谷歌AI大变动，英伟达考虑在Rubin芯片中减少内存，Meta推出新AI编程工具

欢迎大家收听Information的TITV节目。

Welcome everyone to the Information's TITV.

我是Stephanie Palazzolo，今天是8月6日，星期四。

My name is Stephanie Palazzolo and it's Thursday, August 6.

今天节目中，The Information独家报道了英伟达针对内存危机的解决方案，具体来说，就是在其新的Rubin芯片中减少内存的使用。

Today on the show, the Information has exclusive reporting on Nvidia's proposed solution to the Memory Crunch, specifically to incorporate less memory into its new Rubin chips.

接下来，我们来看看Meta的新编程工具如何与Anthropic的Claude Code和OpenAI的Codecs竞争。

We'll then take a look into how Meta's new coding tools aim to compete with Anthropic's Claude Code and OpenAI's Codecs.

我们还会解读谷歌的AI领导层变动，以及这对公司未来可能意味着什么。

We'll also unpack the AI leadership shakeup at Google and what it could mean for future of the company.

节目最后，我们将与投资于谷歌前首席科学家Jeff Dean新创公司的风险投资家们进行对话。

And to close out the show, we'll talk with venture capitalists investing in the new startup founded by Google's former chief scientist, Jeff Dean.

这将是一期精彩的节目，让我们马上开始吧。

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

英伟达正在考虑一个激进的办法来应对先进高带宽内存芯片短缺的问题。

Nvidia is weighing a radical step to deal with a shortage of advanced high bandwidth memory chips.

我们的同事刘菲和刘晨报道了这个消息。

Our colleagues Phoebe Liu and Chenner Liu broke the story.

现在菲比也来到了我们节目。

Phoebe is here to join us now.

欢迎来到节目。

Welcome to the show.

谢谢邀请。

Thanks for having me.

那么菲比，英伟达到底是怎么应对内存危机的呢？

So Phoebe, how exactly is Nvidia dealing with the memory crunch?

对。

Yeah.

我们的报道发现，英伟达正在测试其下一代GPU——Rubin Ultra——的版本，这些版本的内存比最初计划的要少。

So our reporting found that Nvidia is testing versions of its next generation GPU, so Rubin Ultra, with less memory than it initially planned.

我记得在2025年初，黄仁勋说过Rubin Ultra每颗芯片会有1TB的内存，这非常多，比它生产过的任何芯片，甚至比任何人生产过的任何芯片都要多得多。

So I think in early twenty twenty five, Jensen Huang said that Rubin Ultra would have one terabyte of memory per chip, which is a lot, like, a lot more than any chip it's ever produced any chip anyone has ever produced.

自那以后，内存紧张的情况变得相当严重，部分原因是供应链压力，比如英伟达购买了太多先进的高带宽内存，供应商很难生产出足够的量；部分原因可能也是因为客户对内存稍少一些的芯片同样有需求。

The memory crunch has gotten pretty intense since then and partially because of supply chain pressures, like Nvidia's buying so much advanced high bandwidth memory that it's just hard for the suppliers to make enough of it, partially because of that, partially potentially because customers have just as much use for chips with a little bit less memory.

英伟达正在测试内存为256GB甚至192GB的版本，并已提供给一些客户，但重要的是，目前一切都还没有最终确定。

Nvidia is testing versions with two fifty six gigabytes of memory and even one ninety two with some customers, although it's important to know that nothing has been finalized yet.

他们现在只是在测试阶段。

They're just testing for now.

而且，正如你提到的，有点好笑的是，我们之所以陷入这种局面的一个重要原因，是因为在过去一年里，英伟达的GPU变得如此受欢迎，以至于在某种程度上导致了这种内存紧张。

And it's kind of, you know, kind of funny, as you mentioned, that one of the big reasons why we're in this situation to begin with is because over the last year, you know, Nvidia's GPUs have become so popular that in some ways it's kind of led to this memory crunch.

所以这看起来有点像是一个两难困境。

So it's a bit of a catch 22, it seems.

是的。

Yeah.

我也这么认为。

I think so.

而且我觉得英伟达实际上有一个团队，会提前多年研究供应链以及原材料的问题，并根据公司对每种材料需求量的预测来规划，因为作为全球最大的公司之一，它需要的量太大了，有时甚至超过了全世界的总产量。

And I think Nvidia actually has a team that kind of looks at these supply chain and all the way down to the raw materials issues years ahead working with kind of projections the company has for how much of each material it will need because it is one of the world's largest companies and needs so much of this that it sometimes literally wants more than the entire world can produce.

所以他们在这方面考虑得挺多的。

So this is something they think about quite a bit.

但即使他们能做的规划再多，有时候也没法控制流程中的每一个环节，尽管我相信他们很想这么做。

But even with all the planning they can do, sometimes they can't control every single step of the process, even though I'm sure they would like to.

那么，内存减少实际上会怎样影响这款新芯片的性能呢？

And, you know, how does having less memory actually affect this new chip's performance then?

对。

Yeah.

我觉得，因为如今最先进的AI模型都非常庞大，比如有数万亿的参数，训练和运行都需要大量内存。

So I think because today's most advanced AI models are so big, like they have trillions of parameters, that takes a lot of memory to both train and run.

所以，如果某款芯片的高带宽内存较少，要在其他条件相同的情况下运行同等性能的模型，就需要更多的芯片，这其实对英伟达来说算是好事，因为他们反过来也能卖出更多的GPU。

So if one specific chip has less high bandwidth memory, to run a model with the same level of performance, if all other factors were held equal, you would need more chips, which honestly is kind of good for Nvidia because they could also, in turn, sell more GPUs.

但我敢肯定客户对这个想法并不怎么高兴。

But I'm sure customers aren't thrilled about that idea.

对吧？

Right?

感觉有点像，不知道，他们会不会觉得有点措手不及，然后想，哦，现在我得买比我原来想的多得多的芯片了。

It kind of feels like, dunno, do they feel like the rug has kind of been pulled out from under them a little bit and they're like, oh, now I need to buy like so many more chips than I than I originally thought.

是的，显然我没跟每个客户都聊过，但我聊过的那些客户，说实话，不觉得这会影响需求。

Yeah, so obviously I haven't talked to every customer, but the ones that I did talk to honestly don't think it will affect demand.

