"Kimi K2思维"是中国终结美国AI霸权的计划

V世界最新最强大的人工智能模型现在完全免费供您下载并在家中运行。

V World's latest and greatest AI model is a 100% free for you to download and run at home right now.

Kimi K2思维是来自中国前沿AI实验室Moonshot AI的最新推理模型，它在几乎所有基准测试中都超越了OpenAI的GPT-5、Anthropic的Claude和谷歌的Gemini。

Kimi K2 Thinking is the latest reasoning model from Moonshot AI Labs, which is a Chinese frontier AI lab, and it beats OpenAI's GPT five, Anthropix Claude, and Google's Gemini across pretty much all benchmarks.

但这甚至还不是最令人震惊的部分。

But that's not even the most shocking part.

最令人震惊的是它的训练和构建成本仅需460万美元，这仅仅是OpenAI训练GPT所花费数十亿美元的一个零头。

The most shocking part is that it only costs $4,600,000 to train and build, which is only a fraction of the billions of dollars spent by OpenAI to train GPT in the first place.

它还是100%开源的，这意味着您现在就可以坐在家里下载并运行这个前沿AI。

It's also a 100% open source, which means that you can download and run Frontier AI right at home where you're sitting right now.

但当然，这引出了两个非常重要的问题。

But, of course, it begs two very important questions.

第一，开源AI是制胜策略吗？

Number one, is open source AI the winning strategy?

我们一直被引导认为闭源通常是经营企业更好的策略，但中国及其AI模型正在证明我们错了。

We've been led to believe that closed source is typically the better strategy when you run a business, but China and their AI models are proving us wrong here.

第二个问题，更不祥的问题是：美国股市泡沫最终会破裂吗？

And the second question, the more ominous question is, will The US stock market bubble finally pop?

乔希，我们这里有什么发现？

Josh, what have we got here?

这个新模型是什么？为什么我看到的每个社交媒体都在讨论它？

What is this new model, and why is it taking over social media everywhere I look?

他们又做到了。

They did it again.

中国人又做到了。

The Chinese did it again.

他们表现得非常出色。

They knocked it out of the park.

全垒打，本垒打。

Grand slam, home run.

这是个令人难以置信的惊艳模型，而且每次都是这样。

It's an unbelievably impressive model, and this happens every time.

我们拿到了美国这款惊艳的旗舰机型。

We get this amazing flagship model out of The US.

几个月后，我们就能以十分之一的成本获得性能略优的同类产品。

A couple months later, we get the same thing, marginally better, at one tenth of the cost.

成本比当今美国顶尖AI实验室的投入低了整整几个数量级。

Like, full orders of magnitude cost less than what it cost for the leading AI labs in The US today.

参数规格确实令人印象深刻。

The specs are really impressive.

我们会深入探讨所有细节。

We're gonna get into everything.

我们先从高层级的参数表开始讲起。

We'll start with the, I guess, just like the high level spec sheet.

在'人类终极考试'这项基准测试中，它以44.9%的得分刷新了历史记录。

State of the art on Humanity's last exam, which is the reference point that we kind of use in terms of benchmarks, it scored the highest anyone's ever scored, 44.9%.

它实现了一系列突破性进展，但最突出的强项——正如文中所述——在于推理能力、代理搜索和代码生成。

It has a bunch of these really cool breakthroughs, but the big thing that it excels at, like it says in the post here, reasoning, agentic search, and coding.

现在我们可以聊聊几个很酷的点。

Now there's a few cool things that we could talk about here.

EGES，或许我们直接看图表吧，因为我觉得这能直观展示这个模型比其他所有模型优秀多少。

EGES, maybe we'll just get into the charts because I feel like that's an easy way to visualize how much better this model really is than all of the others.

图表显示，虽然GPT-5曾是王者，但Kimi K2现在才是新晋冠军——这体现在思维推理能力上。更惊人的是，第一，这个模型完全开源，你可以免费下载并在本地运行。

And what we're seeing on the chart is that while GPT five was the best, Kimi k two is now the new best, and this is as it relates to thinking and reasoning, and this again, this is so impressive because one, this model is fully open source, you can go download the model and run it yourself locally for free.

你第一反应是什么？

What were your first thoughts when you saw this?

我当时简直惊呆了。

Because to me, I was like, oh my god.

我心想：还有什么理由用别的产品？

I was like, why would I use anything else?

说实话，Josh，我第一反应是去看股市。

My first thought, if I'm being honest, Josh, was, like, to look at the stock market.

我当时想：这会不会让整个美股崩盘？

I was like, is this gonna crash the entire US stock market?

就像当初深度求索发布r1思维模型时，你还记得吗，去年年底那会儿，人们对AI模型训练方式的整个幻想和认知被彻底颠覆了。

Like, when DeepSeek initially released the r one thinking model, do you remember, it was at the end of last year, people's kind of entire bubble and vision of how AI models were trained was completely burst.

自那以后，中国不断推出突破性的前沿模型，其中之一就是月之暗面AI实验室团队开发的Kimi k2。

And since then, China has repeatedly delivered on breaking edge models, one of which is, the Moonshot AI Lab team, which built Kimi k two.

对我来说，这个模型令人印象深刻有几个原因。

It's such an impressive model for a few different reasons, for me.

首先，它现在能与所有顶尖模型竞争。

Number one, it can now compete with all the best.

