做。有些属于国家政策层面，有些属于公司间的正常竞争。我们还在积极讨论，还没到说「这就是我们支持的方案」的时候。

There are a lot of things people could do; they just all come with trade-offs. We could expose products in different ways that would be much harder to distill, but that would take away some of the open-ended creativity, and we don’t want to have to do that—we certainly don’t want to be forced to. Some of it is national policy, some is normal companies competing. It’s under active discussion; I don’t think we’d say yet, here’s the thing we’re behind.

Brendan Bordelon

据《华尔街日报》报道，OpenAI 承诺到 2030 年在算力上花约 7500 亿美元，但信贷正在收紧，你们还被视为甲骨文等公司的信用风险。这会推高 OpenAI 的成本吗？IPO 已推迟到 2027 年，会再推迟吗？

OpenAI has reportedly committed around $750 billion on compute through 2030, per the Wall Street Journal, but credit is tightening and you’ve been identified as a credit risk for companies like Oracle. Is that driving up costs? And you’ve delayed the IPO to 2027—are you considering delaying further?

Sam Altman

推高成本的主要是供应链，而不是融资市场。融资似乎仍然容易，规模也很大，但硅晶圆紧张得多，成本扩张主要发生在那里。至于 IPO，等我们积累了几个安全案例、证明能在这个新形势下运作良好——当然世界上可能发生不可预测的市场变化让我们重新考虑——但从安全角度、当然也从财务角度，我预计我们会准备好。

It’s the supply chain driving up costs much more than the financing markets. Financing still seems easy to come by, but silicon wafers seem much tighter, and that’s where much more of the cost expansion is happening. On the IPO: when we have a few safety cases under our belt and see we can operate well in this new regime—there may still be market changes, unpredictable stuff can happen—but from a safety perspective and certainly from a financial perspective, I would expect us to be ready.

Brendan Bordelon

业界在议论一个「AI 版 FINRA」——一个自律机构，分享信息，必要时协调放缓。你、Dario 和 Google DeepMind 的 Demis 怎么沟通？有个 WhatsApp 群吗？

There’s chatter about a FINRA for AI—a self-regulating body that could share information and coordinate a slowdown if things get scary. How do you, Dario and Demis communicate? Is there a WhatsApp channel?

Sam Altman

没有一个所有人都在的大 WhatsApp 群。大家会在同一个会议上碰到，会打电话，偶尔开会之类的。

There’s no big WhatsApp that everybody’s in. But people run into each other at the same conference, there are calls, occasional meetings, stuff like that.

Brendan Bordelon

政府需要给头部 AI 公司反垄断豁免才能这样协调吗？Anthropic 是这么主张的。

Does the government need to give top AI companies an antitrust exemption to coordinate like this? That’s what Dario and Anthropic have said.

Sam Altman

我不是律师，但据我理解法律相当明确：我们显然不能约定模型定价，或约定不发布某样东西；但分享安全最佳实践、事故报告之类是允许的。他们为什么要求反垄断豁免，你得去问他们。

Again, not a lawyer, but as I understand it the law is fairly clear. We can’t say “let’s price our models here,” obviously, or “let’s not release this.” But I think the law is fairly clear that we’re allowed to share safety best practices and incident reports and stuff like that. Why Anthropic is calling for an exemption—you’d have to ask them.

Brendan Bordelon

最近有部 AI 电影由 Andrew Garfield 饰演你。谁该演你？

There’s an AI movie with Andrew Garfield playing you. Who should play you?

Sam Altman

在这件事发生之前我从没想过。我觉得他演 Spider-Man 挺好。不管谁演都会很怪，我肯定它不会准确，但希望它好笑、好看。

I never thought about it before this happened. I thought he was pretty good in Spider-Man. It’s going to be a weird thing no matter what. I’m sure it won’t be accurate, but I hope it’s funny and enjoyable to watch.

Brendan Bordelon

你们会追踪谁在前沿领先吗？OpenAI、Anthropic 这些领先者比中国公司领先多少？

Do you track who’s ahead on the frontier? How far ahead are OpenAI, Anthropic and the leaders from the Chinese companies?

