#61 - 推动AI走向开源模型的经济因素 - Jeff Morgan

所以我想问你的一个问题是，你知道，现在开源模型在美国引起了这么大的关注，而且大多数领先的开源模型，比如你提到的那些，像阿里巴巴的千问、Kimi、DeepSeek，这些基本上都能和前沿模型竞争的开源模型，大多都在中国。

So one of the questions I have, for for you is, you know, with this moment that we have with with open models gaining so much interest in The US, knowing that most of the open models, the leading open models, like the ones you've mentioned, Quinn from from, you know, Alibaba, Kimmy, Deep Seek, all of these, all of the open models that essentially can compete with the frontier models.

而前沿模型主要在美国。

Those are mostly in China and then the frontier models are in The US.

所以美国这边，AI生态圈里的人，还有政府和其他地方的人都在说，看，我们需要美国本土的开源模型。

So The US has sort of say people in The US in the AI ecosystem and beyond in the government and other places are saying, look, we need open US based open models.

你知道，我们需要有公司来做这件事。

You know, we need companies working on it.

而且我得说，几周前我参加AI活动的时候，确实有几家公司过来跟我说，嘿，我们是做开源模型的。

And I will say when I was at AI for event a couple of weeks ago, I did have a couple of companies that are coming up to me and saying like, hey, we're an open model.

你知道，我们正在打造开源模型。

You know, we're building open models that are here.

所以我也看到了一些苗头，但显然那些公司和团队，他们得来找像你们这样的团队。你不需要说出名字，但我想知道，在最近几个月的这股势头之后，你们是不是已经看到新的，可能来自美国和欧洲的开源模型出现了？

So I sort of have seen a little bit of it, but clearly those companies, and those teams working on that stuff, they they have to come to teams like you and you don't have to say the names, but I guess I'm wondering, are you already seeing after this momentum over the last couple months, are you seeing new maybe US and European based open models arising?

而且我觉得这不只是美国和中国之间的事情。

And it's not a I think it's not just a US versus China thing.

更重要的是，你知道，他们会理解——不管是在加拿大、美国还是欧洲，如果人们在开发开放模型，他们会理解当地的需求，并且能够更协作地与他们合作。

It's more that, you know, they're gonna understand the whether it's in Canada, whether it's in The US, whether it's Europe, if people are making open models, they're gonna understand the needs, right, of some of those local geos, and be able to work more collaboratively, you know, with them and that.

你看到这种情况了吗？

Are you seeing that?

你是否已经看到这里对构建开放模型的兴趣激增？

Are you already seeing a sort of a surge in interest of people building open models sort of, over here?

当然。

Absolutely.

我是说，周一Meta的超级智能实验室推出了Glimmer，并且承诺开放Spark的一个版本。是的。

I mean, Monday with Meta, superintelligence labs launching use Glimmer and then committing as well to open a version of use Spark Yeah.

这是向前迈出的一大步。

Is a huge step forward.

我是说，这些都是最先进的模型。

I mean, these are state of the art models.

我觉得这里有两种不同的方法。

And the the way I think about it is there's kind of two two different, you know, approaches here.

对于这些前沿模型来说，它们追求的是超级智能，探索最前沿，这需要大量的资源。

For these frontier models where, again, they're looking for superintelligence, and seeking the frontier, it requires a tremendous amount of resources.

我觉得周一最令人惊叹的是，最大的实验室之一，也就是Meta超级智能实验室，这个资源最充足的实验室之一，承诺开放模型。

And I think what's so amazing about Monday was that one of the biggest labs, which, you know, is Meta Superintelligence Labs, one of the most resourced labs committed to open.

这对这个市场来说是一个巨大的时刻。

And that's a huge moment for this market.

而且，这个团队正是最初LAMA模型的创始团队。

And again again, this is, you know, the the the team that started it all with the original LAMA models.

所以，这是多么不可思议的时刻。

So what an incredible moment.

然后我们还和一些新的实验室见了面，他们正在开发一些令人难以置信的模型，特别是围绕支持这些不同的应用场景。

And then and then we also meet with, you know, some some of the new Neil labs, which are just working on some incredible models, specifically around enabling these different use cases.

有些可能专注于编码。

So some of them may be focused on coding.

有些可能专注于协作工作。

Some of them may be focused on co work.

有些可能专注于模型定制，而且他们资源也很充足。

Some may be focused on model customization, and they're also well resourced.

这就是硅谷特别的地方之一。

And that's one of the special things about Silicon Valley.

这里可用的资源太多了，不管是人才，还是显然的资金。

It's just the amount of resources that are available here, whether it's talent, obviously funding.

所以我们看到这些资源都集中在开放模型上。

And so we're seeing that and that being focused on open.

一年前你问我，美国前五的实验室里，有多少会专注于开放模型？

When you asked me a year ago, were the next Neil Labs in the top five labs in The US focused on open models?

但过去一个月发生了这样的变化。

With this change that's happened in the last month.

对。

Yeah.

现在其中很大一部分都是了。

A good pro a good portion of them are.

而且他们中的很多人都支持詹森的信和萨蒂亚。

And many of them supported, Jensen's letter and Satya.

我觉得我们正处在一个关键的时刻。

And I think we're at a pivotal moment for that.

然后我也确实看到一些初创公司在探索如何构建非常有差异化的小模型，因为小模型不需要像大模型那样多的资源。

And then I do think there are startups that are approaching it on how can they build very differentiated small models, where small models don't require as many, as many resources as the larger models.

因为一般来说，就像你之前提到的，它们可以被蒸馏。

Because generally, to your point earlier, they can be distilled.

我们看到很多初创公司正在蒸馏模型，这些模型可以在手机上运行，或者有些可以在笔记本电脑上运行。

And we see a great set of startups that are are are distilling models around ones that can run on the phone or one that or some that can run on your laptop.