就我个人而言，我几乎每天都在使用GPT5，无论是处理日常提示请求，还是进行深度思考、研究以及我工作相关的某些领域。

And, personally, GPT five is something that I use pretty much every day, whether it's for, like, kind of casual prompts and requests or whether it's kind of, like, for deeper thinking and research and some of the lines of work that I do.

所以它已经成为我不可或缺的工具。

So it's become kind of, like, quintessential for me.

现在有了这个可以下载并在家用电脑上私密运行的独立模型——就像我这条推文展示的，每百万token输入成本60美分，输出2.5美元——这简直是疯狂的成本削减方案。

Now to have a a separate model that I can download and run privately on my own computer at home that I'm I'm showing on this tweet here that costs 60¢ per million token input and $2.5 output is just an insane cost cutting average way.

如果我要用AI模型运营业务，除了可能需要的维护和设置等工作外，我几乎没有理由不转向这样的方案。

If I was running a business using an AI model, I would there's like very little reason for me not to switch over to something like this aside from maybe like maintenance and setup and and stuff like that.

对我来说另一个令人印象深刻的地方是团队本身，Josh。

The other really impressive thing for me, Josh, was the team itself.

要知道这只是个成立两年的初创公司，这让我想起另一家两年历史的公司——埃隆·马斯克的xAI。

Like, this is only a two year old startup, which reminds me of another two year old startup, which is Elon Musk's x AI.

对吧？

Right?

这两个模型之间有个有趣的关联，Josh——Kimi K2的推理能力，这个思维模型之所以能做到这点，是因为它采用了这种精巧的思维链实验，通过多步骤推导得出逻辑答案，而不是直接抛给你一个结果。

And there's a funny link between these two models, Josh, which is, Kimi K two's reasoning, this thinking model, can do so because it does this, like, really neat little chain of thought experiment where it takes many steps to kind of think to a logical answer versus just kind of like splurging an answer for you.

这正是Grok heavy four在发布新产品时首创的做法。

That's something that Grok heavy four did when they that they pioneered when they launched their new product.

所以Kimi K2某种程度上借鉴了xAI的这些经验来打造类似模型。

So Kimi K two is kind of like drawn on some of these learnings from x AI to to produce a similar model.

另一个很酷的功能是它在思考时能进行工具调用。

The other really cool thing is it does this thing called tool use or tool calling whilst it's thinking.

想象一下，当我尝试思考复杂问题时，会调用不同工具来协助得出答案。

So if you imagine, as I'm kind of like trying to think through a complex problem, I will leverage different tools to be able to help me get to the answer.

所以如果我在参加数学考试，可以使用计算器。

So if I'm doing a maths exam, I can use a calculator.

或者如果我在研究一个深度问题，可能会使用谷歌。

Or if I'm doing a deep research question, I might use Google.

这个AI模型天然具备这种能力，在思考过程中能调用200到300种不同的工具。

This AI model naturally does that and has access to over two to 300 different tool calls and tool uses whilst it does its thinking.

总之，这是一个令人印象深刻的、全新面貌的AI模型。

So just overall, a very impressively new looking AI model.

是啊。

Yeah.

伊托吉，你提到成本是每百万token 60美分。

Itoj, you you mentioned the the cost being 60¢ per million tokens.

没错。

Yes.

我想补充说明这个价格实际上有多低廉。

And I just want to add a little bit of context as to how low that actually is.

我正在查看GPT-5 Pro的输入成本，目前是每百万token 15美元

I was looking at the the GPT five pro cost for per inputs, and it is $15 per

百万15？

million 15?

目前GPT-5 Pro的成本是15美元

15 for the GPT five pro cost currently.

输出部分是每百万token 120美元

The output is $120 per million tokens.

当然这是最顶级的版本

Granted, this is the top of the top.

如果使用GPT-5标准版，输入是每百万token 1.25美元，输出是10美元

If you're using GPT five standard, input is a dollar and 25¢ per million tokens, output is $10.

无论如何计算，成本至少降低了两倍，最高端甚至能达到100倍降幅——前提是它能与GPT-5 Pro竞争，而所有基准测试都表明它完全有这个能力

So any way you you scrape it, it's at least a two x multi cost reduction up to like a 100 x on the highest end, assuming it can compete with GPT five Pro, which all those benchmarks suggest it very well can.

成本运作机制是这样的，屏幕上有个非常复杂的图表

So the cost the way this works is, like, there's this very complicated diagram on the screen.

我不打算解释那到底是什么，但有个有趣的比喻——今早我跟朋友描述时说Kimi K2就像一所巨型学校，里面有所谓的‘专家小组’。

I'm not gonna try to even explain, what that is, but there's this this fun way that I like to describe it when when I was describing it to my friend earlier this morning, which is that, like, Kimi k two, it's like this giant school, and it has these things called specialists.

实际上Kimi K2拥有384个专家小组。

And in fact, Kimmy K2 has three eighty four specialists.

你可以把这些小组想象成数学社、历史社、编程社、辩论社之类的社团。

You could think of these specialists as like a math club or a history club, coding club, debate, whatever it is.

当你提问时，它不会召集全校师生。

And when you ask it a question, it doesn't invite the whole school.

也不会邀请所有社团参与。

It doesn't invite all the clubs.

就像如果你问数学题，它只会咨询数学社——系统会从384个社团中选出8个，整合专家意见来决定解题方案。

It's just eJess, if you ask for a math question, it will query the math club, and it chooses eight out of those three eighty four clubs to help combine their answers, pick the experts, and decide how it's going to solve this problem.

