Elon Musk Loses OpenAI Suit, Amazon Trainium Gaining Ground, Open Source AI Struggles

欢迎各位收看《信息时报》TITV。

Welcome, everyone, to the Information's TITV.

我是阿卡什·帕什里卡。

My name is Akash Pasricha.

今天是5月19日，星期二。

It is Tuesday, May 19.

谷歌年度开发者大会今天在山景城正式开幕。

Google's annual IO Developer Conference kicks off today in Mountain View.

明天的节目里，我们的谷歌记者会带来最新动态。

Our Google reporter will bring us the latest updates on tomorrow's show.

今天首先聊聊埃隆·马斯克对OpenAI的这场重大法律败诉。

First up today, we're talking about Elon Musk's big legal loss against OpenAI.

我们的记者罗克特·德鲁将上线，分享他过去三周在法庭上的见闻。

Our own Rocket Drew is coming on to talk about what he has learned from his last three weeks in the courtroom.

《信息时报》独家报道，亚马逊在AI芯片竞赛中正迅速崛起。

And the information published exclusive reporting about Amazon gaining traction in the AI chip race.

我们的亚马逊记者将加入我们，分享她这段时间的发现。

Our Amazon reporter will join us to talk about what she has learned.

我们还会深入探讨前沿AI模型成本不断攀升的问题，以及这如何促使许多公司转向更便宜的开源AI；最后，我们还将与Sapphire Ventures展开对话。

We'll also dig into the rising cost of frontier AI models and how it's pushing many firms to shift into cheaper open source AI, and we'll wrap the show with a conversation with Sapphire Ventures.

我们很快就会邀请管理合伙人贾伊·达斯上线。

We're bringing on managing partner Jai Das very shortly.

我对这场对话非常期待。

I'm excited for that conversation.

今天这期节目会很精彩，咱们马上开始。

It's going to be a fun show, so let's get right on into it.

周一，埃隆·马斯克在与OpenAI的法庭对决中败诉了。

Elon Musk lost his court battle against OpenAI on Monday.

陪审团认定，马斯克提起诉讼的时间太晚了。

The jury in the case decided Musk waited too long to bring on his lawsuit.

马斯克的律师表示，他们打算对这一判决提出上诉。

Musk's lawyer says he intends to appeal the ruling.

我们自己的罗克特·德鲁在过去三周里一直待在法庭上。

Our own Rocket Drew has been in the courtroom for the last three weeks.

他每天都给我们发视频更新。

He's been giving us daily video updates.

你们都看过了。

You have seen them.

我们都很熟悉这些内容。

We know them well.

但今天，他亲自来到现场了。

But today, he joins us in the flesh.

罗克特，欢迎回来。

Rocket, it is good to have you back.

最近怎么样？

How you been?

我挺好的。

I've been good.

能真正面对面和你们聊天，真的特别好。

It's really great to be back properly talking to you.

是的。

Yeah.

亲临现场。

In the flesh.

虽然隔着三千英里，但至少是真人到场了。

3,000 miles away, but still in the flesh.

感觉

Feels

好多了。

much better.

谢谢你这次的分享，比你之前发给我们的简报生动多了。

Thank you for more interactive than the than the dispatch that that you give us.

很多时候，尤其是深夜，等你有时间消化完庭审和法庭上发生的一切之后。

Oftentimes late at night, I should say, after you've had time to process the trial and everything that happened in the courtroom.

我们一会儿再谈这个判决。

So we'll get into the decision in a minute.

不过我想问问你，回顾过去三周，你看到了这么多证人、证词和证据，有什么总体的感受吗？

But I mean, I just want to ask you, looking back on the last three weeks, you've seen a lot of characters, a lot of testimony, a lot of evidence.

对于你所目睹的这一切，你有什么更深层的想法？

I wonder what your broader reflections are on what you've seen play out.

是的。

Yeah.

关于庭审的内容、出现的证据、出庭的证人，我确实有很多想法，但尤其想对那些没能亲临现场、只通过新闻标题或直播关注这件事的人说，整个事件简直像一场马戏。

You know, I have a lot of reflections on the content of the trial, the evidence that came up, the witnesses on the stand, but especially for people who weren't able to attend in person but maybe followed along with the headlines or even tuned into the livestream, the thing that I would try to communicate is just what a circus this whole event was.

我的意思是，太闹腾了。

I mean, what a spectacle.

有人早上五点就在法院外排队等候。

You had people lining up outside the courthouse at 5AM.

当然，其中一些是记者，为了工作才去的，但也有斯坦福的本科生，纯粹是被这事吸引来了。

Sure, some of those were journalists who were doing it for their jobs, but also you had Stanford undergrads that were just intrigued by what was going on.

他们一大早起来，拼车赶到奥克兰，只为能挤进旁听席，看马斯克作证，或者哪怕只是看他的助手们作证。

And they woke up early and carpooled up to Oakland so that they could sit in just an overflow room even and watch Musk testify or watch even just Musk's associates testify.

而且我

And I

连法院都没料到，这场审判会引来这么大的关注。

think the court was even caught off guard by how big, how much attention this trial got.

相比过去马斯克在这个领域出庭的其他案子，这次的热度简直不可同日而语。

Compared to past trials where Elon even has testified in this area, this case got so much attention.

到了中途，法院干脆开了直播，这下局面彻底变了。

And then halfway through, they introduced the livestream, which kind of changed the dynamics.

有些日子，直播的累计听众超过了上万人。

There were days that that livestream cumulatively got over 10,000 listeners.

这个案子的高关注度，确实非同寻常。

I mean, just the high profile nature of the case was really remarkable.

就连周一，也就是我们拿到判决结果的那天，热度也一点没降。

And that continued even on Monday, even the day that we got the verdict.