他们反正还是会想要这些芯片的。

They're going to want the chips anyway.

而且，希望事情能像我预期的那样发展，如果芯片内存少了，价格也会更便宜。

And, hopefully, if all things shake out as I'm expecting it to, if chips have less memory, they will also be cheaper.

我不知道成本下降是不是一比一的，但这对客户来说会是好事，尤其是那些大型云服务商，他们越来越担心在英伟达GPU和AI上的开销。

I don't know if that will be kind of a one to one cost decrease, but that would be good for customers, especially, like, the big clouds that are getting more and more worried about how much they're spending on Nvidia GPUs and AI in general.

所以听起来，即使他们得买更多芯片，希望也能得到一些折扣。

So it sounds like even if they have to buy more chips, hopefully they can get a bit of a discount on them.

我觉得是这样。

I think so.

不过还有一点很重要，英伟达还没给Rubin Ultra定价。

Although also important to know that Nvidia hasn't set pricing for Rubin Ultra yet.

我还挺好奇的，在这种情况下，正如你所说，内存对于运行和训练这些巨型模型显然非常重要。

And I'm also curious too, having this situation, as you said, memory is obviously very important to run and train these giant models.

你觉得这会不会促使客户转向非英伟达的芯片呢？

Do you think this is going to push customers at all towards maybe non Nvidia chips?

显然我们有TPU，有大型云服务商的芯片，还有那些试图制造芯片的初创公司，你知道的，有几十家。

Obviously we have TPUs, we have chips from the big clouds, we have all these, you know, dozens of startups that are trying to build chips as well.

这足以推动客户转向它们吗，还是说这其实不是什么大问题？

Is this enough to push customers towards them at all or is it not really that big of a deal?

我觉得这可能不是什么大问题。

I think it might not be that big of a deal.

我觉得这取决于英伟达最终会怎么做，以及如果它真的决定大幅降低Rubin Ultra芯片的内存，那他们会做哪些其他优化来弥补这些芯片的性能，这样他们就能……是的。

I think it depends on what Nvidia ends up landing on and kind of if it does decide to go with pretty heavily downgraded or less memory with Rubin Ultra chips, what other optimizations they would make to kind of compensate for those chips performance so they can yeah.

他们还有很多其他事情可以做，比如在芯片的不同部分和数据中心之间更快地移动数据，还有更高效地存储数据。

There there's a lot of other things they can do, like moving the data faster between different parts of the chip and the data center and also kind of storing the data more efficiently.

他们能做的还有很多。

There are lots of other things they can do.

所以我觉得这取决于芯片的整体性能和整体成本。

So I think it just depends on kind of both the overall performance and the overall cost of the chip.

完全同意。

Totally.

而且我知道你提到过，显然这里还有很多变数。

And I know you mentioned obviously it sounds like there are still moving pieces here.

他们还没有最终确定，显然也还没有公布价格。

They haven't necessarily finalized or obviously come out with pricing yet.

我的意思是，英伟达有没有可能在芯片更广泛上市之前，通过某种方式获得更多内存，或者他们还在考虑其他什么办法来应对内存紧张的问题？

I mean, there any way for Nvidia to try to get more memory between now and when they make these chips more widely available or any other things that they're considering to kind of get around this memory crunch at all?

是的，我觉得他们正在尽一切可能来实现这个目标。

Yeah, think they're doing literally everything in their power to make that happen.

所以我想说，要么上周，要么上上周，他们宣布与SK集团延长了5000亿美元的合作。

So I wanna say either last week or the week before, they announced an extended $500,000,000,000 partnership with the SK Group.

SK集团是全球最大的高带宽内存制造商之一的母公司，他们将共同开发下一代高带宽内存。

So that's the parent of one of the biggest high bandwidth memory makers in the world to work on developing this next generation high bandwidth memory together.

具体这5000亿美元会怎么分配还不清楚，但基本上就是SK海力士这家内存制造商，会大幅增加这种先进内存的产能，靠的是额外的投资。

It's kind of unclear exactly how the 500,000,000,000 will shake out, but it's basically SK Hynix, the memory maker, is going to increase its capacity for making this advanced memory pretty significantly with additional investment.

他们的工程师正在和英伟达的工程师合作，共同设计下一代内存，这是英伟达最喜欢用的词之一。

And their engineers are working with Nvidia's engineers to kind of codesign, one of Nvidia's favorite words, that next generation memory together.

所以，是的，情况确实可能会变。

So, yeah, definitely things could change.

而且，英伟达也有可能先推出内存稍小一点的版本。

And, also, it's possible that Nvidia starts with versions that have a little bit less memory.

然后随着他们继续生产更多的Rubin Ultra芯片，我记得那要到明年年底才开始。

And then as they continue to produce more Rubin Ultra chips, I think that's set to start late next year.

他们可以根据客户的需求以及供应链的供应情况，推出各种不同内存规格的版本。

They could create lots of different versions with different memory specs depending on what customers want and depending on kind of what's available in the supply chain.

所以听起来，即使他们现在面临这种情况，他们也在尽一切努力避免将来再发生。

So it sounds like even if they're in this position now, they're doing everything that they can to avoid this happening in the future.

太好了。

Great.

好的。

Okay.

好的，非常感谢你，菲比。

Well, thanks so much, Phoebe.

我们一定会持续关注这些新英伟达芯片的最新动态。

We'll definitely keep in touch on everything that's going on with these new Nvidia chips.

再次介绍，这位是《The Information》负责英伟达报道的记者菲比·刘。

But again, that's Phoebe Liu, our Nvidia reporter here at The Information.

Meta 在周三发布了其新的 AI 编程代理 Muse Code，以及为其提供支持的最新 AI 模型 Muse Spark 1.2。

Meta unveiled its new AI coding agent, Muse Code, on Wednesday alongside its latest AI model, Muse Spark 1.2, that powers it.

我们的下一位嘉宾提前体验了这些产品。

Our next guest had early access.

现在加入我们的是 Creative Strategies 的分析师马克斯·温巴赫。

Joining me now is Max Weinbach, an analyst at Creative Strategies.

马克斯，欢迎来到节目。

Max, welcome to the show.

很高兴你能来，谢谢。

It's great to have Thank

谢谢你的邀请。

you for having me.