所以它虽然拥有万亿级参数，但每次只调用320亿个参数。

So it has a trillion parameters, but it only uses 32,000,000,000 of them at once.

这就是通过‘专家混合’技术实现成本大幅降低的奥秘。

And that's how we're able to get the huge cost reduction because it uses this thing called mixture of experts.

很多人称之为MOE（专家混合），但本质上它并不是用整个模型的智能来回答‘今天早餐该吃什么’这样的问题。

A lot of people describe it as MOE, but basically what it is is instead of using the entire model's intelligence to answer, what should I have for breakfast this morning?

它会调用厨师俱乐部和健康俱乐部，将两者的建议结合起来，最终给出的答案效果与使用整个模型相当，但在成本、能耗和生成token数量方面都更加高效，整体上便宜得多。

It will take the chef club, it will take the health club, it will combine those together, and it will form an answer that should hopefully give you just as good as a result if you took the entire model, but it's much more efficient in terms of cost, in terms of energy, and in terms of the amount of tokens they could generate because it's so much cheaper across the board.

我认为这是中国涌现出的最令人振奋的创新之一。

And I think that's one of the big, really exciting things that has been cool to see coming out of China.

我们在DeepSeek和Kimi上都看到了这种专家混合体架构，它们真正实现了模型的模块化，只针对特定查询调用相关部分。

We saw it with DeepSeek, we see it with Kimi, And it's this mixture of agents architecture where they're really kind of modularizing the entire model and only using the stuff that's important for a specific query.

他们当时处境非常受限，无法获得最新的GPU或英伟达显卡。

They were put in a very constrained position, which is they didn't have access to the latest GPUs or NVIDIA GPUs.

美国对中国实验室实施了一系列关税限制，导致他们难以获取这类硬件。

There's been a bunch of US tariff restrictions on Chinese labs getting access to these kinds of things.

因此他们必须严格在现有条件和资源范围内开展工作。

So they've really needed to kind of, like, work within their bounds and means.

所以研发出专家混合体这样的架构——或者说他们实际采用的方案——就变得极其重要。

And so coming up with an architecture, like, mixture of experts or the one that they did is super important.

这让我想到了一个梗，乔希，就是：我们到底在干什么？

And it brings me to this meme, Josh, which is, what are we doing here?

美国制造的AI模型与中国制造的AI模型之间存在明显的错配。

There is an obvious mismatch between American made AI models and the, Chinese ones.

OpenAI预计在未来五年内将花费1.4万亿美元。

You've got OpenAI, which is now projected to spend $1,400,000,000,000 over the next five years.

那可是以万亿为单位，而Kimi的训练成本仅为460万美元。

That's trillion with a t versus Kimi training for $4,600,000.

我知道这里面有点标题党的成分。

Now I know there's a bit of, like, click baatiness here.

那460万美元是针对一次训练运行的，通常需要多次训练。

That $4,600,000 was relative to one training run and usually takes a few training runs.

但假设进行了20次训练，每次460万美元，那总共也才不到1亿美元。

But let's say it took, like, 20 training training runs, right, at $4,600,000, that's still only, like, a like, a 100 mil, right, or less than that.

当你考虑到传闻中GPT-5的训练成本高达17至24亿美元时，这点钱根本不算什么。

So it doesn't really matter when you put it into the context that GPT five is rumored to have cost 1.7 to $2,400,000,000 for OpenAI to train.

所以这种差距让我很困惑，Josh。

So there's a mismatch that I don't quite understand, Josh.

这让我对美国公司和前沿实验室的所作所为感到最为不安。

And that's what makes me the most nervous when it comes to, what American made companies and Frontier Labs are doing.

我觉得他们偏离了重点。

I feel like they're missing the mark.

我不太确定具体原因，也许是专家混合系统这类东西，但有人正在被谎言蒙蔽。

I don't quite know what it is, whether it's this mixture of experts thing, but there's there's someone's being sold a lie.

我不知道是我有问题，还是我看着这个Kimi K2模型时的反应太过夸张。

And I don't know whether it's me or whether it's me, like, looking at this Kimi k two model and being like, wow.

它真的太惊艳了。

It's so amazing.

是啊。

Yeah.

当我思考中国与美国在开源公司和闭源公司方面的角色时，至少让我感到安慰的是，许多软件层面的创新突破确实发生在这些私人AI实验室里。

When I think about the role that China plays versus The United States in terms of, like, open source companies or closed source companies here in The US, The the thing that is reassuring to me at least is a lot of these innovative breakthroughs that happen on the software level actually do happen in these private AI labs.

我们确实获得了思维链和推理能力，还有一系列迅速成为标准的新创新。

We do get, like, chain of thought and reasoning, and there's, like, this whole slew of new innovation that becomes standard very quickly.

这些都发生在美国的AI实验室里。

That all happens in The United States AI labs.

就我们所知，美国的AI实验室仍在以最快速度取得最多进展，他们创造了最多的创新，然后正如我们在节目早期描述的那样，这些创新开始逐渐渗透——无论是自愿还是被窃取——并被应用到新模型中，它们完全在成本和效率方面淘汰了底层，因为这基本是他们唯一能做的。

And as far as we're concerned, the AI labs in The US still have they're making the most progress the fastest, they are creating the most innovation, and then what you kinda see, like we described earlier in the episode, is that innovation starts to trickle down, whether it's voluntary or whether it's stolen, and it gets implemented into these new models, and they just completely cut out the bottom in terms of cost and efficiency because that's kind of all they're able to do.