这场审判被分成了两个阶段。

The trial was bifurcated into two phases.

第一阶段是责任认定阶段。

The first was the liability phase.

这才是审判的核心部分。

That was the core of the trial.

持续了三周，陪审团听取了证据，来判断OpenAI、微软，还有萨姆·阿尔特曼和格雷格·布罗克曼有没有责任。

It was three weeks where the jury heard evidence that would establish whether or not OpenAI, Microsoft, as well as Sam Altman and Greg Brockman were liable or not.

但这个阶段已经在周一结束了。

But that phase had already ended on Monday.

陪审团现在在另一个房间进行商议，而法庭则进入了第二阶段——救济措施阶段。

The jury was off deliberating in another room, and the court was proceeding with a second phase, which was the remedies phase.

这个阶段的目的就是，如果陪审团最终裁定确实存在责任，那接下来该怎么办？

And the purpose of that phase was, if the jury comes back and their verdict is there is liability here, then what?

该采取什么措施？

What should happen?

法院和法官该采取什么样的措施？

What kind of actions should the court and the judge take?

这一阶段依然引起了很大关注。

And there was still a lot of interest in this phase.

我本来以为第二阶段会特别无聊。

I expected that this second phase would be kind of a snooze fest.

毕竟证人又不是萨姆·阿尔特曼。

I mean, witnesses aren't Sam Altman.

这些证人都是OpenAI或者马斯克拉来法庭的一些普通经济学家。

The witnesses are some random economist that OpenAI or Musk is dragging into the courtroom.

但现场还是坐满了律师和记者。

But still, it was packed with lawyers and journalists.

就在这一阶段第一天开始一个多小时后，陪审团就作出了裁决。

And just a little over maybe an hour and a half into the first day of this portion, the verdict was returned.

当时法庭画师正画到一位作证的经济学家，结果突然被打断了。

I mean, the the courtroom sketch artist was halfway through sketching an economist who was testifying, and she got interrupted.

她连水彩都没开始画呢。

She hadn't even moved on to the watercolors.

还只是铅笔稿。

It was just pencil.

她待会儿得把这稿子重画一遍。

She's gonna have to flush it out later.

所以这个判决来得这么早，我想谁都没想到。

So it was a surprise that the decision came so early, I think, to everyone.

但我最想说的是，当时法庭里的那种氛围。

But that's the thing I really want to communicate, is just the atmosphere in the room.

某种程度上，这也反映了过去两三年这项技术飞速发展的样子。

Well, in some ways, it seems like a little bit of a microcosm of the way that this technology has evolved so quickly too over the past two, three years.

我们节目里的人还在讨论模型授权和相关影响。

I mean, we have people on the show talking about a model release and talking about the implications.

结果还没来得及消化一个模型，下一个又来了。

And then before even have time to digest a single model, another one comes through.

听好了，这是一起法律案件。

Look, this is a legal case.

我们在这讨论的是两件完全不同的事。

These are two separate things we're talking about.

但庭审节奏这么快，事情随时在变，简直太能代表我们当下这个时代的特征了。

But the frenetic nature of the trial and things changing in real time, it sounds to be very emblematic of the moment that we're in.

确实如此。

That's true.

没错。

That's true.

整个氛围特别有硅谷的味道。

It is a very Silicon Valley kind of vibe to it.

再说内容上，听到律师们也整天说这些词，还挺有意思的。

And then also in terms of content, I mean, it's funny to hear the lawyers talking about these same terms.

他们动不动就提GPU、Nvidia H100，AGI挂在嘴边，还点名各种公司具体用的模型。

They're talking about GPUs and Nvidia h one hundreds, and they're talking about AGI all the time, and they're name dropping specific models the different companies.

我的意思是，听他们说话，简直像在看TITV节目一样。

I mean, it sounds like they could be on TITV when you hear them talk some of the time.

这一点也特别有意思。

So that's also been really entertaining.

那我们来谈谈这个判决本身吧。

So let's get to the decision itself.

我们看到的结果是OpenAI胜诉，马斯克败诉。

So we saw the decision OpenAI wins, Musk loses.

我们还谈到了诉讼时效的问题。

We saw the statute of limitations.

昨天我们跟一位律师聊了聊，关于这个判决的核心内容。

We talked with a lawyer yesterday talking about really the substance of the decision itself.

所以这部分大家应该都比较清楚了。

And so I think that part is sort of well understood.

我其实更想听听你的看法：埃隆·马斯克在X上发了回应。

What I actually wanted to get your thoughts on was So Elon Musk, he issued a response on X.

他说，这根本就是关于诉讼时效的问题。

He said, This is all about the statute of limitations.

法院其实根本没有裁定OpenAI是否错了——比如从一个慈善机构变成非慈善机构，侵占了非营利组织的资产。

The court never actually ruled on the substance of whether or not OpenAI, basically if they were in the wrong by sort of going from a charity to not so charity, stealing from the nonprofit.

我想知道，你对马斯克那篇帖子的反应是什么？

My question is, what was your reaction to Elon Musk's post?

是的。

Yeah.

我要说，他发了两篇帖子。

Well, I will say he had two posts.

有一篇他后来删了，里面指责法官是激进派，可能后来觉得不妥，就删掉了。

There was one post that he deleted where he called out the judge for being an activist, and then maybe wisely removed that post.

但他还发了另一篇，说这个判决只是基于技术性细节，根本没有认真评估他对OpenAI和微软提出的指控是否成立。

But he had another one where he called out the decision for being a narrow technical call rather than really evaluating the merits of his claims against OpenAI and against Microsoft.

说实话，这确实是个很狭隘的判决。

And frankly, it's true that it was a narrow decision.