我觉得这对所有层级的模型来说都是一个不可思议的时刻，你知道，现在所有引擎都在全力运转。

And I think that's I think it's an incredible moment for all tiers of models that, you know, all pistons are firing right now.

所以这真的非常令人兴奋。

And so it's incredibly exciting.

但一年前，情况绝对不是这样。

But a year ago, that definitely wasn't the case.

我希望的是，而且我认为这会发生，因为客户们非常重视拥有自己的AI，所以我觉得一年后我们再回头看，会说，哇，那只是开放模型浪潮的开始，整个生态系统已经建立起来，实际上每个模型实验室都在发布开放模型。

What I can hope for, and I think will happen because customers are, are so importantly focused on owning their AI that I think that, like, a year from now, we'll be looking back and saying, wow, that was just the beginning of the open model wave, that there's an entire ecosystem that's built, that effectively every model lab is open is releasing open models.

我觉得我们很容易在一年内实现这个目标。

And I think I think we could easily get there in a year.

所以你认为这个时刻正在变成一场运动，这就是现在的发展方向吗？

So you see the moment becoming a movement, as the direction it's going right now?

我们用Ollama发布模型已经三年了，但我觉得正如你所说，真正看到市场发生关键变化的是最近三周。

It's been three years since we've been launching models with Ollama, but I think it's really only in your point the last three weeks that we've seen a pivotal change in the market.

而且，这实际上始于这个领域最大的参与者。

And, you know, and and it really starts with ultimately the largest, players in the space.

所以当那些拥有顶尖人才和资源的实验室开始行动，再加上Nvidia的硬件支持来推动开放，我觉得你就有了启动一个生态系统所需的一切。

So when when the lab's online with the talent and the resources come online and then you've got the hardware support from Nvidia to go open, I think you've got everything you need to to jump start an ecosystem.

这是过去三年里我们从未见过的。

That's unlike anything we've seen in the last three years.

而且开放模型已经被部署到企业中了。

And open models have already been deployed in enterprises.

你知道吗，财富500强中有85%的公司都在用Llama。

You know, llama's used by 85% of the Fortune 500.

但说到底，我觉得这才刚刚开始。

But, ultimately, I think it's just at the beginning.

你看一下token的使用量，有多少token实际上是通过开放模型处理的？

You know, if you look at the token rate, how many tokens are actually going through open models?

这个数字还相当小。

That's still quite small.

我觉得我们即将看到一个惊人的反转，开放模型将占据token的绝大多数。

And I think we're about to see an incredible inversion where open models are the super majority of tokens.

好的。

Okay.

你刚才说，目前全球20%的模型流量是通过开放模型处理的？

You said 20% of models today globally are going through are going through open models?

20%的token流量，对，是20%的流量。

20% of the token volume is is Oh, 20% volume.

我和团队一起，就是处理token量。

The team and I have, and and just have token volume.

对吧？

Right?

好的。

Okay.

因为你想啊，如果你想在AI之上建立下一个业务，一个AI平台，这在过去三十天之前其实是不可能的，因为模型、因为行业的支持、因为模型本身，还有硬件合作伙伴的支持。

Because if you think about, you know, if you're gonna build the next business on top of AI, an AI platform, that really wasn't available to you until the last thirty days because of the models, because of the backing from the industry, because the model and because of the, the hardware partner support.

所以我觉得这些因素都汇聚在一起了，这会把那20%彻底变成80%到90%。

And so I think that's all come together, and that's gonna completely change that 20% into an 80 to 90%.

你很容易看到这样一个世界，在日常任务中，除了最前沿的研究，开放模型都是最佳选择。

You know, you see a world easily where for everyday tasks, everything but the very frontier research, open models are the best choice.

而且到了那个时候，你会看到一个焦点，世界上大部分token都通过日常的开放模型来处理。

And, you know, at that point, you have a a focal point where most of the tokens in the world are just processed through everyday open models.

这对企业来说非常好。

And that's great for businesses.

对吧？

Right?

因为成本会降低。

Because the costs get lower.

他们在定制化方面有更多控制权。

They have more control over customization.

这会开启全新的创新浪潮。

It unlocks a whole new wave of innovation.

它不会被少数人垄断。

It's not locked up in in, you know, one group of people.

是的。

Yeah.

好的。

Alright.

我想花点时间谈谈开发者，因为显然他们是你们的核心受众。

I wanna talk for a moment about developers, because, obviously, it's a core audience for you all.

过去几个月里，我看到了这股趋势。

And I saw this movement happen, over the last few months.

大概三四个月前，大家开始真正感受到token成本和推理带来的压力。

This is maybe three or four months ago when people really started to feel the pain of all of token costs and inference.

比如，像Uber这样的公司，今年前四个月就把全年的token预算用完了。

And one of the things, you know, we had businesses like Uber saying they ran through their whole, token budget in four months for the first four months of the year.

这其实跟我们之前聊到的智能体有很大关系。

This is really driven by what we talked about earlier by agents.

这些智能体消耗大量token，但同时也干了很多活。

You know, these these agents, they eat up a lot of tokens, but they also do a lot.

它们能做的事情多得多。

They can do a lot more.

到了2026年，AI的能力突然大幅提升，光靠智能体和它们的工具框架就做到了。

All of a sudden AI has gotten enormously more capable in 2026 just because of the agents, the agent harnesses, you know, alone.

这带来的一个后果就是，开发者有时候要花掉非常惊人的钱。

One of the consequences of that is that developers are spending sometimes crazy amounts of money.

有个首席技术官告诉我，他们开发团队里的负责人，

I had a CTO tell me that they're the sort of leads on their develop on their development team.