他们无法获得像黄仁勋那样的GPU资源。

They don't have access to the resources of, like, of GPUs from Jensen Huang.

也无法获得500亿美元资本支出，仅用于支付员工薪资和薪酬。

And NVIDIA, they don't have the access to $50,000,000,000 of CapEx just to spend on employees, just to spend on salaries and compensation.

所以在我看来，我们其实做得相当不错。

So it seems to me like, I mean, we're still doing very well.

只是中国非常擅长将技术规模化应用，并以开源方式实现。

It's just China is very good at implementing the technology and applying it at scale in a way that's open sourced.

关于开源这件事，有很多值得称道之处，因为它非常令人印象深刻，有点像我们早期在美国看到的那种社区协作，但美国一旦做得更好就将其封闭起来了。

And the open source thing, there's there's a lot to say for that because it's it's very impressive, and it's kind of this community effort that we saw early days with The United States, but once they became better, closed it off.

于是情况就变成：创新先出现在Kimi这样的公司，接着你看到DeepSeek实现了它，然后又看到Quen也采用了，突然间这项技术就在三者间同步发展——因为完全开源，他们公开所有代码和权重参数，这让它们更容易蓬勃发展。而美国的创新则基本发生在封闭高墙内，只有在新模型发布时才会泄露到外界，人们才能逆向解析其工作原理。

So what happens is you get innovation in one company like Kimi, and then you see it implemented in DeepSeek, and then you see it implemented in Quen, and then suddenly this technology is is kind of synchronously growing between the three because it's all open source, they're publishing all the code, all the open weights, and it's much more easier for them to thrive, whereas innovation in The United States very much happens behind a closed wall, and it's only leaked out at the advent of a new model when they release it to the world and people kind of reverse engineer how it works.

嗯。

Mhmm.

我最近在《金融时报》上读到一篇采访黄仁勋的文章，他原话说'如果中国继续当前的发展道路，而美国不加大能源生产力度，中国将赢得AI竞赛'。

I was reading an article in the Financial Times where they interviewed Jensen Huang, and he said verbatim that China will win the AI race if they continue, down the path that they're currently on and if The US doesn't kind of ramp up their energy production.

他更广泛地指出，中国的开源策略非常有效——在你刚提到的种种限制条件下，他们正是用这种方式构建这些新AI模型的。

He was making a wider point that their open source strategy is, pretty effective in the way that they're that they're building these new AI models with the constraints that you just mentioned.

再谈谈开源性的优势，我这里有条推文显示Kimi K2这个新模型基本能在两台MacBook M3 Ultra上运行，总成本才几千美元——这简直疯狂，相当于能在家里私有部署

Kind of speaking more about the open sourceness and the benefits of this, I've I've got a tweet up here which shows that Kimi K two thinking this new model can basically run on two MacBook m three Ultras, which is the like a couple of thousand dollars worth of cost, which is an insane thing to do to run Frontier

AI模型

AI model

进行训练，并用自己的私有数据微调。

at home privately in your house, trained, and fine tuned on any of your own private data.

所以你完全不需要把数据卖给Sam Altman那些人。

So you don't need to kind of, like, sell that data to Sam Ortman or whoever.

简直酷毙了还便宜得要命。

Just super cool and super cheap.

对吧？

Right?

因为你在家里本地运行推理，完全不用担心有人窥探你的查询、提示或研究内容。

Because you're running local inference at home, so you don't have to worry about anyone kind of, like, spying on any of your queries or your prompts or your research.

所有操作都在家里完成，我觉得这简直太酷了。

It's just all at home, which I thought was super cool.

关于开源部分还有个有趣的点，乔什，就是他们这次新发布采用了MIT许可证或调整版的MIT许可证。

The other part of the open sourceness, which I found interesting, Josh, was the fact that they had an MIT license with this new release or an adjusted MIT license.

我们稍后会深入探讨这点。

And we'll dig into that in a second.

但重点是，当DeepSeek发布首个重磅开源模型震撼世界时，当时根本没有配套的重要许可证。

But the point being, when DeepSeek released their first major open source model and took the world by storm, there wasn't any major licenses around that.

所以你基本可以下载后对它为所欲为。

So you could pretty much download and do whatever the hell you wanted to it for it.

你可以将其集成到自己的产品中，无论你是美国创业者。

You could implement it into your own product, whether you're an American founder.

假设你将这个功能扩展到百万用户规模，而这个功能正是基于DeepSeek模型开发的，你完全不需要向该团队致谢。

If, let's say, you scale that up to a million users that use a feature that was, leveraging that deep seek model, you wouldn't have to credit that team at all.

Kimi K Two在这方面采取了不同策略，他们发布的MIT许可证规定，当产品用户量达到1000万或2000万时，必须展示Kimi K Two标识并声明底层使用了该模型。

Kimi k two kind of, like, takes a step in a different direction here where they've released an MIT license where I think if you hit I think it's either 10,000,000 or 20,000,000 users for your product, you need to show the Kimi K two label and say that, listen, I'm using this model under the hood.

但这个许可证确实存在一些特殊之处。

But there's there's some differences with this license.

对吧，Josh？

Right, Josh?

我们能深入探讨这个问题吗？

Can we can we dig into that?