你看看陪审团的表格，只要他们认定诉讼时效问题，就可以直接结案了。

Like, if you look at the jury form, as soon as they rule on statute of limitations, they get to move on.

他们根本不用回答其他问题，这就是为什么判决来得这么快。

Like, they don't have to answer the other questions, which is why the decision came so quickly.

说实话，才花了两个小时，结果本来就应该这样。

You know, really, after only two hours, it couldn't have been any other way.

他们实际花的时间就这么点。

Like, that's the amount of time they spent

如果诉讼时效成立，整个案子根本没必要进入下一阶段。

the whole even proceed to the next to the next legal matter if the statute of limitations is is what the decision is.

对。

Right.

面对那么多证据要梳理，其他指控本来需要花大量时间，但很可能他们一进陪审室，就听到了OpenAI的结案陈词——他们说马斯克抱怨的那些事，都发生在诉讼时效之前，他早该知道的。

When there's mountains of evidence you would have to sift through for those other claims, probably what happened is they all walked into that room, they had all heard OpenAI's closing arguments where they argued that the things Musk is complaining about happened earlier than the statute of limitations, and reasonably, he should have known about them before the statute of limitations.

他们肯定都听到了那段结案陈词。

They probably all heard that closing statement.

他们一进陪审团室，发现大家的想法都差不多，这大概就是为什么判决这么快就出来了。

They came into the jury room, and they realized they were all roughly on the same page is probably how that went down so quickly.

但确实，这个判决范围很窄，如果法官给陪审团的指示不一样，结果可能会不同。

But it's true that it was a pretty narrow decision, and it could have gone otherwise if the judge had given different jury instructions.

所以，他们现在抓住的就是这一点。

So this is the thing that they're pointing to.

你知道，马斯克的律师说这是个技术性判决，打算上诉，其实陪审团本来是可以被给出不同指引的。

You know, Musk's lawyers, when they say that it was a technical decision and they want to appeal, the jury could have been instructed otherwise.

他们本来可以被明确告知，诉讼时效的起算时间到底该从什么时候算起。

They could have been given different instructions about when the statute of limitations clock starts running.

比如，按照马斯克的说法，每次发生违规，时效都该重新计算。

For example, it could have been the case that the clock resets every time there's a violation, according to Musk.

他们想请更高一级的法院重新审视这一点，这正是他们上诉的理由。

So that's the thing that they would like a second opinion on, and that's sort of the grounds for their appeal.

当你回过头来看这场审判，我们其实学到了很多。

When you think about the trial, we learned a lot.

有哪些问题还没得到解答？

What questions didn't get answered?

或者你还在期待哪些问题能有答案？

Or what questions are you still hoping for answers to?

也许不是关于马斯克和阿尔特曼的关系，也不是关于XAI和OpenAI的关系，但接下来我们该往哪儿走？

Maybe not with respect to the Musk Altman relationship, or the XAI OpenAI relationship, but where do we go from here?

对。

Yeah.

我觉得这里有个问题，是关于公司该如何成立的先例。

I think there is a question here about precedent for for how companies get founded.

这正是马斯克在整个庭审中一直在强调的观点：如果放任OpenAI这样行事，那人人都会跟着学。

This was sort of the case that Musk has been advancing throughout the trial, is that if you let OpenAI get away with this, then everyone's going to do it.

我觉得这种说法在几个方面被夸大了。

I think that's overblown in a few ways.

首先，这个案子本身可能并不会确立多少真正的法律先例。

One is that the court case itself might not result in much true legal precedent.

而且，就算真有先例，OpenAI 经历的这些麻烦本身就已经树立了一个先例。

But also, to the extent there is precedent, it's already been created just by the amount of hassle that OpenAI has had to go through.

有律师跟我指出，现在没有哪家理智的初创公司，也没有哪家像 OpenAI 那样搞科技的非营利组织，会在注册章程里把自己的使命和目标限制得这么死。

I've had lawyers point out to me that no startup in its right mind or no nonprofit even that's going to be doing tech the way OpenAI is, is going to constrain themselves so much in their mission and purpose when they set out their articles of incorporation these days.

他们会看到 OpenAI 的例子，然后说，不管官司最后怎么判，我们可不想落到那种境地。

They will see OpenAI's example and they'll say, We don't want to end up in that situation, regardless of which way the court case ultimately goes.

所以，先例问题确实很重要，这一点依然成立，但可能没马斯克说的那么具有普遍影响。

So the question of precedent an important one still, remains important, but is not as sweeping maybe as Musk would make it out to be.

不过我觉得，其他问题还是集中在违反慈善信托责任这一点上。

I think the other questions, though, are still on this breach of charitable trust.

这里有个问题，就是慈善机构在与营利性关联公司，比如像微软这样的大型企业合作时，能做什么、不能做什么才算合理。双方请来的法律专家都是非营利组织惯例方面的行家，但他们对什么是正常操作、什么该被接受，看法完全相反。

There's a question of what can a charity reasonably do in its relationships with a for profit affiliate, in relation with a large corporate player like Microsoft, and the legal experts that both sides brought in, who are well versed in this, they're the legal experts on nonprofit custom and practice, they really disagreed about what is normal and what should be accepted practice here.

所以显然还有一些悬而未决的问题。如果案件上诉并继续推进，这些问题就会冒出来，到时候就得搞清楚，在这个案子里，慈善信托责任到底有没有被违反。

So there are clearly some unresolved questions, and those are the kind of things that would come up if there's an appeal and it ends up moving forward and there has to be a question about whether a charitable trust was breached in this case.

好的。

Great.

罗杰特，非常感谢你今天来做客。

Well, Rocket, I want to thank you for coming on.