那你觉得 Muse Code 怎么样？

So what did you make of Muse Code?

这是一个有趣的编程工具。

It is an interesting new coding harness.

Meta 真的很想进入编程这个领域。

Meta is really trying to get into this, coding game.

编程代理，我觉得是吸引资金和兴趣的主要驱动力，而且它们也能推广到一般的工作场景中。

The coding agents are, I guess, the big driver of interest money, and they do generalize into just general workplace as well.

但这是一个不错的编程工具，由 Muse Spark 2.1.2 驱动。

But it was it's a good coding harness, powered by Muse Spark two point 1.2.

它让我想起很多，不是最前沿的，但在模型上可能比大实验室落后一代或两代。

It's, it reminds me a lot of, a lot of the not frontier frontier, but that's a step behind maybe a generation or two model wise from the big labs.

这是一个不错的编码工具。

It's a good good coding harness.

好工具。

Good tool.

对。

Mhmm.

听起来结果不错，但也许并没有让你特别惊艳？

So sounds like positive results, but maybe you're not blown out of the water per se?

我觉得它有点像GPT-5.2的水平。

I would say it reminded me of something closer to a GPT five dot two.

需要很多人工指导。

It needs a lot of handholding.

结果可能不错。

The results can be good.

可能会非常好。

They can be very good.

它在设计方面不太行。

It's not great at design.

这似乎是Anthropic模型特别擅长的地方，而它需要很多很多的指导、细节和方向。

That really seems to be something that Anthropic models are great at, and it needs a lot of a lot of handholding, a lot of detail, and direction.

新模型（比如Cloud 5系列和GPT 5.6）的一个主要特点是，它们大致能理解你想让它做什么，你给一个非常简短的提示，它就能自己跑起来。

One of the major things of the newer models, the Cloud five family as well as GPT 5.6, is you can it kind it understands more or less what you're trying to tell it to do and just knows you can get a very brief prompt and it will go.

这个模型需要很多细节。

This one needs a lot of detail.

你真的需要手把手地引导它，才能让它完成该做的事情。

You need to really try to hold its hand to push it through what it needs to do.

所以感觉它落后了一两代，尽管它能力很强。

So it does feel a generation or two behind even though it's very capable.

听起来它有些不足之处，比如你给它一个非常简短的提示，它就能自己发挥。

So sounds like maybe some some places where it falls a little short is this, you know, ability to give it a really brief prompt and have it run from there.

你还提到了设计方面的问题。

And also you mentioned design as well.

还有哪些地方你觉得它可能不太行，或者哪些地方它特别突出，让你印象深刻？

Are there any other areas where you felt like maybe it fell short and where did it maybe stand out and really impress you?

要说它哪里不行，它确实有个毛病，尤其是在软件开发方面，它不太愿意用真实数据。

It's really to to fall short, it does have this tendency, especially for software development, of not wanting to work with real data.

它特别喜欢自己编造模拟数据，所以想在实际应用里用它就特别难，因为它总想造假数据，而不是用真实数据。

It loves to create its own mock data, so it makes trying to actually use it in an actual application very difficult because it's trying to fake the data rather than using real data.

比如你想创建用户账户，它就做不好。

So if you're trying to make user accounts, it won't do it well.

这又是老一代模型的问题了。

And that's kind of, again, a generation or two old model thing.

大多数新模型都不会这样。

Most of new ones don't do it.

它真正擅长的是当个勤恳的干活的，就像DeepCV四Flash、GPT 5.6 Luna那样，你给它一个小而具体的任务，比如去更新一些东西，它能做得很好。

Where it is just really good is it's a good workhorse in the same way DeepCV four Flashes, GPT 5.6 Luna, where if you give it a small detailed task like going to update some stuff for you, it will do a great job at it.

它们能做这些小事，而且你可以为它用很多子代理，这很棒。

They can do these little things, and you can use a ton of sub agents for it, which are great.

它非常快。

It's very fast.

每秒大约180次，而且根据你使用的版本，价格非常便宜。

It's about 180 per second, and it's, depending on the version you use, really cheap.

对。

Mhmm.

好的。

Okay.

抱歉。

I'm sorry.

我得双击那个。

I have to double click on that.

假数据是怎么回事？

What's up with the fake data?

比如，在什么情况下它会生成假数据？我不太清楚。

Like, in what situations is it even making fake data and like what I don't know.

训练数据里有什么东西吗？

Is there something in the training data?

比如，显然我们是在猜测，但什么可能导致这种情况呢？

Like, obviously, we're speculating, but what what could cause that?

所以这不是什么，这是正常操作。

So it's not it's a normal practice.

如果你没有完整的数据库设置，它喜欢生成模拟数据，这样它就能创建界面，确保所有逻辑正常，并完成所有这些。

If you don't have a full database setup doing whatever, it likes to do mock data so it can create its UI, make sure all the logic works and do all of this.

这是很正常的事情。

That's a normal thing.

但是很多新模型，比如GPT和Claude，甚至其他一些模型，在Cursor的辅助下，能很好地引导出这种行为。

But a lot of the newer models, again, the GPTs and the clods and even some of the other ones in cur their cursor's harness does a great job of kinda pulling this behavior out of it.

它会开始真正尝试分离这些，试图模仿你在实践中真正想做的，而不是仅仅为了先让它工作而构建。

It it will start to actually try to separate it, so it's trying to mimic the way you would actually wanna do something in practice rather than just building it to make sure it works at first.

它通常会把模拟数据放到不同的文件里，这样就不会弄乱代码。

And the way it will do it is usually putting it off to different files so it's not muddying up code to do it.

Muse Spark 不会那样做。

Muse Spark doesn't do that.

它有点像，再说一次，我不会这么说，但它让我想起了大约九个月前我们在模型中看到的东西，然后试图去做。

It kinda takes this, again, I wouldn't say it, but it reminded me of stuff we saw about nine months ago in the models and try to do it.

但你是在把九个月前的那种奇怪行为，和今天非常强大的模型混在一起，结果就是过去一年里我们在很多不同模型中看到的各种行为的一个奇怪混合体。

But you're taking this weird behavior from, like, nine months ago, mixing it with a very capable model today, and it's just kind of a weird amalgamation of different behaviors we've seen over many different models over the past year or so.