我认为这是经过修改的。

I believe it's it's modified.

虽然不清楚具体修改程度，但确实能感觉到这里的条款有所不同。

I don't know to the extent that it is modified, but I know that there is something different going on here.

这上面写了什么？

What does this say?

我们唯一的修改条款是：如果该软件或其衍生作品用于月活跃用户超过1亿或月收入超过2000万美元（或其他货币等值）的商业产品或服务，您必须在相关产品或服务的用户界面上显著展示Kimi K Two标识。

Our only modification part is that if the software or any derivative works thereof is used for any of your commercial products or services that have more than a 100,000,000 monthly active users or more than 20,000,000 US dollars or equivalent in other currencies in monthly revenue, you shall prominently display Kimmy k two on the user interface of such product or service.

这倒是个有趣的营销策略。

That's a fun little marketing ploy.

有道理。

Fair enough.

有道理。

Fair enough.

知道这让我想起什么吗，乔什？

You know what it reminds me of, Josh?

什么？

What's that?

这就像Meta试图用他们的Llama模型做的事。

It's what Meta tried to do with their llama models.

对吧？

Right?

所以Meta是我能想到的唯一一家走开源AI路线的美国大公司。

So Meta is the only other major American company that I can think of that went down this open source AI route.

当时的目标或者说初衷，基本上是想拉平Meta与OpenAI及其他前沿AI实验室之间的差距，后者已经遥遥领先。

And the goal or the intended goal at the time was to basically level the playing field between Meta and OpenAI and other frontier model AI labs, which had raced so far ahead.

因此，如果你免费开放所有这些尖端AI技术，让任何人都能使用，那么OpenAI和其他前沿AI实验室对此收取的溢价就会相应降低。

So if you released all this cutting edge AI tech for free and accessible to anyone, then it kind of drives down the cost premium that OpenAI and all these other Frontier AI Labs can charge you, to access this thing.

中国正在美国AI股市上大举做空。

China's doing that as a vast hole on the, on the American AI stock market.

对吧？

Right?

所以我们看到，消息一出，英伟达股价就暴跌了4.2%左右。

So that's why we saw, like, NVIDIA crash, I think, 4.2% on the news getting released and such.

乔什，我很好奇这是否会戳破美国的资本支出泡沫。

I'm curious whether this kind of pops the bubble and the CapEx bubble in America, Josh.

这么说是不是有点疯狂？

Is that a crazy thing to say?

我是说，市场对这消息的反应相当本能。

I mean, the market's reacted pretty viscerally to this news.

我...我觉得我对此没什么意见。

I I don't think I have a problem with this.

我不认为这会戳破泡沫。

I don't think it's popping a bubble.

我不认为我们会有麻烦。

I don't think we're in trouble.

我认为只要我们能继续保持略微领先或至少持平，就完全没问题。

I think this is just totally fine so long as we continue to stay slightly ahead or at least at par.

我觉得我们在软件开发、软件分发和产品创造方面确实非常出色。

I think we're really excellent at making software, distributing software, creating products.

我认为中国非常擅长毫无顾忌地创新和部署，不需要经历美国主要面临的那些繁琐流程和知识产权问题。

I think China's really good at shamelessly innovating and deploying without needing to go through all of the hoops and intellectual problems that The United States mostly has.

因此我不认为这会导致任何形式的泡沫破裂。

So I don't think this will lead to any sort of bubble popping.

我认为许多前沿创新仍在美国发生。

I think a lot of the frontier innovative stuff still happens in The US.

我开始感到些许担忧的时刻，是当技术转向具身人工智能时。

The place where I will begin to start to get a little worried is when this switches to embodied AI.

一旦我们从大型语言模型转向将这些技术植入机器人或实体硬件，我认为问题就会显现。

Once we start moving from large language models to implementing these into robots, or implementing these into physical hardware, that's where I think we have problems.

在软件领域，我们表现出色，占据优势，所有人都在投入巨额资金。

On the software front, we're good, we're crushing it, everyone's spending tons of money.

但在硬件方面，我们不具备同样的领先地位。过去三十到五十年间，我们已将制造能力外包给其他地区，众所周知，美国本土已难以实现成本效益生产。

On the hardware front, we don't have the same lead, and over the last, what, thirty to fifty years, we've kind of outsourced our manufacturing capabilities to other places, and therefore are just kind of I mean, everyone knows, we just can't really make things cost effectively here in The United States.

若与中国展开具身AI（如人形机器人、特种机器人等）的竞速，情况将变得棘手——因为中国在原子层面的制造能力具有显著优势，而实体制造比代码复制困难得多。你可以窃取开源代码稍加创新，一夜之间推广给十亿用户，但这套模式不适用于机器人迭代。

If we are at a foot race with China, when it comes to making embodied AI, like humanoid robots, specialized robots, whatever it may be, that's where things start to get a little bit scary, because that's where there is a significant lead, and that lead just comes in the form of atoms, which are much more difficult to move than bits, because you can steal some open source code, create this slight innovation on top, roll it out to a billion users overnight, and that's innovation.

人形机器人从二代升级到三代的过程，完全不是这种发展模式。

That does not happen between version two and version three of your humanoid robot.

实际上需要用工厂、真实的材料、人员和场地来建造，这非常困难且具有挑战性。

Actually have to build it with a factory, with real materials and people and places, and it's it's very difficult and challenging to do.

中国极有可能成为其中的最大赢家。

China very much stands to be the largest winner in that.