接下来几天好好休息一下，希望再别有官司了。

Get some well earned rest the next couple days, and let's just hope that there's no more lawsuits.

不过我得说，我觉得还会有更多官司找上门。

Although, I will tell you, I think there's more lawsuits coming.

那肯定少不了。

Oh, I bet there are.

如果你还想回法庭，他们肯定还会请你去的。

So to the extent that you want back in the courtroom, I'm sure they'll have you back.

罗杰特，非常感谢你今天来做客。

Rocket, I want to thank you for coming on.

这位是我们的记者罗杰特·德鲁，他报道人工智能、机器人，以及法庭上的各种事。

That is Rocket Drew, our reporter covering AI, robotics, and everything in the courtroom here

谢谢，阿卡什。

Thanks, on Akash.

亚马逊曾经被认为是人工智能竞赛中的落后者，但随着它的Trainium芯片逐渐成为英伟达GPU的有力竞争者，加上公司芯片软件的不断优化，它在人工智能领域已经站稳了脚跟。

Amazon was once seen as a laggard in the AI race, but as its Trainium chip becomes a more popular challenger to Nvidia's GPUs and the company's chip software improves, it has very much gained traction in the AI arena.

我的同事凯瑟琳·佩洛夫今天发表了一篇深度报道，讲的就是这件事，我想请她来聊聊这个话题。

My colleague Catherine Perloff published a deep dive about that today, and I want to bring her on to talk all about it.

凯瑟琳，欢迎来到节目。

Catherine, welcome to the show.

很高兴你能来。

It's great to have you here.

嗨，阿卡什。

Hi, Akash.

你能不能说说，Trainium芯片现在发展得怎么样了？

Tell me, how much traction has Trainium been getting?

是的。

Yeah.

我觉得，过去几个月，Anthropic和OpenAI都签了大单，决定使用Trainium芯片。

So I guess like, you know, obviously Anthropic and OpenAI signed these big deals in the past couple months to use Trainium.

但我发现，一些规模较小的开发者也意识到这款芯片的性能在提升，尤其是软件方面进步明显。

But what I found is that there are smaller developers that have also found that the chip is getting better and particularly the software is getting better.

所以我觉得，这意味着确实有更多人开始关注 Trainium 芯片。部分原因是软件变好了，另一部分原因是 Nvidia 的芯片越来越难搞到。

And, you know, so I think that that means that, yeah, more people are looking at the Trainium chips and partly that's because the software is getting better, partly because it's getting harder to access Nvidia chips.

不过关于更广泛的市场接受度，比如到底有多少客户在使用，我们稍后可以再深入聊聊。

But I guess, and we can talk more about this later in terms of broader traction, in terms of like, you know, how many customers are using that.

我觉得目前这还是个不太确定的问题。

I think that's still a bit of an open question.

亚马逊并没有公布具体数据，说明他们的 Trainium 收入中有多少是来自哪些客户的。

Amazon doesn't break down like what percent of their Trainium revenue comes from what customers.

不过我觉得我们可以说，越来越多的 AI 初创公司开始把 Trainium 视为英伟达的一个可行替代方案。

But I think we can say that more AI startups are starting to see Trainium as, a a viable alternative for Nvidia.

那我们来深入聊聊你提到的这几个点。

So let's dig into a couple of these pieces here that you mentioned.

软件方面是你报道的重点，你也采访了好几位使用过这些芯片的人。

The software element to this was a big part of your story, and you spoke with a number of people who have used these chips.

我想他们一开始对软件有些抱怨，说软件的表现没达到预期。

And I guess they had initial complaints about, Hey, the software wasn't working as well as it needed to.

这就是之前大家不愿意用 Trainium 芯片的原因。

That was what was preventing people from using the Trainium chip.

他们具体遇到了哪些问题？

What were the exact problems that they were facing?

然后我们再聊聊这些问题有没有得到改善。

And then we'll talk about if they've gotten better.

对。

Yeah.

所以我觉得，首先的问题是，Trainium 有一套自己的软件，而长期以来，Nvidia 的软件一直是开发者的默认选择。

So I think, you know, the first issue is that, you know, Trainium had its own set of software that, was sort of, not the Nvidia software's been sort of like the default for developers for a while.

另外，Nvidia 的软件与 PyTorch 这个开源平台集成得非常好。

Also, Nvidia software integrates very well with PyTorch, an open source platform.

PyTorch 在开发者群体中也非常普遍。

That's also pretty common for developers.

开发者就是用 PyTorch 来调用芯片的计算能力吗？

PyTorch is what is what the developers would use to basically access the compute on the chip?

差不多是这个意思吧？

Is that the idea?

这套软件让你能编写模型，让模型在芯片上运行。

This software is that you can make your model you can code your model to work on the chip, basically.

Trainium 的软件有自己的语言，你得专门去学它。

And the Trainium software kind of had its own language and you had to learn that language.

而且一旦出问题，很难找到人解决。

And also if problems came up, it was hard to get those problems solved.

软件里头也没多少文档，很难找到怎么处理各种问题的指南。

So there wasn't a ton of documentation within the software to, you know, kind of give you kind of manuals for how to solve different problems.

遇到问题时，亚马逊的工作人员响应有时也挺慢的。

And sometimes Amazon staff could be slow in responding when you ran into problems.

所以这些开发者一边要学一门新语言，一边却没人给现成的代码参考。

So, you know, these developers are trying to learn a new language and no one's really giving them the code.

那阵子确实是个问题。

That was sort of the problem for a while.

而且软件本身也有不少限制。

And then, know, also the software itself could be limited.