每个月在令牌上的花费，基本上是他们总薪酬的1.5倍。

We're spending 1.5 X their total compensation in tokens essentially, you know, every month.

所以我们看到，这也是你们参与的一个领域。

And so we've seen this is another one of the areas where where that you all play in.

我们就是这么认识的。

It was how we first met.

我想问问你关于M5的事情，你知道，MacBook Pro M5 Max发布了，突然你就能运行更大的模型了，比如300亿参数的模型，700亿参数的模型。

I wanted to ask you about this M five, you know, MacBook Pro M five Max came out, and all of a sudden you were able to run these larger models, 30,000,000,000 parameter models, 70,000,000,000 parameter models.

对于不熟悉的人来说，以前这通常需要数据中心才能做到，而在性能很强的笔记本电脑上，可能只能运行22亿参数的模型。

For those who aren't familiar, that's typically you would need a data center to do that before you could only run maybe a 2,200,000,000 parameter model on a really powerful laptop.

但现在，你可以在自己的笔记本电脑上运行这些模型了。

Well, now you could run this on your laptop.

所以突然之间，你有了所有这些好处，你有了很多好处。

And so all of a sudden, had all of these benefits of, you had a number of benefits.

你在token成本上省了很多。

You are saving a lot in token costs.

你的token成本可能为零，你知道，可能只是硬件的成本，即使那台笔记本要五千美元，如果它在一个月内帮你省下的token成本超过这个数，那也值了。

Your token costs might be zero, you know, might be whatever the cost of the hardware is, even if that if that laptop cost $5, you know, if you were spending, if it saved you, you know, in that month, know, might have saved you more than that in one month, you know, in token cost.

而且你还有好处，比如速度更快，隐私性更好。

But then you also had the benefit of like, it was faster, It was more private.

你还能在离线的时候运行它。

And you also, you know, had the ability to to run it, you know, when you were offline.

比如，有些推文说，我在两万英尺的高空用笔记本做所有工作，连接着本地模型。

Like, you could be there were these tweets that went live of like, I'm, you know, doing all my work on a laptop at 20,000 feet, you know, and connected to this local model.

所以我很想听听你们的看法，因为今年硬件突然变得非常有趣和强大。

And so I'm interested to hear your all's perspective on that because all of a sudden this year, the hardware has gotten really interesting and good.

你们和硬件供应商合作，包括苹果、AMD、英伟达。

You all work with the hardware providers, including Apple, AMD, Nvidia.

你们看到了这种情况吗？

Are you seeing that?

这只是在一些推文里看到的噱头吗？

Is this sort of just a flashy thing we see in some tweets?

还是说你也看到了真正的势头，开发者和人们想以这种方式运行本地模型？

Or are you also seeing some real momentum around developers and people wanting to run local models in that way?

是的。

Yeah.

工作站硬件市场正在出现一个转折点。

It's a there's an inflection point occurring in the, you know, workstation hardware market.

这始于苹果芯片的笔记本电脑。

And that started with the Apple silicon laptops.

关键的变化是什么？

And what are the key things that have changed?

首先是内存的极大增加。

The first one is that you've got an extreme increase in memory.

你知道，就是电脑自带的内存容量非常大。

You know, just the sheer amount of memory that comes with your computer.

而过去游戏和工作站市场的固定显卡，可能只有8、10、16，或者32GB的显存。

Whereas, you know, a fixed graphics card for the last for the gaming and, workstation market of the past might come with eight, ten, 16, maybe 32 gigabytes of VRAM.

现在花几千美元就能买到一台128GB内存的笔记本电脑。

You can get a laptop for a couple thousand dollars that is a 128 gigabytes.

这是统一内存的第一步，因为你可以把系统和图形内存整合到同一个地方。

And that's the first step of this unified memory because you're able to pull the system and the the, graphics memory into into one place.

所以这是第一个变化。

So that's the first thing that changed.

第二个变化，在M5上非常新，就是这些新的硬件指令。

The second thing, which is really new at the m fives, is these new, hardware instructions.

过去的显卡主要专注于渲染图形，也就是三维图形。

So the the graphics cards of the past were really focused on rendering graphic you know, they focused on three d graphics.

由于像CUDA这样的技术，英伟达多年来一直开发，直到第一个神经网络出现，这确实带来了灵活性，让这些神经网络得以运行，但这并不意味着硬件是为它优化的。

And that that that has the you know, because of technology like CUDA, you know, which Nvidia built for many years leading up to the first neural network, you know, that enabled the flexibility to enable these neural networks, but that doesn't mean the hardware was optimized for it.

现在我们看到M5、英伟达的新DGX Spark和RTX Spark的情况。

And now we're seeing a case with m five and the new DGX Spark from Nvidia and the RTX Spark.

我们看到这种统一内存与专门加速神经网络的硬件指令相结合，也就是苹果的神经加速器。

We're seeing this unified memory be met with these hardware instructions that are specific for accelerating neural networks, the neural accelerators in Apple.

然后，当然，还有英伟达推出的Blackwell和Vera Rubin芯片，它们专门用于加速神经网络。

And then, of course, you've got the Blackwell and and, you know, Vera Rubin chips coming from Nvidia that are purposefully focused on accelerating neural networks.

对。

Yeah.

而且它配备了

And it comes with

AI工作负载。

the AI workloads.

就像是一个单独的处理器，专门用来运行AI工作负载。

Like, it's like a separate processor processor just just to to run run AI AI workloads.

工作负载。

Workloads.

没错。

Exactly.

你知道，就是为了处理注意力机制，这是个非常耗资源的操作。

You know, just to handle, attention, which is a very expensive operation.

然后最后，还有一系列软件发布出来，用来加速，把这些能力变得非常容易使用，比如像Ollama这样的工具，以及其他开发者，可以把它整合到开发流程中。

And, and then lastly, you've got this suite of software that's being released to accelerate, to take that and then make it extremely, accessible for tools like Ollama and other developers to then integrate that into the developer workflow.