所以在软件方面，我感到非常有信心，就目前而言，这就是我们正在全力竞争的领域。

So I think on the software front, I feel really confident, and as of now, that's all that we're battling on.

但在不远的将来，当人工智能开始具身化，当AI以物理形态出现在我们周围的世界时，那似乎是一个我会更关注中国投资而非美国投资的领域。

But in this near future where things start to become embodied, where AI b card becomes physically manifested in the world around us, that that seems like a place where I would start looking at Chinese investments a little bit more than the American ones.

好的。

Okay.

我...我觉得我可能要稍微反驳一下，有合理证据表明在发展到具身AI之前，软件方面就值得看跌。

I I I think I might push back a little bit and say that there is reasonable evidence to be bearish on the software side before it gets to embodied AI.

我是说，可以从几个角度来考虑这个问题。

I mean so so a few ways to think about it.

在这些领域的资本支出方面存在着巨大的差异。

There is such a gross discrepancy when it comes to capital expenditure for these things.

一方面，美国花费数万亿美元专门用于训练AGI或最顶尖的AI模型。

On one side, you've got The US spending trillions of dollars literally to train AGI or the best AI models.

而另一方面，你们的投入仅在数亿美元量级，相差了一个数量级。

And on this side, you're you're in, like, the hundreds of millions of dollars, which is like an order of magnitude less.

对吧？

Right?

所以这里存在明显的资源不对等现象，无论是训练架构、训练设计还是硬件制造环节。

So there's an obvious mismatch here that we aren't seeing, whether it comes down to training architecture, training design, or just kind of like hardware manufacturing.

我不清楚这种优势具体体现在哪里，但中国已经找到了突破口，并能有效利用这个杠杆实现赶超或与美国并驾齐驱。

I don't know where that, kind of advantage is being played, but the Chinese have found it, and they are able to kind of really push down on that lever to get ahead or on par with The US.

而且他们已经持续多年成功保持这种发展态势。

And they've been able to successfully do this for years now at this point.

深度求索（DeepSeek）可以说是第一个测试案例。

DeepSeek was kind of like test case one.

自那以后，我已经看到至少50个开源模型来自中国的前沿AI实验室。

Now I've seen, like, you know, at least 50 open source models come out of, Chinese frontier AI labs since then.

第二点，美国政府并非没有尝试过限制他们。

Number two, it's not like the US government has kind of like not tried to to constrain them.

我们已经实施了一系列不同的制裁措施，包括限制英伟达和美国其他制造商可以向中国出售哪些GPU，但这依然没能阻止他们。

We've imposed a number of different sanctions, which include, you know, constraining which GPUs NVIDIA and other manufacturers within The US can sell to China, but that still hasn't stopped them.

尽管面临这些种种限制，他们依然能够维持并训练这些前沿的人工智能。

They've been able to maintain and train these frontier AI intelligences despite all of these different things.

所以我认为，如果从另一方面来看，即便你拥有一个非常酷的开源模型，那又怎样？

So I think if I were to look on the other side of this, it would be so what if you have an open source model that is super cool?

为什么你现在不使用它呢？

Why aren't you using it right now?

就像我虽然使用GPT-5，但并没有经常使用Kimi K2，尽管它可能比GPT-5更优秀。

Like, I'm not using Kimi k two regularly even though I use GPT five and it might be better than GPT five.

对我来说，答案很简单。

And the answer for me is pretty simple.

我已经深深融入OpenAI的生态系统，对此相当满意，因为它能记住我的使用习惯。

I'm locked into an ecosystem in OpenAI that I'm pretty happy with, which is it has memory on me.

它了解我是谁。

It understands who I am.

它掌握了我与它之前所有对话的上下文。

It has a context of all the previous chats that I have with it.

但最重要的是，Josh，如果我的账户或使用过程中出现问题，有一个我可以求助的社区。

But also most importantly, Josh, if there's an issue with something on my account or something that I'm trying to use, there's a community that I can access.

有一个我可以联系的支持团队。

There's a support team that I can speak to.

有一个支持我的软件生态系统。

There's a software ecosystem that supports me.

对吧？

Right?

相比之下，如果我转投Kimi K two，自己安装配置，然后还得自行解决问题。

Versus me jumping ship to kind of Kimi K two, setting it up on my own and then having to like troubleshoot it myself.

我认为很多人都会因此望而却步。

I think a lot of people will be disincentivized to to do that.

这确实很困难，但我的意思是，我们看到市场力量来自双方。

It's it is difficult, but oh, I mean, we're seeing market forces from both sides.

对吧？

Right?

就像，我看到你在这里某个地方放了一个链接，提到Kursar和Windsurf的新AI模型，他们正在使用某种中文模型。

Like, I I saw you included a link here somewhere where Kursar and Windsurf's new AI models, they they were using some sort of Chinese models.

事实上，它们是用中文思考的，我觉得这非常有趣，美国制造的产品现在居然用中文思考。

In fact, they were thinking in Chinese, and I found this really fascinating that, like, American made products are now thinking in the Chinese language.

所以在商业层面上这确实是个问题，API成本至关重要——如果能用60美分获得百万token而不是10美元，这会直接影响企业利润。

So that's certainly a concern in terms of the commercial side, where those API costs really matter, where if you can get a million tokens for 60¢ versus $10, that's that really affects the margins of your business.

对我们这样的普通用户来说，确实没有使用Kimi k two的动力。

For consumers like us, there there's no real interest to use Kimi k two.