有些功能和调整，开发者想做却发现Trainium不太擅长，有时候还得把模型写两遍，而本来用一种方式就能搞定。

Like there were certain, capabilities, and sort of tweaks that, developers wanted to make that they found that Trainium wasn't that good at, or you'd have to sort of code the model twice to do the same thing that you could maybe do in one kind of coding session for.

但现在好多了。

With But now it's now it's better.

现在确实好多了。

Now it's now it's better.

我觉得部分原因是他们更好地融入了开源平台，比如PyTorch。

I, you know, I think part of it is they've integrated better with open source platforms, developed like, like PyTorch.

几个月前他们宣布了对PyTorch的原生支持，这真是个重大突破——现在不用再先用Trainium自己的语言写代码，再想办法迁移到Trainium，省了一大步。

They announced the native integration with PyTorch a couple months ago, which is a real big game changer because now instead of having to sort of code in Trainium's language and you would really code in PyTorch and then try to transfer to Trainium, which would be kind of an extra step.

现在可以直接一步到位了。

But now you can kind of do it all at once.

另外，开发者们表示，一些bug已经被修复，而且亚马逊的开发团队响应速度也更快了。

Also developers say that they fix some of the bugs and that also Amazon developers are a lot more responsive.

他们会在GitHub上积极回复大家的问题。

They respond to stuff on GitHub.

一般来说，你的问题大概一两周就能得到解决。

You know you can kind of get your problem solved maybe in a week or two in general.

所以我觉得，这些问题的确已经解决了不少。

So I think that you know a lot of these issues are fixed.

不过说实话，现在肯定还是更多开发者对英伟达的软件更熟悉，但要把模型迁移过去，难度已经小多了。

It's still probably you know, more developers are comfortable with Nvidia software, but it is less of a big lift to be able

可以开始进行这些训练了。

to start these training.

另外得提一下，亚马逊最近受到这么多关注，背后有个很重要的背景：现在显卡严重短缺，大家根本抢不到想要的英伟达芯片，因为需求实在太大了。

And we should say, I mean, one of the one of the, backstories here that is really important to Amazon getting a lot more attention is the fact that there is a GPU crunch and people cannot get their hands on the Nvidia chips that they want because there is so much demand.

这种情况可能正好给亚马逊提供了一个机会，让他们能把 Trainium 推向市场。

And that maybe could provide a bit of a window here for Amazon to get Trainium out there.

亚马逊有足够产能满足需求吗？

Does Amazon have capacity to satisfy the demand?

他们能实际供货并及时交付客户，这会不会成为他们的竞争优势？

Could this be a competitive advantage for them in that they actually have supply and they can get it to customers?

是的。

Yeah.

我认识的一些开发者说，他们选择 Trainium 的一个重要原因是很难拿到最新的英伟达芯片。

I think that is I mean, the developers I talked to, they said a big pull factor towards Trainium was the fact that it was hard to access the most recent, Nvidia chips.

他们还提到，如果你跟 AWS 的人谈英伟达芯片，他们可能在 Trainium 上更愿意谈价格，或者多给一些积分。

And they said that, you know, if you talk to AWS reps about Nvidia chips, they're they might be more flexible on pricing with Trainium or offer you more credits, than they would with Nvidia chips.

现在亚马逊表示，他们的多代 Trainium 芯片几乎都卖光了，或者已经被全部预订。

Now Amazon says that, you know, a lot of their generations of Trainium are either almost sold out or, you know, fully subscribed.

所以这些 Trainium 芯片并不是闲着没用的。

So it's not like these chips like the Trainium chips are like just sitting around.

实际上，它们的需求也很高。

Like they're also, I think, in pretty high demand.

但我感觉，它们的供应量比英伟达的芯片更充足。

But the impression I get is that they're more available than the Nvidia chips.

而且在整个背景故事里，我觉得很重要的一点是，芯片和软件之所以进步，很大程度上要归功于Anthropic。

And also just kind of in talk in kind of in this whole sort of backstory thing, I think it's important to note that part of the reason that the chips and the software have gotten better is Anthropic.

Anthropic和亚马逊合作得非常紧密。

Anthropic and Amazon worked really closely together.

他们经常沟通交流。

They would talk very frequently.

Anthropic为了满足自身需求对Trainium做的优化，也惠及了所有非Anthropic的Trainium用户，因为这些改进让软件变得更高效、运行更快，整体性能提升了不少。

And some of the ways that Anthropic helped optimize Trainium for their own needs has helped all Trainium customers that aren't Anthropic, because they help to kind of make the software better and work faster and sort of more efficiently in lots of ways.

那亚马逊自己的基础模型呢？

What about Amazon's own foundation models?

你提到过Bedrock和Nova，我觉得那是一系列模型。

I mean, you talked about Bedrock and Nova, I think, is the is a family of models.

这些模型相比以前有进步吗？

I mean, have those gotten any better compared to what they once were?

因为那确实是个大问题。

Because that was a big thing.

你知道，大概一年前，当亚马逊还被排除在讨论之外时，我们一直有个疑问：谷歌推出了自己的前沿模型，那亚马逊到底处在什么位置？

You know, that was a question we had probably a year ago when Amazon was sort of out of the conversation was you have Google coming out with its own models that are cutting edge.

说实话，我当时真不清楚亚马逊在这件事上是什么角色。

I mean, I didn't know where Amazon fell in the conversation.

亚马逊现在还是有Nova。

So Amazon still has Nova.

我接触过一些使用Nova的AWS客户，他们说这东西很便宜。

And, I talked to AWS customers who are using Nova, and, they're cheap.

真的很便宜。

It's cheap.

但Nova算不算一个被大家广泛讨论的顶尖AI模型呢？

But is Nova a, like a leading AI model that people talk about?

就像大家常提OpenAI、Anthropic和Gemini那样？

Like they talk about OpenAI, Anthropic, and Gemini?