对吧？

Right?

所以对苹果来说，有MLX项目，Ollama利用它，在Mac上达到顶尖速度。

And so with Apple, there's the MLX project, which Ollama leverages to be state of the art speeds on Mac.

然后当然，还有英伟达新的CUDA栈和Nemotron栈，属于应用层。

And then, of course, there's the the new CUDA stack and Nemotron stack of the application layer from Nvidia.

所以你把这三样东西放在一起，重点不再是说，嘿，我们看到桌面电脑变得足够好了。

And so you you take these three things and you bring them together, and it's less about like, hey, we're we're seeing desktops get good enough.

而是说，这些原本驱动数据中心的技术，现在来到了你的笔记本电脑上。

It's like, this is the the stuff that's powering data centers that's coming onto your laptop.

而且这非常强大，价格却和过去十年的笔记本电脑差不多，甚至没有贵多少。

And that's extremely powerful at a price point which is not much more expensive than the laptops of the last ten years.

对吧？

Right?

是贵了一点，但也没有贵出一个数量级。

It's a little bit more, but it's it's not it's not order of magnitude more.

这就形成了一股潮流。

That creates a movement.

好。

K.

对吧？

Right?

就是说，你能买到一台很好的。

Like, you can get a really good one.

三千到五千美元。

Three to five k.

而且而且

And and

那就是你把所有钱都花在5千到1万的代币费用上的时候。

that's when you're spending all you're spending that on 5 or 10 k on token costs.

比如，你可以说，太好了。

Like, you can you can be like, great.

我花了那些钱，现在我在代币费用上花了一大笔。

I spend that, and now I'm I'm spending a ton of token cost.

没错。

Exactly.

而且令人印象深刻的是，对于一些模型，比如Gemma 4的260亿参数模型和QEN 35B模型。

And what's so impressive is for some of the models, you take the Gemma four, 26,000,000,000 parameter model and the QEN 35 b model.

这些模型都是为高吞吐量优化的。

And these are models that are optimized for high throughput.

所以在M5 Mac或者RTX Spark或者DGX Spark上，你可以在提示处理中每秒处理数千个代币，输出速度也能达到每秒100到200个代币。

And so on a m five Mac or an RTX Spark or DGX Spark, you're able to get thousands of tokens a seconds processed in prompt processing and over a 100 to 200 tokens a second in output.

所以你能在这台3000到5000美元的机器上获得数据中心级别的速度。

So you're getting data center speeds to all that happening on this 3,000 to $5,000 machine.

而且最棒的是，这是固定成本，而且是私密的。

And so and and what's great about that is that it's a fixed cost and it's it's private.

所以你不用去操心，嘿，算力在哪里？

And so you don't have to go figure out, hey, where's the compute?

它是不是在安全的地方？

Is it in a secure location?

是不是在合适的国家？

Is it in the right country?

它就在你身边。

It's right there with you.

所以我们对这个的看法是，这是个人化的时刻。

And so the the way we think about this is this the personal moment for it.

AI的个人电脑时刻。

Personal computer moment for AI.

这是个人电脑的时刻。

It's a PC moment.

因为所有那些从当年的IBM大型机到个人电脑的变化，你可能还记得你拥有第一台电脑的时候。

Because all the things that went from a mainframe back IBM mainframe back in the day to the PC, you probably remember when you got your first

个人电脑。

PC.

那是一次难以置信的体验。

It was an incredible experience.

对吧？

Right?

而且它并不是最强大的机器，跟银行用的那些大型机比起来肯定差远了，但它是你的，数据是你的，应用程序也是你的，而且你不用担心按量计费的问题。

And it wasn't the most powerful machine compared to what you could you know, what I'm sure what banks were using for mainframes, But it was yours, and it was your data, and it was your applications, and, you know, you didn't have to worry about metered costs.

现在同样的时刻正在发生，它将彻底改变开发者、消费者，但最重要的是，企业。

And that same moment's happening right now, and it's gonna completely transform, developers, consumers as well, but most importantly, you know, businesses.

所以我们其实还处在某种算力危机之中。

So a lot of the we we're in this sort of compute crisis still.

对吧？

Right?

比如说，内存成本太高了。

Like, cost of memory is so high.

这很罕见。

It's rare.

GPU也很难搞到。

GPUs are are hard to get ahold of.

但我看到的一个趋势是，在2026年下半年的趋势中，有一个动向就是本地处理，无论是在台式机、笔记本电脑上，还是甚至在本地。

But one of the things that is one of the things that I see as one of the movements in the second half of the trends for the second half of twenty twenty six is this local processing, both sort of on the sort of desktop, laptop, and, you know, even kind of locally.

所有这些都有可能减轻一些负担，这个数据中心的负担，因为大家都在建数据中心，担心AI发展太快，我们没有足够的电力来支撑。

All of that has the potential to ease up some of the this burden, this data center burden that, you know, everyone's running building data centers because they're worried that, you know, AI is growing so quickly and we won't have the power to, to do it.

而这些开放模型加上这个硬件，有机会帮助缓解一些这种负担。

And these so open models and plus this hardware has a the opportunity to help ease some of that, that burden.

是的。

Yeah.

没错。

Exactly.

而且我觉得，就像你说的，这也不是全部的情况。

And I think it's not all of it, to your point.

但是，对。

But Yeah.

有一类使用场景，在本地运行是非常合理的。

There's a set of use cases where it makes so much sense to run it locally.

就目前而言，我希望这会改变，而且我相信将来会改变，但最难的编程任务，你还是得用非常大的模型。

And as we stand today, and I I hope this changes, and I believe it will in the future, the hardest coding tasks, you still gotta use a really large model.