你之前提到的现象——实际上可以在两台搭载M3 Ultra芯片的Mac Studio上运行量化版的Kimi k two，生成速度大约是每秒13到15个token。

And the phenomenon you spoke about earlier where you can actually run a quantized version of Kimi k two on two Mac studios running the m three ultra chips, it generates tokens at, like, 13 to 15 tokens per second.

所以速度非常慢。

So it's very slow.

慢。

Slow.

你每秒只能得到一两句话，这速度要慢得多。

You're you're getting like a second a sentence or two every second, which it's it's much slower.

会感觉很迟钝。

It's gonna feel groggy.

体验不会好。

It's not gonna feel well.

有理由认为这种情况会改变，因为今年——有趣的是苹果目前是唯一支持这个的电脑厂商。

There's a case to be made that that changes because this year and it's funny that Apple's really the only computer that that supports this now.

他们将发布m5 Ultra芯片作为新版本，观察其表现会很有趣。

They're releasing the m five Ultra, which will be the new version, and it's gonna be interesting to see how it plays out.

有个有趣的插曲想和你分享EJess，你可能会觉得酷——运行在这些苹果电脑（Apple Studios）上的版本，

What I found interesting, this one side note actually that I wanted to share with you, EJess, because you might find it cool too, is the version that runs on these Apple computers, the Apple Studios.

是略微量化过的版本。嗯。

It's a it's a slightly quantized version, and Mhmm.

我最近在特斯拉的财报电话会议上听说了这件事，他们最近召开了股东大会，我们本周晚些时候会有一期节目专门讨论这个。

I heard about this, and I learned about this recently in the Tesla earnings call that they had the shareholder meeting recently, we're and gonna have an episode on this later this week.

但有个有趣的事情是埃隆在节目中提到量化AI与浮点AI的区别。

But there's this interesting thing that Elon mentioned during the episode where he was talking about quantized versus floating point AI.

我当时就想，这到底是什么玩意儿？

And I was like, what the hell is that?

你为什么要花这么多时间讨论这个？

Like, why why are you spending so much time talking about this?

这根本说不通啊。

It doesn't make sense.

后来我了解到，很多AI模型会使用小数点后多位数据来获得更精确的结果，这就是浮点运算。

And what I realized is a lot of AI models, they they use, like, many, many points after the decimal in terms of data to get more precise results, and that is floating point.

当你对模型进行量化时，会去掉小数点右边的所有数据，只保留整数部分。这样可能会损失高达60%的精度，但能获得更快的效率、更显著的性能提升和成本优化，而且可以直接在本地设备上运行。

When you quantize a model, you remove all of the data to the right of the model, and you just go to single integers, so you lose the variance of maybe up to 60%, but you gain so much faster efficiency, so much better speed improvements, cost improvements, and you can actually run it locally on these things.

所以我觉得观察人们如何权衡模型精度与成本效益、运行效率之间的关系很有意思。

So I think it's interesting to see the different decisions that people are making in terms of, well, how precise does the model have to be versus how cost effective and how efficient does it need to be?

我们在Kimi Gay身上也看到的是，很容易过度追求效率，但这可能不是OpenAI的既定目标，如果他们真想这么做，完全可以对这些模型进行签名和量化处理。

And what we're seeing with Kimi Gay too is it's very easy to to over index on the efficiency, but maybe that's not the stated goal of OpenAI, where if they really wanted to, they could sign and quantize these models.

他们可以更多地采用整数类型计算。

They could go more to integer type compute.

我刚才在想的是他们的处理方式，因为可能只是Kimi在优化速度和效率，而下游效应就是它确实非常快，而OpenAI目前还没有专门针对这方面进行优化。

And it was just something I was thinking about is how they approach them, because it could just be, well, Kimi's just kinda optimizing for speed and efficiency, and the downstream effect is it's also really fast, whereas OpenAI kinda hasn't really optimized for that specifically yet.

没错。

Right.

对此的反驳观点可能是，乔希，它在所有我们评估其他美国模型的基准测试中都表现优异。

And the the counterargument to to that point would be, well, Josh, it's crushing all the benchmarks that we've evaluated all the other American models on.

对吧？

Right?

是的。

Yeah.

所以它显然要好得多。

So surely it's much better.

而我对此的反驳会是，基准测试在实际使用中并不真正体现价值。

And my my pushback on that would be like, well, benchmarks don't really materialize in real life use.

就算它在人文学科上次考试中碾压50%又怎样？

So what if it crushes 50% on humanities last exam?

对我来说实用吗？

Is it useful for me to use?

它能理解我想表达的意思吗？

Does it understand what I'm trying to say?

它能理解我输入的提示上下文吗？

Does it understand the context of the prompts that I'm putting into it?

另一方面，关于量化这点，乔什，我认为很多美国前沿AI实验室如OpenAI、谷歌等，实际上拥有足够的算力来提供最佳体验——用最高精度的浮点运算来呈现，但他们把大部分算力都用于训练我们尚未得见的下一个大型模型。

The other side of this, you know, on the point of quantization, Josh, is I think that a lot of frontier American AI labs like OpenAI, Google, etcetera, actually have enough compute to give you the best experience, the the highest floating point experience to put it to put into that context, but they're using the majority of that compute to train the next big model that we haven't even seen yet.

对吧？

Right?

上周有消息爆出OpenAI正在这么做。

There were there was news that broke last week that OpenAI is doing this.