不是。

No.

所以，Nova 的未来嘛，你知道的，它还在。

So I, you know, I guess the future of Nova, you know, it still exists.

不过，我觉得有了 Trainium 芯片，亚马逊还是能在 AI 竞赛中保持一定的领先地位。只要他们的芯片是那些顶尖 AI 实验室想用，或者至少愿意用的——特别是如果亚马逊肯砸钱的话。

But, I guess, you know, I think with these Trainium chips though, I think though it's sort of a way for Amazon to still be somewhat of a leader in the AI race if they can have a chip that, you know, that the leading AI labs want to use, or at least, you know, will also use if Amazon invests a bunch of money in them.

也就是说，即便 Nova 模型不被视为前沿技术，亚马逊依然能通过这种方式在 AI 竞赛中保持领先，并拥有重要的创新成果。

But you know, that that it's a way for Amazon to still be leading in the AI race and have an important innovation in the AI race, even if their Nova models aren't seen as frontier.

这并不意味着没人用，但据我了解，大家好像对 Trainium 芯片的关注度比 Nova 模型要高一些，尽管肯定也有客户在用 Nova。

That doesn't mean nobody's using them, but you know, they're I think they're not there seems, you know, kind of in my reporting, maybe a bit more excitement about Trainium chips than there are about Nova models, even though there are definitely, you know, customers using those too.

最后我再问一个问题。

And let me ask you one last question here.

你觉得这个故事，至少在过去一年关于 Trainium 的发展中，关键的转折点是什么？

You know, if you think about the the turning point in this story, at least in the in the past year or so with respect to Trainium.

我脑海里有一个关键节点，就是亚马逊投资 OpenAI。我稍微停顿一下，因为我想确认 OpenAI 当时是不是说过要用 Trainium 芯片。所以，考虑到他们和 Anthropic 的关系，这算是 Trainium 普及过程中的一个大转折点吗？

One inflection point that I have in my mind is Amazon's investment into OpenAI and the extent to which I'm pausing here because I'm trying to remember if OpenAI said they were going to use Trainium chip for So my was that sort of a big turning point in this Trainium adoption story that because they had the relationship with Anthropic.

他们现在正在和OpenAI取得进展。

Here is them making inroads with OpenAI.

我的意思是，这两件事大概就是亚马逊芯片走向市场的关键认可吧，对吧？

I mean, these two probably seem to be like the stamps of approval that Amazon needs to get their chips out there, right?

不过说实话，从开发者角度来看，真正的转折点其实是软件的改进。

Honestly, though, I think from the developer point of view, what was really the turning point was just this improvement in the software.

因为我接触过一些开发者，他们说，亚马逊当然会回应OpenAI和Anthropic的需求，但未必会重视我们这样的公司。

Because like, you know, I, I, I talked to developers who are like, well, you know, Amazon's going to be responsive to OpenAI and Anthropic, but they won't be responsive necessarily to us.

不过我不确定这是否属实。

Now, I don't know if true.

我相信亚马逊也能对小型公司做出积极响应。

I'm sure like Amazon can be also responsive to smaller companies as well.

但对眼下那些小型公司来说，最重要的是能用上这套软件。

But like what, you know, smaller companies need to get going today is to be able to use the software.

好像从去年下半年开始，这方面才真正有了明显改善。

It seems like that sort of really started to improve in the back half of last year.

我觉得 OpenAI 和 Anthropic 的情况，比如 OpenAI 今年早些时候成为客户，Anthropic 今年早些时候加大了对 Trainium 的投入，这些肯定对 Trainium 的故事有帮助，也确实起到了推动作用。

I think OpenAI and Anthropic, OpenAI signing up as a customer earlier this year, Anthropic, enlarging their commitment to Trainium earlier this year, definitely like helps the Trainium story and it definitely helps.

就像杰西在最近一次财报电话会议上说的，他们定制芯片业务的年化收入规模达到了 200 亿美元，这不仅仅包括 Trainium，还涵盖了其他产品。

Like Jassy said in the most recent earnings call that the run rate of their custom chips business, which isn't only Trainium but includes Trainium is 20,000,000,000.

所以这肯定对营收数字有帮助。

So definitely helps those revenue numbers.

但在更广泛的行业采用方面，我觉得关键问题其实是：我们真的能用得上这个吗？

But in terms of the like broader industry adoption story, I think it's really kind of like, can we use this?

你懂的？

You know?

对。

Right.

我的意思是，至少在开发者群体中，大家的感觉是，你知道的，它

I mean, perception, certainly among developers, like, you know, it's

排序

sort

收入和口碑、态度有时候是两码事，光靠几笔大单子，很难撑起持续的需求故事。

of almost revenue and vibe or attitude are sort of two separate things sometimes, and oftentimes you can't just bank on a couple deals for a sustainable demand story.

你得真正在开发者社区里下功夫，就像亚马逊在做的那样。

You have to really cater to a community of developers like you're saying Amazon is doing.

我觉得这一点特别关键。

So I think that's a really important point.

我想快速问你一下，Catherine，关于昨晚看到的另一个消息。

I want to ask you, Catherine, really quickly about another story we saw last night.

谷歌和黑石正在合作打造他们自己的云计算业务，这挺有意思的，因为谷歌一直想把他们的TPU推到市场上。

So Google and Blackstone are partnering together to create their own type of cloud computing venture, and this is something that's interesting because Google has been trying to get its TPUs out there on the market.

我们这个节目聊过很多次了。

We've talked a lot on this show.

我们几乎问过每一位新兴云服务商的高管：你们会用TPU吗？

We've asked almost every Neo Cloud company executive here on the show, will you use TPUs?