它挺好的，你知道，如果你把它放在正确的方向上，它是有能力的。

It's good and, you know, capable if you put it in the right direction.

只是感觉有点不一样，在用过来自其他前沿实验室的明显更好的模型之后，很难喜欢上它。

Just kind of feels different and hard to hard to like it after using the significantly better models from the other frontier labs.

那么，就不同的编码代理而言，显然我们有，你知道，Cloud Code、Codex、Cursor，现在还有 Muse Code。

So in terms of the different coding agents, obviously we have, you know, Cloud Code, Codex, Cursor, now Muse Code.

我的意思是，我知道你最近才开始测试 Meta 的新产品，但你对这些工具在哪些用例下你会用哪个，有没有一个概念？

I mean, I know you have only recently kind of started testing the new Meta products, but do you have a sense of, you know, which one of those you're reaching for for which use cases?

我基本上都是换着用的。

I kind of use them all interchangeably.

我觉得Anthropic的模型擅长处理大型任务，就是那种非常大的项目。

I I view the Anthropic models are good at large, just like very large thing.

如果你要处理很多文件，Claude模型很擅长创建工作流，还有子代理来检查所有内容。

If you're trying to do something over a lot of files, the cloud models are very good at making workflows and having sub agents to go over everything.

如果我从零开始，或者需要非常复杂的东西，我只想要一个模型，我可以给它一个任务，相信它能搞定。

If I'm starting from scratch or need something super complicated and I just want a model, I can give it a task and trust it'll figure it out.

那基本上就是Codex和ChatGPT。

That's kind of codex in chat GBT.

Cursor同时有Claude和Codex，但它的框架感觉更贴近我实际认为的真正工程师的工作方式。

Cursor has both Claude and codex, but its harness feels more directed towards the way I would I actually think real engineers do work.

所以它更有方向性，带有这些功能，而其他工具，Codex确实很好，但它也适用于任何任务。

So it's kind of more directional with these features in there versus the others, which codex is really good, but it's also for anything.

这让它有点奇怪，你不太会像信任Claude那样信任它去做大型重构。

And that makes it kind of funky where you don't necessarily trust it as much to the large refactors you would with Claude.

所以我全都用，我觉得它们各有各的用武之地。

So it it I use all of them, and I feel like they all have a different place.

而Muse Code，我觉得它更像是介于Cloud Code和Codex之间的东西，但它的模型能力不一定有那两个强。

And the Muse Code kinda fills, I would say, something closer to, like, in between Cloud Code and Codex, but it's not necessarily as capable of a model as the two of them.

所以，我可能更倾向于在Cursor或者其他编码工具里用Muse Spark 1.2，而不是直接用Muse Code。

So maybe I would probably just use Muse Spark 1.2 in Cursor or one of the other coding harnesses rather than using Muse Code itself.

当然。

Sure.

显然，性能只是问题的一个方面。

And so obviously, you know, performance is kind of one side of the equation.

另一方面就是成本、速度这些你刚才提到的因素。

The other side is, you know, cost, speed, some of these that you've already talked about.

从他们公布的定价来看，Muse Spark 1.2比大多数Anthropic和OpenAI的模型都便宜，除了他们最轻量的那些选项。

It looks like from the pricing they've released that Muse Spark 1.2 is cheaper than most Anthropic and OpenAI models other than kind of their most lightweight options.

也就是GPT 5.6 Luna和HYCU 4.5。

So that's GPT 5.6 Luna and HYCU 4.5.

你觉得新的Meta模型在成本方面怎么样？

How did you feel about the new Meta model cost wise?

我其实昨天才看到价格，挺有意思的。

I didn't actually see the cost until yesterday, which was interesting.

但我觉得贡献者层级，就是如果你分享数据给他们，他们会给你95%的折扣，这真的很划算。

But I think the contributor tier, which is they give you a 95% discount if you share your data with them, is a really good deal.

我日常做的很多事情，其实我不太在意。

A lot of the things that I'm doing day to day, I don't necessarily care.

Meta会拿这些数据做训练。

Meta gets the training data for it.

虽然这些数据是内部专有的，但如果将来能帮我用这些模型获得更好的性能，我不介意。

It's it's, you know, internal proprietary, but if that can then help me in the future using these models get better performance out of it, I don't mind.

而且我会接受那95%的折扣。

And I'll take the 95% discount on it.

对。

Yeah.

在付费层级上，我可能还是会选择那些高端的顶尖模型，因为到了那个价位，比如每百万输出token 4.2美元，它确实很强大。

On the the pain tier, I think I would still probably go for one of the the higher end frontier models just because it gets into that weird point where, yes, it's capable at $4.20 per million output tokens.

我觉得大概是每百万输入一美元吧。

I think it's like a dollar per million in.

而且，是的，它确实有能力。

And, yes, it's capable.

它很好。

It's good.

但我也可以用Grok builds或者Cursor，那里我得到的是补贴很多的令牌，实际上我并没有按令牌费率付费。

But I can also use Grok builds and or Cursor where I'm getting very subsidized tokens on that, and I'm not actually paying the token rate for it.

所以你就处于一个奇怪的位置：是的，我可以作为API付费使用。

So you're kind of in this weird position where, yes, I can do pay for it as an API.

如果我不太在意，我可以得到很大的折扣，或者我可以去其他那些补贴很大的地方，每月花200美元就能获得价值1万美元的使用量。

I can get a massive discount if I don't really care, or I can just go to one of the other ones where it's hugely subsidized and I'm getting, dollars 10,000 worth of usage for $200 a month.

我很好奇，其他模型提供商是否愿意提供类似的商业模式，如果你愿意分享数据，就给你很大的折扣。

I'm curious if other model providers were willing to offer that same kind of, you know, business model where they give you a huge discount if you're willing to share data.

你觉得这对用户有吸引力吗？

Do you think that would be attractive to users?

或者你觉得，到了现在这个阶段，他们已经习惯了不分享数据，这样做可能会让他们反感？

Or do you think at this point they're already used to, you know, not sharing their data and it might rub them the wrong way perhaps?

我觉得Meta的做法其实挺好的。

I think the way Meta actually did is a great way.

它有两个不同的模型版本。

It's a two different model slugs.

你可以非常明确地选择。

You can choose very specifically.