你看到了那个QMK3模型，那个2.8万亿参数的模型，那个CoinMax模型。

You saw the q m k three models, the 2,800,000,000,000 parameter model, the coin max model.

据说未来的模型会非常大。

Future models are rumored to be very large

对。

Yeah.

参数达到万亿级别，而且是开源模型。

In the trillions of parameters, the open models.

所以这些模型就是为了和那些大型的封闭前沿模型竞争。

And so and those are there to rival these large frontier closed models.

而且它们在那些非常难的编程问题上确实表现出色。

And and they really shine with these really hard coding problems.

不过，对于很多日常工作，比如处理文档，用200亿到1000亿参数的模型就已经完全足够了，这些模型在我们讨论的这种硬件上运行得很好。

However, for a lot of the co work trying to get work done, trying to do document processing, that's already absolutely possible with the 20 to 100,000,000,000 parameter models that run great on this the kind of hardware we're talking about.

所以我觉得，对于这类常规工作负载，绝对会优先在本地运行。

And so I I see that for the run of the mill workloads like this, it will absolutely be local first.

你会进入一个像iCloud那样的世界，有些操作在本地完成，有些则卸载到云端。

And you'll be in a world just like with, you know, iCloud, for example, where some of it will happen on your machine and some of it you'll offload to the cloud.

对。

Yeah.

随着硬件的发展，我们需要把这种切换做得非常无缝。

And we'll just need to get really good at making that seamless, you know, as as the hardware catches up.

所以你可以想象，这时候路由就派上用场了。

And so you can imagine a world where This is where routing comes in.

路由现在是个大话题，不仅从成本角度看，从延迟和客户体验角度也一样。

And routing is a huge topic right now from a cost standpoint, but also from a latency and just customer experience standpoint.

路由有很多潜在的应用场景。

Routing has so many potential applications.

非常好。

Very good.

那你现在也在和其他供应商合作吗？

Now, are you working with other providers too?

我看看，我拿一个出来。

I'm looking, I'm gonna hold one up.

比如这个是AMD的Ryzen AI Halo。

Like this is the AMD Ryzen AI Halo.

当AMD把这个交到我手里，说‘来，试试我们的版本’的时候。

And when AMD sort of put this in my hands and said, okay, give try our version of this.

这基本上就是他们版本的，你之前提到的那个Nvidia的盒子。

This is essentially their version of the one you mentioned earlier, the Nvidia's Nvidia's box.

你提到了英伟达和苹果，它们是这个行业的两大巨头，但还有其他公司在做类似的事情。

You mentioned Nvidia and Apple, they're two heavyweights in this industry, but there are others doing.

你们也在和其他硬件厂商合作吗？

Are you you working with other hardware?

未来还会有更多吗？

Is there more coming?

你知道，你不必说出名字，但你是否能看到更多硬件能够做到我们刚才谈到的英伟达和苹果那样的功能？

You know, you don't have to say the names, but do you have a line of sight into more hardware being able to do the same kind of things like we talked about with Nvidia and Apple?

当然。

Absolutely.

你知道，我刚才提到的这两家，可以说是按市值排的。

You know, I just mentioned the two, you know, maybe call it sorted by market cap.

但我们看到每个主要硬件厂商都在经历这种转型，因为应用正在转向这些神经网络、这些模型，而这种统一架构正成为支持这些模型的最强大方式。

But we're seeing this transformation in every major hardware vendor because the applications are shifting to these neural networks, these these models, and this unified architecture is becoming the most powerful way to to support these models.

我们看到，AMD和英特尔都是我们的合作伙伴，我们非常兴奋能与它们合作。

And we're seeing you know, both AMD and Intel are two example partners of ours that we're so excited to collaborate on.

因为你说到成本要三千到五千美元，那如果它没那么贵呢？

Because to your point about it costing 3,000 to $5,000 well, what if it wasn't even that expensive?

而且我觉得，当硬件厂商们合作解决这个问题时，不管是入门级笔记本还是高性能顶级机器，都会有各种选择。

And I think when you have a collaboration of hardware vendors come together to solve this problem, whether it's an introductory laptop or a really powerful high end machine, there'll be there'll be options across the board.

所以我们真的非常兴奋，从AMD、英特尔，到英伟达和苹果，我们都在合作，确保当有人买了这些机器，打开就能用，直接跑模型，马上开始干活。

And so we're we're just really excited across the board from, you know, both AMD, Intel, as well as Nvidia and Apple, and partnering with them and making sure that, you know, when somebody buys one of these machines, they can open it up, get up and running with a model, and start doing work with it.

但这两个只是更多合作伙伴中的例子。

But those are two examples of more partners.

我们和高通也是合作伙伴，对他们做的事情也感到非常兴奋。

We're also partners with Qualcomm and extremely excited about what they're doing as well.

好的。

Okay.

这些差不多就是该领域所有主要的芯片和硬件制造商了。

Those are pretty much all the big chip and, you know, hardware makers in the in the space.

对吧？

You know?

然后还有一些笔记本电脑制造商。

And then there's some, you know, laptop, you know, makers as well.

所以，这真的是一个非常有趣的发展。

So, that's really that's a really interesting development.

我们看到AI生态系统正在形成。

We're seeing the AI ecosystem sort of it.

大概一年前，可能只有几个名字主导这个领域，但就像我们今天聊到的，你可以看到，会有越来越多的参与者。

There was a lot, a few names, you know, that dominated, maybe a year ago and and kind of to your point, as we've been going through the conversation today, you can see that a lot, there's gonna be a lot more participation.

而且从某种意义上说，这是必要的，因为会有更多不同的用例，也会有更多的自由和选择，让人们去定制和做自己的事情。

And in a sense, you're gonna need that because there are gonna be more distinct use cases, they're gonna be more, there's gonna be some more freedom and options for people to customize and do some of their their own things.