答案总是：我们对英伟达的GPU挺满意的。

And the answer is, well, we're okay with our Nvidia GPUs.

你刚报道完Trainium的故事，对这件事有什么看法？

I wondered what your reaction to this story was, given that you're coming out of reporting a story on Trainium.

说实话，我也不太确定。

You know, I don't know.

你觉得亚马逊长远来看，会不会也跟其他公司做类似的合作？

Could you see Amazon doing something like this with another provider in the long run too?

是的。

Yeah.

上个月贾伊在致投资者的信里也提到，他们未来有可能把芯片卖给云服务以外的客户。

You know, Jai also said last month in his investor letter that they could envision selling chips someday outside of their own cloud service.

看来他们确实在考虑这件事。

So it seems like it's something they're thinking of.

不过，我也不太清楚他们具体会怎么操作。

I I, you know, I don't know how they would structure that.

会不会是那些自建数据中心的公司直接采购？

Could that just be in individual companies that own data centers?

会不会是和黑石这样的金融合作伙伴合作呢？

Could it be with a, you know, a financial partner like Blackstone?

这 definitely 是有可能的。

It's definitely possible.

我觉得得记住，亚马逊是最大的云服务商，而谷歌大概还是排第三。

I do think it's good to keep in mind that Amazon is like the biggest cloud and Google is sort of like, I'm pretty sure it's still number three.

所以我在想——

So, you know, I I wonder- Do

你真这么认为吗？

really say that?

别忘了，谷歌的立场很清楚。

It's just remember Google's Right.

远远落后。

A distant

亚马逊确实

bird Amazon did

没人让我这么说，但他们确实希望——嗯，没错，就是这么想的。

not pay me to say this, but, they would like, but yeah, yeah.

我觉得他们更喜欢自己掌控整个流程。

So I think that they like to like control a lot of the process themselves.

不过说实话，听到他们居然会考虑让芯片脱离自己的云服务，我还挺惊讶的。

But yeah, I was surprised to hear that they would even like talk about like, you know, having chips exist outside of their own cloud service.

我有点怀疑，他们真会把芯片放进谷歌云或者微软云吧？毕竟那是他们的竞争对手。

I kind of doubt that that would mean that they would put chips in like Google Cloud or Microsoft because those are their competitors.

所以，他们找别的合作伙伴来做这件事，倒也说得通，明显是看中了背后的商机。

So it kinda does make sense that they would maybe work with some sort of other partner to to to do this because it seems like they they see the dollar signs.

他们看得见这个故事的价值。

They see the good story.

他们不想让谷歌独占风头。

They don't want Google to get all the shine.

不过这我也说不准。

But don't know.

他们自己运营数据中心也很在行，可能根本不需要外人帮忙。

They're also really good at operating data centers themselves and might not need the help of another

资产管理人。

asset manager.

对。

Yeah.

凯瑟琳，非常感谢你今天来做客。

Well, Catherine, I wanna thank you for coming on.

这位是凯瑟琳·佩洛夫，我们《信息》杂志的亚马逊记者。

That is Catherine Perloff, our Amazon reporter here at The Information.

随着前沿人工智能模型的成本持续高企，开源模型正越来越受到关注。

As the cost of frontier AI models remains an issue for tech companies, open source models are getting a lot more attention.

我的同事斯蒂芬妮·帕拉佐洛在我们的《AI动态》通讯里写过这个话题，我想请她来分享一下她的见解。

My colleague Stephanie Palazzolo wrote about that in our AI agenda newsletter I want to bring her on to share with us what she knows.

斯蒂芬妮，欢迎再次回到节目。

Stephanie, welcome back to the show.

很高兴你能来。

It's great to have you here.

先给我们介绍一下背景。

Set the stage for us.

最近开源模型领域都在发生什么？

What is going on in the land of open source models these days?

对。

Yeah.

所以正如阿卡什提到的，很多公司的AI成本都在上涨，这已经不是什么秘密了。

So Akash, as you mentioned, it is definitely no secret that the costs of AI are going up for many companies.

事实上，今年早些时候，Uber的首席技术官就告诉我们，像Uber这样技术实力雄厚的公司，仅仅几个月就把AI预算花光了。

In fact, earlier this year, the Uber CTO actually told us that the company, you know, even a company sophisticated as Uber have basically blown through their AI budget in just a matter of months.

为了应对这种情况，许多公司开始转向这些更便宜的开源模型，比如月之暗面的Kimi K 2.6或者DeepSeek v4。

And so to kind of counteract this, many companies are switching to these cheaper open source models, models like Kimi K two point six from Moonshot AI or DeepSeek v4.

虽然这些模型的能力可能不如前沿的闭源模型，但它们依然价格便宜，非常适合用来自动化处理一些简单的任务。

And even though these models are maybe not as capable as frontier closed source models, they are still again cheap and really good for automating more simple tasks.

所以你刚才说这些开源模型适合做简单任务，是吧？

So are these open source models you said they're good for simple tasks.

我想弄清楚，简单任务和大型科技公司需要它们完成的任务之间到底有什么区别。

I'm trying to understand the delta between the simple tasks and then the things that big public tech companies need them to do.

你提到了优步。

You mentioned Uber.

那这些开源模型够不够好，能应付大科技公司的需求吗？

I mean, is there a clash of open source models that is good enough to do what these larger tech companies need them to do as well?

是的。

Yeah.

这其实要看具体情况。

So again, it really depends.

比如，一个开源模型能做的简单任务，可能是查看客户支持邮件，然后把它们分类。

Like, you can think of maybe a simple task that an open source model could do is maybe looking at customer support emails and categorizing them into different categories.

比如说，有人要退款，或者有人对配送司机有意见。

So like maybe people that want a refund or people that had a problem with their delivery driver.