要么按正常价格，我们不获取数据，要么你选择折扣价。

Normal normal price, we don't take it, or you can take a discount for it.

OpenAI，实际上，我相信他们还在这么做。

OpenAI, actually, I believe they still do this.

如果你愿意免费分享数据，他们会每天或每周给你一定数量的token。

They give you a certain amount of tokens per day or week if you're willing to share data for free.

我觉得他们已经这样做大概六到九个月了，或者可能快一年了。

They've been doing this, I think, for six to nine months now or may been almost a year at this point.

谷歌也提供了慷慨的免费额度。

Google gives you a generous free tier as well doing it.

很多公司都这么做了，但与其只是给95%的折扣让你买尽可能多的token，不如每天、每周或每月给你一定数量的免费额度。

A lot of them have done it, but instead of just doing the 95% discount for as many tokens as you're willing to buy, it was will give you a certain amount free per day, week, month.

这样你就同意让我们用你的数据来训练。

And with this, you agree to let us train on your data.

我觉得，这其实是做同一件事的两种方式。

I think that's, you know, that's two ways of doing the same thing.

我不介意。

I don't mind it.

我认为这是个好办法。

I I think it's a great way to do it.

毕竟，AI很贵。

I mean, AI is expensive.

大家都在说，知识便宜到不值一提。

Everyone's saying, you know, knowledge that's too cheap to even matter.

是的。

It's yeah.

它的智能就在那里。

It it's intelligence is there.

如果你能拿到折扣，人们显然会对这个新定价感到非常兴奋，尤其是深度求索还说要涨价。

If you can get the discount on it, people will obviously seem really excited by this new pricing, especially with deep sea saying they're gonna increase rates as well.

我敢肯定，这比每天给你有限免费令牌要好得多。

I'm sure that's a much better way of doing it than just saying you get a limited amount of free tokens per day.

而且，退一步看，看看所有这些不同的编程代理，包括Meta的新代理，我们显然在过去六到十二个月里看到了从编程助手到能够独立工作、有时甚至连续工作数小时的代理的转变。

And, you know, taking a step back, looking at all these different coding agents, including Meta's new ones, we're obviously seeing this shift in the last, know, six to twelve months from coding assistants into agents that can work, you know, independently, sometimes for hours at a time.

我很好奇，根据你的经验，现在用编程代理能做哪些三、六、九个月前做不到的事情？

I mean, I'm curious just in your experience, what are some things that you're able to do today with coding agents that maybe you weren't able to do, you know, three, six, you know, nine months ago?

我想说，如果你理解软件的工作原理，模型架构或软件架构，你几乎什么都能做。

I would say doing it would be you're able to do if you understand how software works, the model architect just architecture of software or the models, you could do essentially anything.

当然，在速度、时间、可靠性方面肯定有限制。

I mean, obviously, there's limitations to it in terms of speed, time, reliability.

但我觉得九个月前，我连现在做的一半都做不了。

But I don't think nine months ago, I would have been able to do half of what I'm doing now.

而且这很大程度上不只是因为模型更好了，代理更好了什么的。

And a lot of that isn't just because the models are better, the agents are better, whatever.

就是太费时间了。

It's just it would take so long.

以前基本上是一次只能处理一个文件。

Originally, it would be basically one file at a time.

如果你想跨多个文件做点什么，编辑三个文件可能要花十分钟，因为你得去用那个OpenAI o1或o3模型。

You want to do something over multiple files, it would take ten minutes to edit three files because you have to go into one of, like, the OpenAI o one or o three models.

它没法使用工具。

It wasn't able to use the tools.

虽然你有能力做，但模型、代理，它们配合得不太好。

While you were capable of doing it, it just the the models, the agents, the way they all worked didn't come together really well.

现在你可以做大规模的改动，用完全不同的语言重写东西，一夜之间效率提高一百倍。

Now you can do sweeping changes, rewrite things in completely different languages, make it a 100 times more efficient kind of overnight.

所以你现在进入了一个高度高效软件的领域，而九个月前我们可能只有软件。

So you're getting into this area of just highly efficient software, which maybe nine months ago, we had software.

现在，假设你知道如何使用一切，我们开始拥有优秀、高效的软件。

Now we're getting into good, highly efficient software, assuming you know how to use everything.

而且，你知道，也许再过九个月，我们就能进入这样的阶段：你只需给出一个简单的想法，就能在几小时甚至几分钟内得到你能想象到的任何东西。

And, you know, maybe another nine months from now, we're just getting into you can give it a quick idea of what you want and get anything you could ever imagine built for you in a couple hours, minutes even.

听起来

Sounds

对我来说是个美好的未来。

like a great future to me.

我想我们得在六个月或九个月后再来看看情况如何。

I guess we'll have to check back in in six, nine months, see what see what things look like then.

好的，再次非常感谢你，Max。

Well, again, thank you so much, Max.

刚才那是来自Creative Strategies的Max Weinbach。

And that was Max Weinbach from Creative Strategies.

谷歌周三宣布了其AI业务的重大变动。

Google announced major changes to its AI operations on Wednesday.

谷歌DeepMind的首席执行官戴密斯·哈萨比斯将卸任，转而担任该实验室的董事长兼Alphabet的首席科学家。

Google DeepMind CEO, Demis Asabas, is stepping aside to become Chair of the Lab and Chief Scientist at Alphabet.

接替他的是现任DeepMind首席技术官的卡拉伊·卡瓦丘格鲁。

Taking his place will be Karai Kavachuoglu, current CTO at DeepMind.

他将以高级副总裁的身份领导该实验室。

He'll now lead the lab as senior vice president.

一些人将这次人事调整解读为谷歌的坏消息，消息公布后，谷歌股价收跌。

Some interpreted the shakeup as bad news for Google and the stock closed in the red following the news.

但The Information的联合执行主编马丁·皮尔斯在他的晚间简报中写道，这些变动可能更多关乎Alphabet未来的继任计划。

But the information's co executive editor, Martin Peers, wrote his nightly briefing arguing that the moves could say more about future succession plans at Alphabet.

马丁现在加入了我们的节目。

Martin joins me now.

马丁，欢迎来到我们的节目。

Martin, welcome to the show.

斯蒂芬，你好吗？

Steph, how are you?