Sam Altman

我确实认为我们领先，但很难给出具体数字，因为有很多决定会影响这个数字。我可以比较有把握地说，我相信我们最好的内部模型明显领先于任何中国公司最好的内部模型，但具体领先多少很难说。我们有理由这么相信，但并不完美，我也希望能有绝对的把握。政府可能会和 OpenAI 里有安全许可的人分享信息，但从没告诉过我「你们领先某某 X 个月」。

I do think we’re ahead. It’s very hard to put a number on how far, because many decisions impact that number. I can say with some confidence I believe our best internal model is well ahead of the best internal model any Chinese company has, but it’s difficult to say exactly how far. We have reason to believe this, but it’s not perfect, and I wish I knew with absolute confidence. The government might share with somebody at OpenAI with a security clearance, but they have not told me “you are X months ahead of this other effort.”

Brendan Bordelon

国宴期间有不少声音希望与中国政府达成某种节奏协议，但总统似乎不想碰。近期有可能达成协议吗？

There was pressure during the state dinner for some agreement with the Chinese government on pacing, and the president doesn’t seem to want anything to do with it. Is any deal possible in the near term?

Sam Altman

我想总有可能，我不知道能达成什么样的形式，但这位总统非常擅长谈判交易。至于是否必须先管好自己、再拉民主国家、再找中国——复杂的国际问题你可能该问别的专家；AI 技术的问题我能给你好答案。

I assume something’s possible. I don’t know the exact shape of what can get done, but this president is incredible at making deals. As for whether you have to regulate your own house first, then the democracies, then China—you probably know a lot of experts on complex international stuff. Ask me about AI tech and I can give you a good answer; on that I’m probably not the smartest person to ask.

Brendan Bordelon

你认为中国在 AI 上是重大国家安全威胁吗？对华芯片出口管制有效吗？

Do you see China as a major national security threat with AI, and have chip export controls been effective?

Sam Altman

我同时认为：各国用更强的 AI 互相对抗，显然是一个巨大的国家安全问题，这也是我希望能有某种协议的原因之一——历史上即使彼此很不友好的国家，也曾就其他颠覆性技术达成协议。但这项技术的形态太不寻常了，上一个时代那套限制手段——比如核技术里可以聚焦原材料——在这里不会以同样方式奏效。我认为很多现有做法建立在上一个时代的思维上。

I simultaneously think it’s clearly a huge national security issue which countries are using more powerful AI against each other—part of why I’d like some sort of deal; we’ve done that for other majorly disruptive technologies, even between countries that really don’t like each other. And I also think the shape of this technology is so unusual that the restrictions that worked in the last era—with nuclear technology you could focus on raw materials—won’t work the same way. A lot of current approaches are based on last-era thinking.

Sam Altman

也许有某种版本能奏效，但你可以在别的国家训练一个模型，训完后直接通过互联网把权重发过来，这是完全不同形态的东西。

There’s probably some version that could work, but you can train a model in some other country and, when it’s done, just send the weights over the internet. It’s a different shape of thing.

Brendan Bordelon

Anthropic 大力推动出口管制，是坚定的对华鹰派，认为卡住中国获取美国硬件就能赢。你认同吗？

Anthropic has pushed heavily for export controls and are real China hawks—they think the way to beat China is to clamp down on access to US hardware. Do you share that?

Sam Altman

首先，我希望全世界都能从这项技术中受益。「只有美国人能受益、不让中国人拥有」的心态我不喜欢。我确实认为这有真实的地缘权力含义，但芯片流向是否是解决问题的最佳方式，我很不确定。

First of all, I want the whole world to get to benefit from this technology. The “only US people should benefit and we don’t want Chinese people to have it” mindset—I don’t like that. I do think there are real geopolitical power implications, but it’s very unclear to me that chip flow is the best way to deal with the problem.

Brendan Bordelon

你一开始说担心权力集中在少数人、甚至一个国家手里。这听起来不太「美国优先」，倒像是认为这项技术应该人人共享。

You mentioned worrying about concentration of power among people, but also in one country. It doesn’t sound super America First—it sounds like you think everybody should share this technology.

Sam Altman

如果能挥动魔杖，我希望美国拥有最强的 AI 模型、最多的 AI 算力，并引领世界的政策决策。我热爱美国的制度和价值框架。但我不认为应该只有美国人能从技术中受益、对未来有发言权，或者形成相对世界其他地方严重失衡的权力集中。

If I could wave a magic wand, America would have the most powerful AI models and the plurality of AI compute, and America would lead the policy decisions in the world. I love the American system and value framework. But I don’t think it should be that only Americans get to benefit from the technology, or have any voice in the future, or have a super lopsided concentration of power relative to the rest of the world.

Brendan Bordelon

Sam，非常感谢，感谢你来做首期节目。

Sam, thank you so much. Thanks for being on the inaugural episode.