接下来的步骤，比如规划下一步该做什么，或者回复这些邮件，就可以交给更先进的AI模型来处理，毕竟它们的能力更强。

And then maybe the next step of the next step of planning what to do next or responding to those emails could be done with a more, you know, with a, with a better frontier AI model that's able to handle that.

所以没错，像Uber和Airbnb这样公司的高管们确实公开表示，他们一直在大量使用开源AI。

And so yes, like a number of execs from companies like Uber and Airbnb have really come out and said, we've been using open source AI a lot.

它在自动化一些相对简单的任务方面非常有用。

It's super helpful for automating maybe these more simple tasks.

不过最近我们也听到开发者们反映，开源模型在某些方面仍然存在不足。

But we have also heard from developers more recently that there are still areas where open source is falling short.

比如，我之前跟一家 OpenAI 和 Anthropic 的大客户的高管聊过，他一直在尝试使用 Kimi K 2.6 和 DeepSeek v4 这类开源模型。

So for instance, I spoke with an exec at a big customer of OpenAI and Anthropic who's been trying to use open source models like Kimi K 2.6 and DeepSeek v4.

这些名字念起来真拗口。

These names are really a mouthful.

不过你每次都能准确念出来。

You nailed them every time though.

说实话，你对这些名字挺熟的。

Mean, you're pretty good at them.

但这位高管告诉我们，尽管这些模型在基准测试和表面问题上表现不错，但一旦你追问下去，或者问一些更深入的问题，它们就开始吃力了。

But basically this exec was telling us that even though these models perform well on benchmarks and kind of surface level questions, they do start to struggle when you try to ask some follow-up questions or these kind of deeper lines of questioning.

那为什么会这样呢？

Why is that though?

你刚才提到的这个细节挺有意思的。

Mean, this was an interesting line of detail that you had in your story.

我很好奇，为什么这些开源模型在实际使用中表现没那么好？

I mean, looked at why it is that these open source models aren't as good in real life.

你能稍微解释一下，这些模型的训练方式有什么不同，导致了这样的结果吗？

Can you explain a little bit about how these models are trained differently and why that might be the case?

是的。

Yeah.

当然，我并不在这些公司内部，所以没法确定，但我们有几个推测，可能解释了为什么会这样。

So obviously, I'm not at these companies, so we don't know for sure, but we've a couple of guesses of why this might be the case.

我们其中一个猜测是，这些模型的训练数据覆盖面还不够广。

So one hypothesis that we had is that the models training datasets don't have enough coverage.

用大白话说，这意味着在训练这些模型时，你并没有给它们展示足够多的例子，让它们了解在现实世界中可能遇到的各种情况和任务。

Basically what that means in normal English is that whenever you're training these models, you're not really showing them enough examples of all the different sorts of situations and tasks that they might encounter out in the real world.

举个例子，如果一个模型主要是在公共网络上训练的，那么在网络上会有很多类似入门博客文章的内容，用来解释一些概念，比如这里有一份关于模型训练技术的说明。

So for instance, if a model is mostly trained on the public web, on the web there are lots of examples of these kind of like intro blog posts that are explaining concepts like, you know, here's an explanation of this type of training technique for models.

但网络上其实并没有太多关于更复杂情况的例子。

But the web doesn't really include lots of examples of more complex things.

比如，你知道的，如果某种模型训练技术出了问题，一个人该怎么办？

Like for instance, you know, what should a person do if one of these training techniques for models like goes wrong?

另一种可能性是，这些模型在训练时采用了业界常说的「前沿模型蒸馏」方法，但如果操作不当，反而会适得其反。

Another possibility is that these models were trained by basically what people call distilling frontier AI, which can backfire when it's not done correctly.

不过总结来看，斯蒂芬妮，这些开源模型现在是不是还在持续迭代进步呢？

Taken together though, Stephanie, I mean, these open source models getting better?

还有你专栏里提到的另一个问题，现在顶尖大模型和开源模型之间的差距是在缩小吗？

Mean, this was the other part of your column was that the delta between the frontier models and then the open source models, is it narrowing?

还是在不断拉大？

Is it widening?

所以有一些证据表明，开源和闭源人工智能之间的差距实际上在拉大。

So there is some evidence that that gap between open and closed source AI is actually getting bigger.

美国国家标准与技术研究院（NIST）做了一项分析，追踪了不同模型随时间的性能变化。

And so there was an analysis done by by NIST where they basically charted the performance of various models over time.

在一个理想、支持开源的世界里，你希望看到这两条趋势线保持平行，甚至逐渐靠拢，最终开源模型能追上来。

And, you know, in a kind of perfect world that's like pro open source, you would want to see those two trend lines either, you know, kind of moving in parallel or maybe even kind of converging over time where you see them kind of meeting and open source catching up.

但这项分析显示的恰恰相反。

But instead what we saw in this analysis is the exact opposite.

这两条趋势线实际上正在逐渐分离，它们之间的差距随着时间推移越来越大——如果你是开源的坚定支持者，认为开源最终能赶超，这显然不是你希望看到的局面。

So those two trend lines are actually kind of diverging, and the gap between them is getting wider over time, which is not what you wanna see if you are somebody who's a big proponent of open source and thinks that open source is gonna catch up in the long run.

接下来，你关注这个话题的哪些动向，Stephanie？

What are you watching for, Stephanie, on this story moving forward?

一种看法就是，我们只看这些基准测试数据怎么变化。

I mean, one way to look at it is we just see how the benchmarks do over time.

但正如你刚才说的，基准测试并不能真实反映模型在现实中的表现。

But as you just said, the benchmarks are not really the best indication of reality and what they can do in real life.