我很好。

I'm doing great.

所以——

So-

我得恭喜你，你把科尔的姓念得那么准，因为我完全不知道怎么念。

have to congratulate you on your pronunciation of Core's last name because I have no idea how to pronounce it.

没有啦。

No.

你不知道，今天节目前我练了多少遍。

You have no idea how many times I was practicing that before the show today.

那你觉得，这次高管变动到底说明了什么？比如，对谷歌和整个字母表公司未来的领导层有什么影响？

So what do you think exactly what do you think this executive shakeup could say about, you know, future company leadership for Google and Alphabet more more broadly?

嗯，之前有一种看法是，不管什么时候，现任CEO皮查伊卸任，接替他的可能是谷歌云的主管托马斯·库里安，或者是戴米斯。

Well, there had been a view that the, whenever, the CEO, the current CEO, Pichai, steps down, that he could be succeeded either by Google Cloud's chief, Thomas Purion, or by Demis.

但我觉得很明显，德米斯其实对管理整个Alphabet不感兴趣。

But I think it's pretty clear that Themis is not interested in actually running Alphabet.

他真正想做的是继续他过去一直在做的事，也就是作为一名备受赞誉的科学家。

He really wants to go off and continue what he has done in the past, which is a very acclaimed scientist.

我觉得那才是他真正感兴趣和充满热情的地方。

I think that's where his interests and passions really lie.

所以我认为他不会参与接任的竞争。

So I don't think he is in the run to take over.

而且每当这个问题被提出来的时候——可能短期内不会——Alphabet的董事会就得考虑其他人选了。

And whenever this issue comes up, which probably won't be for a while, the Alphabet boy will have to think of other people.

而且我觉得你也许能从昨天的变动中看出这一点，对吧？

And I think maybe you can see that even in yesterday's moves, right?

德米斯有点退居幕后，而卡莱真的站到了台前。

Dennis kind of taking a step back and Karai really stepping forward.

根据我们过去一年的报道，卡莱实际上一直是更深度参与日常运营的人。

As we've reported over the last year, Karai has really been actually the person more involved maybe in day to day operations.

你觉得这样理解情况公平吗？你觉得那里的人其实并不那么惊讶吗？

I mean, do you think that's, like, that a fair read of the situation and do you feel like people there are not really that surprised then?

是的，我觉得完全是这样。

Yes, I think that's absolutely the case.

我们去年报道过，实际上是科瑞在管理一切，而特米斯更专注于大局。

I I think what we had reported last year is that Coray was really the one running everything and then Themis was much more focused on the big picture.

虽然他在公众场合出现很多，但我觉得他实际上并没有真正参与日常运营。

I mean, while he appears a lot in public, I don't think he was really that involved in actually running the place today.

所以我觉得发生的一切只是他们公开明确了实际情况。

So I think all that has happened is that they've publicly made clear what was actually the reality.

所以我觉得其他媒体的报道有点过度了，尤其是一些新闻媒体说这是件大事。

So I think the coverage in other outlets was a little overdone, particularly some news outlets saying it's a huge deal.

我完全不觉得这是什么大事。

I don't think it was a huge deal at all.

不过，我想从另一个角度看，很多时候要激励AI研究人员，他们需要觉得自己有一个像灯塔一样的人物，可以指引他们，并且对AI的发展方向有一套哲学。

I mean, think just to take the other side of the argument though, I think you know, a lot of the times to inspire AI researchers, they need to feel that they have this sort of, like lighthouse type person that they can, you know, that's kind of guiding them and has this, philosophy for where AI is going.

你觉得丹尼斯在新的职位上还能提供那种指导吗？还是说……

I Do you feel like Dennis can still offer that in his new position or do

哦，我觉得完全可以。

you Oh, think absolutely.

我觉得事实上他们说的就是他会继续这么做。

I think that in fact that's what they were saying is that he will continue to do that.

所以我认为谷歌会受益于他的远见，而且他还会在谈论AGI发展方向等方面发挥作用。

So I think Google will get the benefit of his vision and the fact that he will play a role in talking about where AGI is headed, for instance.

他只是不再负责那些他本来就没在做的事情。

He's just not going to be responsible for the things that he was not already doing anyway.

所以，是的，官方说法就是这样。

So yeah, official.

对。

Yeah.

你提到过，每当你看谁可能接替现任CEO时，托马斯·库兰，也就是领导谷歌云的那位，似乎是热门人选之一。

So you mentioned whenever you look at people that could succeed the current CEO, it seems like Thomas Curran, who obviously leads Google Cloud, is one of the front runners there.

在我们的报道或讨论中，还有没有其他人可能成为Alphabet未来的CEO？

Are there any other people that have come up in our reporting or in discussions that could be potential future CEOs for Alphabet?

我听到的更多是那些据说不在候选名单里的名字。

I've heard more about the names of people who I'm told are not in the run.

我觉得有可能。

I think it's possible.

我的意思是，我确实想过Core是否可能成为CEO。

I mean, I did wonder whether Core could be a potential CEO.

我的意思是，昨天有人告诉我，嗯，也许将来某个时候吧。

I mean, somebody told me yesterday, Well, maybe at some point in the future.

但我真的不知道。

But I really know.

我不想把那些据说不在候选名单里的人的名字提出来。

Don't want to put forward names of people that I'm told are not in the running.

所以我会说，目前来看，Curian处于最有利的位置。

So I would say right now, Curian is in the best position.

但我还要说，这不是一个需要马上解决的问题。

But I will also say this is not an issue that has to be resolved anytime soon.

我觉得没有任何迹象表明CEO皮查伊有离开的打算。

I don't think there's any sign that the CEO, Pichai, has got any plans to move on.

他现在地位非常稳固，所以这个问题可能几年内都不需要解决。

He's in a very strong position right now, so this is not an issue that has to be resolved for years probably.

不过有意思的是，我们大约一年前对卡莱做过一次专题报道，从那次报道来看，他在公司内部确实非常懂政治，很擅长获得领导层的支持，同时也善于把团队拉到自己麾下，从而扩大影响力和权力。

I mean, it is interesting though that I think we obviously did a profile on Karai about a year ago and it did seem like from that he is, you know, super politically savvy within the company, very good at kind of getting support from leadership, but also moving teams under him and gaining more influence and power in that way.