关于这个生态系统，我最后想问的一个问题是，Jeff，我最近遇到了汤森路透的首席技术官，他说他们构建了自己的模型。

One of the one of the last sort of questions I I had for for now, Jeff, for, you know, all of this ecosystem stuff is I recently ran into, the CTO of Thomson Reuters, and, and he was saying that they built their own model.

他们认为，智能将成为他们现在所做事情中最重要的部分。

And they had this view that, you know, that intelligence was gonna become the most important part of what they do now.

我们想到汤森路透。

We think of Thompson Reuters.

路透社其实是个新闻机构。

Reuters is like a a news organization.

他们还有更大的业务。

They have a bigger business.

他们做的其中一件事就是法律。

One of the things they do is like law.

他们有很多自己的法律服务。

They have a lot of, like, their own kind of law services that they offer.

他们的结论是，如果我们真的要把这些模型作为我们工作的核心，我们就必须拥有智能。

And their conclusion was, you know, if we're going to if we're gonna really make these models a key part of what we do, we have to own the intelligence.

我们不能把它外包给别人。

Like, we can't outsource that to somebody else.

这让我开阔了眼界，觉得这确实是合乎逻辑的结论。

And and that sort of broadened my perspective to be like, you know, that is sort of the logical conclusion.

你会看到越来越多的公司得出这样的结论：好吧，我们正在把重要的东西放在这上面。

You can see a lot more, a lot more companies get coming to the conclusion, okay, this is doing import we're we're putting important stuff on this.

而且我们也得特别小心，到底要把这些内容给谁。

Like, we also we wanna be really careful about who we wanna send that to.

对吧？

Right?

这可能会成为我们最重要、最有价值的资产。

That becomes our maybe most important, most valuable IP.

而且因为你跟80%的财富500强公司合作，那这怎么说呢？自己做模型可能只是比拿一个开源模型然后针对我们的需求定制更进一步而已。

And so because you work with 80% of the Fortune 500, Where does that I mean, making our own model is only maybe a step further from like taking an open model and customizing it for for what we do.

但我就是想听听你的看法。

But I just wanted to get your perspective on that.

你是不是看到越来越多的公司基本上都得出了这个结论，就是我们需要拥有自己的智能层。

Are you seeing more and more companies basically coming to that conclusion like, oh, we need to own the intelligence layer.

我们不能只是把它外包给供应商。

We can't just outsource that to a provider.

不管怎么说，把我们业务中最核心的部分交给别人做，那也太傻了。

That would be that would be silly to take our most important part of our business and do that no matter what it was.

对。

Yeah.

我看到这些大型企业正在经历三个阶段，他们非常想专注于拥有和定制自己的AI。

I'm seeing three phases from these very large businesses who are very much wanna focus on owning owning their AI and customizing it.

第一阶段是，你知道的，就是如何切换到开放模型。

The first is you know, the the first phase is how do you do you switch how to open models?

基本上就是切换到一个你可以控制的架构。

So it's basically switched to a stack where you control it.

然后第二步是如何定制它。

And then step two is then how do you customize it?

然后最令人兴奋的是第三步，这周早些时候我和一家大型金融公司讨论过，就是如何与你的合作伙伴一起实现这一点。

And then what's so exciting is the step three that I was whiteboarding earlier this week with a major, financial firm, which is then how do you give it how do you do that with your partners?

以及如何以一种有利于生态系统和你的合作伙伴生态系统的方式来分享它。

And how do you share that in a way that benefits the ecosystem and your partner ecosystem?

而这一切最终都会回到第一点上。

And this will all come back kind of to the first point.

但是，你知道，第一步就是确保你构建在最终能自己运行的智能之上。

But, you know, step one is just making sure that you've you're you're building on intelligence that you can ultimately run it yourself.

你可以掌控它。

You can take control of it.

你知道，激励机制对你和那些智能提供者是一致的，都是为了推动你的业务前进。

You know, the incentives are aligned for you and, and the providers of that intelligence to, you know, move your business forward.

如果你回顾一下最大的云服务提供商，比如AWS、谷歌，它们都是建立在开源之上的。

And if you look back to, you know, the largest cloud providers, AWS, Google, they're built on open source.

这为这些企业提供了极大的自主权。

And that provides an incredible amount of agency to these businesses.

所以这是第一步。

So that's step one.

第二步就是如何定制那个循环，我们之前和萨蒂亚讨论过，对吧，就是萨蒂亚发的那条推文。

Step two then is how do you customize that loop we stop spoke about with Satya, right, with Satya's tweet?

那涉及到如何利用你的数据。

And that involves then how do you take your data?

比如，路透社有海量数据，对吧，关于所有新闻稿，还有那个部门的所有法律数据。

Like, Reuters has so much data, right, around all the press releases, all this legal data for that part of the division.

然后，你怎么去定制模型和工具，真正确保你能以前所未有的水平解决客户的问题。

And then how do you customize, the models, the harnesses, to really make sure that you're able to solve your customer's problem in a, in a level you're not before.

但最让我兴奋的是第三步，回馈。那家大型金融公司当时在说，如果我们自己把模型开源呢？

But this third step of giving back is what excites me the most, which is this major financial firm was talking about, well, what if we, what if we open sourced the model ourselves?

因为我们可以构建自己的AI，交付给客户，但那样的话，我们的客户就会陷入我们之前面对封闭模型供应商时的同样处境。

Because we could build our own AI, deliver that to customers, but then our customers are now in the same shoes as us with the previous closed model providers.