所以我确实认为，我完全同意。

So I do think, you know, I totally agree.

现在考虑桑达尔的继任者确实为时过早，但卡莱看起来确实有足够的政治智慧，如果真到了那一步，他或许能成功。

It does seem way too early to even think about successor for Sundar, but Karai does seem like somebody who's politically savvy enough to maybe pull it off if were to come to

我觉得没错。

I think that that's correct.

我认为问题在于，既然他现在正式负责那个部门，谷歌的表现会如何。

I think it's just a question of how Google does under now that he is formally in charge of that division, how they do.

但我认为他绝对是我们应该关注的人。

But I think he's definitely somebody that we should be paying attention to.

听起来，根据你的分析，你觉得华尔街可能对昨天的消息反应过度了，因为那在公司内部可能已经是事实了。

I mean, sounds like from your analysis that you feel like Wall Street maybe overreacted to yesterday's news, given that it was perhaps already the reality within the company.

不过我在想，华尔街的反应也许还因为，你知道，谷歌去年十二月有过那个时刻，但之后可能并没有像人们预期的那样成为领跑者。

I guess I'm kind of wondering though if maybe the reaction from Wall Street was also maybe due to people, you know, Google maybe hasn't had this moment last December, but since then perhaps has not really pulled forward as the front runner as people thought they would.

所以我也在想，这些领导层变动也许被解读为谷歌在AI方面内部挣扎，或者没有像人们想象的那样领先。

So I guess I also wonder if maybe these leadership changes was maybe the- people were kind of reading that as like a sign of maybe Google internally struggling with AI or not playing to the front like people thought they were.

嗯，我觉得你得稍微分析一下。

Well, I think you have to unpack it a little bit.

首先，这些股票，你注意到了吗，它们每天都会大幅波动。

So for one thing, these stocks, I don't know if you've noticed, but they tend to move up and down a lot every day.

有些股票，比如SpaceX就是个很好的例子，一天跌10%，第二天涨，第三天又跌。

And there are some stocks, SpaceX is a really good example, 10% down one day, up the next day, down the next day.

它们就是不断上下波动。

I mean, they just keep moving up and down.

这是一个波动很大的时期。

It's a very volatile period.

所以谷歌的股票，我觉得今天早上稍微跌了一点，但我相信它肯定会反弹的。

So Google stock, I think this morning is down a tiny bit, but really, I'm sure it will bounce back.

我觉得这种反应可能是对头条新闻的回应，那些新闻对这一切都非常负面。

I think that the reaction was probably in response to the headlines, which were very negative about all of this.

我认为至少有很多分析师指出，谷歌是否拥有领先的AI模型，并不会真正决定它能否成功将AI转化为巨大的业务，因为它拥有极其丰富的资产，可以用来驱动AI，也可以被AI赋能，它比任何其他公司都更有优势，我真的不认为它必须拥有排名第一的AI模型，尤其是大家都似乎同意，你不需要在所有事情上都用最好的模型。

I think there are many analysts out at least who make the point that whether or not Google has got the leading AI model is not going to really determine whether it succeeds in turning AI into a giant business because it's got such a vast array of assets that it can use to power AI, it can be powered with AI, that it's much better off than anybody else and I really don't think it has to have the number one AI model, particularly as everyone seems to agree, you don't need really the best model for everything.

所以我的猜测是，投资者并不是根据这类事情来回应头条新闻的，这只是非常短期的反应，随着时间的推移，一切都会平稳下来。

So my guess is investors are not they respond on this kind of thing to the headlines, but it's just a very short term reaction and over time things smooth out.

所以是的，我不觉得这是什么大不了的事，而且我觉得随着时间的推移，人们会认识到这一点的。

So yeah, I don't think this is such a big deal and I think over time people will recognize that.

对，看来我们不应该对股票一天的表现过度解读。

Yeah, it seems like we maybe shouldn't read too much into the performance of a stock over one single day.

我觉得昨天除了我们刚才讨论的，另一个大新闻是谷歌DeepMind几位相当高层的人员离职，包括长期从事AI研究的科学家杰夫·迪恩，还有几位参与Gemini模型开发的其他人。

I think another big piece of news yesterday alongside what we've just talked about are the departures of pretty senior folks at Google DeepMind, including longtime AI scientists, Jeff Dean, several others that worked on the Gemini model.

他们离开是为了去创办一家名为Discovery Loop的新公司，这家公司正试图构建用于科学发现的人工智能。

They were leaving to go start this new company called Discovery Loop that's trying to build AI for scientific discovery.

你觉得谷歌会不会因此出现人才流失的担忧呢？

I mean, do you think there's a worry about kind of brain drain here from Google at all?

我觉得这虽然不是理想情况，但说实话，谷歌是一家非常大的公司，拥有大量人才，而且有才华的AI人才一直在流动。

I mean, I guess it's not ideal, but honestly, Google is such a large company and they've got an enormous amount of talent and talented AI people are moving around all the time.

我认为对此反应过度并不明智。

I don't think it's that smart to overreact to things.

杰夫·迪恩显然是一个非常受尊敬的人，但我们听说他已经不像以前那么重要了。

Jeff Dean is obviously a very highly thought after person, highly thought of person, but what we had heard was that he was not as important as he had been once.

我认为他在作为人们敬仰的对象方面很重要。

I think he's important in terms of being that person that people look up to.

不过，这很难说，我再次表示，我不会对他的离开感到太担心。

Mean, look, it's a bit hard to say, again, I wouldn't be that concerned about his departure.

正如我所说，他们有很多人，我相信他们承受得起失去一些人才。

Mean, as I said, they've got many people and I'm sure they can afford to lose some.

太好了。

Great.

好的。

Okay.

好的，非常感谢你，马丁。

Well, thanks so much, Martin.

再次感谢，刚才那位是马丁·皮尔斯，The Information的联合执行编辑。

And again, that was Martin Peers, co executive editor here at The Information.

说到最近谷歌的离职潮，我们的下一位嘉宾，Rob Toews，是Radical Ventures的合伙人，他实际上投资了他们新创立的公司Discovery Loop。

Speaking of those recent Google departures, our next guest, Rob Toews, a partner at Radical Ventures, is an investor who's actually backing their new startup Discovery Loop.