所以这个见解很有意思，我很喜欢，因为它意味着，与其把智能攥在自己手里，藏着掖着，不提供给客户，不如想想怎么真正回馈。

And so it's this interesting I I love that insight because what it meant was instead of taking the intelligence and keeping it to yourself and then, you know, hide you know, and then not provide it to your customers, how do you actually give back?

这个想法就是，如果你能构建自己的模型，然后把它开源，整个循环就能继续下去。

And this idea that, well, if you take it if you're able to build your own model and then actually open source that, the whole the whole cycle continues.

现在你的客户可以说，嘿。

And now your customers can say, hey.

比如，我可以和这家金融公司合作，用他们的模型，而且我相信他们会支持我。

I can work with this financial firm, for example, use their models, and I can trust that they they've got my back.

然后那家金融公司就有很好的方式来服务这些客户，围绕那个模型提供大量的基础设施和增值服务。

And then that financial firm has a great way to service these customers with a ton of infrastructure and value added services around that model.

这真的是开放模型的特别之处，你总是在为你的客户着想，因为每个人都是别人的客户。

And this is really what's special about open models is you're always paying it forward to your customers because everyone's a customer of somebody.

如果你能在整个链条上都有开放模型，从基础模型一直到垂直定制的模型，再到最终面向消费者的终端提供商，那会怎么样。

And if and if you can have open models all the way down the stack from the front from the foundation models all the way to the vertically customized models, all the way to the end, you know, provider, end retail, you know, end end provider to the consumer.

然后猜猜谁受益最大？

And then guess who benefits the most?

消费者。

The consumer.

因为他们在整个链条上都有激励，最终消费者真正拥有自己的智能。

Because they've got incentives along all the way down the stack, and then the consumer ultimately owns their intelligence.

那个想法对我来说太有力量了。

And that I that thinking was so powerful to me.

我觉得这是每个企业都会接受的事情。

And I thought that was something that, you know, every business will embrace.

对。

Yeah.

非常好。

Very good.

我很喜欢这种利他主义的观点，关于生态系统的未来走向。

Love that love that altruistic note, in terms of where where the ecosystem is going.

而且总的来说，谈到开放模型，确实有更多的赋权。

And and the the conversation in general, like, thinking about open models, there's a lot more empowerment there.

自上而下的东西少了，自下而上的多了，这些都是好事。

There's a lot less top down, a lot more bottom up, and all of that, you know, are are good things.

好的。

Okay.

杰夫，我想用我问每个人的两个问题来结束这次播客。第一个问题是，你现在使用的AI工具中，有没有哪个让你感到惊喜，但可能知道的人不多，你希望大家去了解一下的？

Jeff, I'm gonna let's let us, sort of wrap with the two questions I ask everybody, you know, at the end of the podcast, which is the first one is what's the the AI tool that you're using, you know, today that's surprising you that maybe not as many people know about that you'd love for for people to take a look at?

我真的很兴奋。

I'm really excited.

你知道，我整天泡在开发者圈子里。

You know, I'm I'm neck deep in the developer world.

我对这个新的软件层感到非常兴奋，它

I'm just excited by this new layer of software that's

正在

being

而且，我觉得有两个工具我们花了很多时间在Ollama里面。

And, I'd say there's kind of two that we're spending a lot of time with within, Ollama.

一个是苹果的MLX软件，很多人不知道它来自最初的PyTorch团队，而PyTorch是目前全球运行推理的软件。

One is the MLX software from Apple, which many people don't know is from the founding PyTorch team, which is kind of the software that runs all the inference in the world right now.

他们去了苹果，开始了这个叫MLX的新项目。

And they went to Apple and started this new project called MLX.

最棒的是，很多人不知道的是，它表面上看起来是另一种用GPU运行模型的方式，实际上它是一个灵活的工具，可以用来训练、推理，甚至构建任何你想要的东西。

And what's so awesome and a lot of people don't know is it's it's what comes across as another way to use GPUs to run models is actually this flexible tool for you to do training, for you to do, inference, to to kinda build anything you want.

我觉得这非常强大，因为它在普通的消费级笔记本电脑上就能运行。

And and I think that's so powerful because it can be done on a consumer laptop.

所以这让我超级兴奋。

And so that that I'm I'm super excited by that.

只要有空，我就去读相关资料，解解馋。

And with any spare time I have, I'm just reading about it, scratching the itch.

而且我觉得，在英伟达这边，DGX Spark 也让我们看到了同样的趋势。

And and I think very much on the, you know, with with Nvidia, we're seeing the same thing with the DGX Spark.

如果要排第二的话，那就是 DGX Spark。

And I'd say if there's a number two, it's the DGX Spark.

我们很幸运，是最早拿到它的人之一。

We're so lucky to be one of the first people to to receive one.

你知道，它最初可能是一款面向研究人员和开发者的产品，但现在已经有点火了，因为它太强大了，几乎可以运行任何类型的人工智能工作负载。

And, you know, what started as kind of a a product maybe focused on researchers and developers, I think, has become a little bit of a sensation because it's so powerful, and you can run almost any kind of AI workload on it.

我觉得无论是搭载 MLX 的苹果芯片笔记本，还是配备新 CUDA 栈的 DGX Spark，它们都是强大的工具。

And I think both whether it's an Apple silicon laptop with MLX or it's a DGX Spark with the new CUDA stack for it, I think they're just powerful tools.

这就是给开发者的答案。

That's the answer for developers.

好的。

Okay.

我想说，让我兴奋的是我的Gmail搜索框终于有了AI功能。

I think I'll say the one that I'm excited about is my Gmail search box finally has AI.

我对此非常期待。

I'm so excited for that.

这和开源模型关系不大，但我很高兴现在处理邮件能高效多了。

It doesn't have much to do with open models, but, I'm just glad I I can get through my email much more effectively now.

很好。

Very good.

这些都是AI工具。

Those are AI tools.

好的，谢谢。

Well, thank you.

非常感谢。

Appreciate it.

我们总是尽量去了解，你知道的，行业里大家都在用的那些AI工具。

We we always try to get the, you know, tools that people in the industry are are using the AI tools.

所以这两个都是很棒的选择。

So those are two great ones.

对。

Yeah.

从开发者的角度，一直到普通人都能用的东西，都有涵盖。

From from the sort of developer standpoint, all the way down to the stuff that everybody, you know, can use.

我也是，嗯，对。

I also, yeah.

非常高兴现在甚至能直接搜索并找到Gmail里的邮件。

Very glad to have the ability even just to search and find, you know, emails in Gmail.

这是一个很大的升级。

It's a big, it's a big upgrade.

好的。

Okay.

最后一个问题，我问每个人的问题是，你知道，AI一直有个承诺，就是当AI发展到一定程度时，它会帮我们处理很多琐碎的任务，我们就会有更多时间去做生活中的其他事情。

The last one, last question that I, that I ask everyone is, you know, there was always this promise in, in AI that AI, when AI sort of advanced, that it would do a lot of the mundane tasks for us, and we would have more time to sort of do other things in life.

但在过去几年里，我发现情况恰恰相反。

And I found over the last couple of years, more of the opposite issue.

我认识的大多数领导者，他们现在的时间更少了，而不是更多，也许是因为他们看到了机会，并且对能做得更多感到非常兴奋，AI能帮他们做很多事情。

Most of the people, the leaders I know, they have less time and not more, at the moment, maybe because they see the opportunity and are very excited by the opportunity that they can do so much more, and AI can do a lot of the things for them.

但正因为如此，每个人都越来越在意时间。

But because of that, everyone is more and more conscious about time.

作为领导者，他们经常说，如何让我的时间发挥最大杠杆效应？

And as leaders often, phrase that they talk about, how do I get maximum leverage for my time?

所以我总是喜欢问领导者，你最好的建议是什么？

And so I love to always ask leaders, you know, what's your best tip?

你现在在做什么来让你的时间发挥最大杠杆效应？

What are you doing right now that's helping you get maximum leverage for your time?

你会给其他领导者分享什么建议？

You know, what's the tip that you would share, to for other leaders?

我觉得每个人，你知道，我们刚才聊的是业务层面的循环，但我觉得每个人每天都有自己的循环。

I think everyone, you know, we talked about a loop at a business level, but I think everyone individually has their own loop that they go through in the day.

我鼓励自己和团队做的事情就是，想办法让这个循环尽可能自动化。

And what I'm encouraging myself and my team to do is to say, how do you how do you make that loop as automatic as possible?

当然，要用AI来实现。

And obviously use AI for that.

这确实不容易，因为你得找到合适的连接器、模型和运行环境，还得找到合适的工具。

And and, you know and it's hard because you gotta get the right connectors and model and runtime, and you gotta find the right harness.

但我觉得，我最大的建议就是，如果你能把那些日常中不特别的部分自动化掉，那会很有帮助。毕竟，经营一家初创公司真的很难。

But I think ultimately, the biggest tip I have is if you can automate that away, if you can take these parts of your your your day to day life that are not special, you know, running a startup is really hard.

我经常问自己，有哪些事情可以自动化，这样我就能有更多时间和团队面对面交流，也能专注于个人生活，多陪陪我的妻子。

I always ask myself, what are things I can do to automate so I can spend more face time with my team so I can focus on, you know, my personal life and spend more time with with my wife?

我觉得这就是关键，鼓励每个人建立自己的循环，然后真正去优化它，把它搞清楚。

I think that's where, encouraging everyone to build their loop and actually optimize it and figure that out.

我最好的例子其实是Ollama的一位主要投资人。

The best example I have for this actually is, one of Ollama's lead investors.

他叫托马斯·通格斯。

His name is Tomas Tungus.

他虽然是AI和数据领域知名的投资人，但他用了一个叫Pi的编程代理，把生活里很多事情都自动化了。

He while he's an you know, a a well known investor in AI and data, he's automated so much of his life with a coding agent named Pi.

从研究到邮件，再到文档和演示文稿，他全都用这个搞定。

And everything from research to emails to, documents to presentations, he does it all.

他整天就待在终端里。

He lives in his terminal.

我觉得这就是未来。

And I think that's that's the future.

而且我觉得，可能不一定是在终端里，也许是在Codex、Claude或者某个开源替代品里。

And I think, you know, it may not be in a terminal, maybe in the codex or Claude or an open source alternative.

我觉得那个世界会特别令人兴奋，以至于你甚至都不需要再参与其中了。

I think that world is is gonna be so exciting to the point where, like, you just don't even have to be in the loop anymore.

你从早到晚，都有一个像好朋友一样的同事，帮你把生活中大部分无聊的事情都做完了。

You can start your day and end your day, you have a whole best friend coworker that's done most of the boring work of your life.

所以我的建议就是努力实现那种闭环，因为我觉得不只是企业在这么做。

So that's my tip is to strive for that loop because I think it's not just businesses doing it.

我觉得是从企业里的个人开始的。

I think it starts with the individuals at the business.

非常好。

Very good.

好的，杰夫，感谢你抽出时间。

Well, Jeff, thank you for the time.

聊得很愉快。

Great conversation.

我相信我们还会再聊的，因为这个领域变化太快了，能有机会聊这些真的很棒，特别感谢你对我们讨论的所有话题的见解，尤其是关于开源模型未来走向的部分。

I'm sure we're gonna talk again because there's so much happening in this space, but what a what a moment to be able to do this and really appreciate your insights on on all the things that we talked about, especially sort of where open models go from here.

这是一次很棒的讨论。

It's a great discussion.

非常感谢你邀请我。

Thank you so much for having me.

好的。

Alright.