欢迎来到节目，Rob。

Welcome to the show, Rob.

谢谢，Stephanie。

Thanks, Stephanie.

好的。

Good.

很高兴来到这里。

Great to be here.

那跟我说说，Discovery Loop 到底是做什么的？

So tell me, what exactly is Discovery Loop doing?

显然，现在还在保密阶段，但你能不能稍微透露一点？

Obviously, pretty still in stealth, but is there a sneak peek you can give us at at all?

是的。

Yeah.

我想先强调一下，这个创始团队的阵容真的是前所未有的强大。

So I would start by just highlighting the the really unprecedented caliber of the of the founding team.

四位创始人一起，在过去十五到三十年里，真正打造了谷歌的技术基础设施和人工智能基础。

There's four founders who collectively really built Google's techno technical infrastructure and AI foundation over the past fifteen to thirty years.

杰夫·迪恩将担任CEO，他之前是Alphabet的首席科学家，创立了谷歌大脑，创造了TPU，从一开始就推动了谷歌在AI领域的崛起。

Jeff Dean will be the CEO who is the chief scientist at at Alphabet, founded Google Brain, created the TPU, was has been behind, Google's ascendance in AI from the very beginning.

奥里尔·维尼奥尔斯也是谷歌和DeepMind在AI方面的长期领导者。

Oriel Vignoles is, is also a longtime leader at Google and DeepMind on the AI side.

他最近是Gemini的联合负责人，但多年来也是谷歌许多最重要突破性研究的首席研究员，你知道，可以追溯到像最初的专家混合论文或Chinchilla论文等。

He was the co lead of Gemini most recently, but also, was a lead researcher on a lot of Google's most important breakthrough research over the years, you know, dating back to things like the original mixture of experts paper or the chinchilla paper, etcetera.

Sanjay Gimmelwat，他也是谷歌基础设施和分布式系统领域非常长期的领导者和先驱，然后他也是谷歌最资深、任职时间最长的AI研究员之一。

Sanjay Gimmelwat, who also has been a a very long time leader and pioneer of Google's infrastructure and distributed systems, And then who is, you know, another one of the of the senior most and and longest tenured AI researchers at Google.

如果你在Google Scholar上按引用次数排序，看看世界上AI领域引用最多的研究者，那真是了不起。

It real it is remarkable if you sort on Google Scholar by, like, the most citations of of any researcher in the world in AI.

这个团队，他们加起来，排名大概是第三、第六和第七。

This this collect this team collectively is, like, number three, number six, and number seven.

就像杰夫·辛顿、扬·勒昆，然后是杰夫·迪恩，然后奥里尔和克拉克也在上面。

It's, like, Jeff Hinton, Jan McCune, and then Jeff Dean, and then Oriel and Clark are also on there.

然后如果你单独按分布式系统领域引用最多的人排序，杰夫和桑杰也都是前五名。

And then if you separately sort the most cited folks in terms of, distributed systems, Jeff and Sanjay are also, like both both of them are top five.

所以这真的是，老实说，我认为可能客观上是AI初创公司有史以来最强的创始团队。

So it really is like a honestly, I think possibly objectively the strongest founding team for an AI startup ever.

你可能会有一些争论。

You could have some debates around that.

但这确实是一个独特的机会。

But, so it really is a unique opportunity.

至于他们在做什么，宏观愿景是构建能够递归自我改进的AI系统，因此命名为“发现循环”，就是构建这种递归自我改进的循环。

In terms of what they're doing, the big picture vision is to build recursively self improving AI systems, hence the name Discovery Loop, you know, kind of build building these recursively self improving loops.

他们最初的重点是开发能够改进AI的AI，也就是自动化机器学习研究的过程。

The initial area that they are focusing on is building AI that can develop better AI, so kind of automating the the process of machine learning research.

但更宏大的愿景是，将这个递归自我改进的系统应用于科学和工程领域的许多重大挑战，从气候变化到健康，再到半导体设计等等。

But the bigger picture vision is to point this recursively self improving system at a lot of the biggest challenges in science and engineering from climate change to health to semiconductor design and so forth.

所以，真正的宏观愿景，或者说SAR，是利用前沿AI来应对人类面临的一些最大挑战以及科学中尚未解决的问题。

And so it really is the big picture vision or SAR is about using cutting edge AI to tackle some of the biggest challenges facing humanity and unresolved challenges in the sciences.

不过，这确实很有趣。

I mean, it is interesting though.

我觉得这里有很多初创公司，比如Periodic Labs、Future House，甚至OpenAI、Anthropic，还有谷歌——这些联合创始人显然来自谷歌——都在类似地探索如何构建能够进行更好AI研究的AI。

I feel like there are a number of startups here like Periodic Labs, Future House, but even OpenAI, Anthropic, and actually Google, which obviously these co founders came from, that are similarly tackling this question of how do we build AI that can do better AI research?

所以，我想问的是，为什么要单独去做，而不是留在这些现有的初创公司里呢？

So, know, I guess my question is why kind of go do that separately versus staying within one of these other existing startups?

那你为什么有信心支持这个独立的创业项目，而不是担心OpenAI、Anthropic、谷歌会先做到呢？

And what makes you confident to kind of back this separate venture versus like, you know, are you worried at all that OpenAI, Anthropic, Google are going to reach that first?

是的。

Yeah.

确实有其他团队在研究这些问题。

There certainly are other groups working on this set of problems.

因为坦白说，我认为这是应用AI最重要也最有趣的领域。

And it's because frankly, I think it's like the most important and the most interesting area to which to apply AI.

所以我觉得这是一个巨大的机会，我不认为任何一家公司能独自在AI应用于科学领域胜出。

And so I I think it's, you know, it's a it's a vast opportunity set, and I don't think any one company is going to, you know, win applying AI for sciences.

在科学领域，你可以研究材料科学、生物学、化学。

You know, with within the sciences, you know, you can pursue material science, you can pursue biology, you can pursue chemistry.

在生物学里，有无数不同的治疗领域和方式可以探索。

Within biology, there's, a million different therapeutic areas and modalities you can pursue.

所以我们谈论的是一个非常广阔的领域，涵盖了人类大部分的探索和努力。

And so we're really talking about, like, a a very vast space that encompasses, like, much of the human, you know, endeavor and undertaking.