对吧？

Right?

所以从技术上讲，他们有足够的算力全年为你提供优质服务，但他们却将70%的算力用于训练GPT-6。

So, technically, they have enough compute to give you, like, amazing service all year round, but they're using 70% of that compute to train GPT six.

因此我认为这只是当前阶段的优先级问题，直到我们达到某种平衡——这些AI模型足够好用为止。

So I think it's just a matter of prioritization right now until we reach some kind of parity that these AI models are are good enough.

但我要说，根据本期节目讨论的所有内容，有一个明确的赢家，那就是消费者。

But, I will say from all of the things that we've discussed on this episode so far, there is one clear winner, and that is the consumer.

就是你、我和所有听众，我们几乎不用花费什么就能获得前沿级别的人工智能。

It's you, I, and everyone listening to the show, which basically gets access to frontier level intelligence for the cost of next to nothing.

完全免费下载，在家私有化运行。

Download it completely free and run it privately at home.

在我打开的这条推文里，基本上说每个闭源模型都有对应的开源替代方案，并列举了诸如Sonnet 4.5这样的例子。

On this tweet that I have pulled up here, it basically says for every closed model, there is an open source alternative in it, and it goes through a list like Sonnet 4.5.

还有GLM 4.6。

You've got GLM 4.6.

Grok CodeFast，你有GPT开源版

Grok CodeFast, you've got GPT OSS.

GPT五，你有Kimi K2思维版

GPT five, you've got Kimi k two thinking.

这样的例子还在不断增加

And it just goes on and on and on.

如果我们回顾一年半前，甚至两年前，这份清单根本不存在

And, if we look at this kind of like a year and a half, ago, maybe even two years ago, this list would be nonexistent.

那时候只有闭源阵营的Frontier AI实验室，开源阵营为零

It would just be Frontier AI Labs on the closed source side and zero open source side.

所以看到这种进展真的非常令人鼓舞

So to see this kind of progress is really, really encouraging.

嗯

Mhmm.

是啊

Yeah.

这将是一场竞赛。

It's gonna be a race.

这将是开源与闭源之间的较量。

It's gonna be a battle between open and closed source.

或许这甚至算不上真正的较量。

And and perhaps that's not even the battle.

可能开源会一直存在，直到它们赶上闭源，然后全面转向闭源。

Perhaps it's open source until they catch up to closed source, and then it's closed source across the board.

所以观察这些发展会很有趣。

So it's gonna be interesting to see the developments.

我们即将迎来一批新模型。

We have a new batch of models that are coming.

我们现在处于一个奇怪的过渡期，Gemini 3有望很快发布。

We're kind of in this weird limbo where Gemini three is hopefully coming soon.

我们将获得一些新的基准测试数据，而最让我难以接受的一个残酷现实——正如你刚才提到的eJaz——就是所有人都受限于算力。

We'll have some new benchmarks, and and one of the things that that was this harsh truth to kind of wrap my head around, which is what you just mentioned, eJaz, and the fact that everyone's just compute constrained.

比如，OpenAI本可以做出比GPT-5还要惊艳两倍的模型，如果他们真想这么做的话。

Like, OpenAI could have made GPT five probably twice as impressive if they really wanted to.

他们只是没有足够的算力来支撑，那样成本会高得离谱，运行速度也会慢得无法接受。

They just have no compute to serve that, It would've been way too expensive and way too slow.

所以不是技术上实现不了。

So it's not that it's it can't it can't be done.

只是人们目前没有足够的资源来实现它。

It's just that people don't have the resources to do it.

所以这是一场持续的平衡游戏，看看各公司如何在这条曲线上定位自己会很有趣——他们愿意在计算资源上投入多少成本，与他们实际可用于训练这些模型并大规模部署给用户的资源之间的权衡。

So it's this constant balancing act, and it's gonna be fun to see how how companies kinda slot themselves into that that curve of, like, how much they wanna spend on compute versus cost versus just what they have available to actually use to train these models and deploy them at scale to users.

今天的节目就到这里，各位。

And that's it for today, folks.

非常有趣的一期节目。

Super fun episode.

开源AI能如此迅速地追赶闭源集中式AI，总是让我感到惊讶。

I it is always surprising to me how quickly open source catches up with closed source centralized AI.

我总以为开源会滞后几年，但现在看来只滞后几周了。

I always think kind of like it's gonna lag a few years, and now it's come down to the fact that it's lagging a few weeks.

我们这周行程满满当当。

We have a jam packed week.

明天谷歌可能会发布新的Nano Banana模型。

We have potentially a new nano banana model being released by Google tomorrow.

祈祷吧。

Crossed.

我正为此祈祷呢。

I'm praying for that.

但愿如此。

Fingers crossed.

我也在为此祈祷。

I'm also praying for that as well.

我们还有第二期节目，将讨论特斯拉投资者日的重磅新闻。

And we have a second episode based on Tesla's Investor Day, which had some really jam packed exciting news.

现在听着。

Now listen.

如果你希望美国赢得这场AI竞赛，别搞错了，这确实是一场竞赛。

If you want The US to win this AI race, and make no mistake, it is a race.

你需要订阅美国的AI YouTube频道，我们就是其中之一。

You need to subscribe to American AI YouTube channels, one of which is us.

请订阅我们。

Please subscribe.

点击通知按钮。

Hit the notification buttons.

无论你在哪里收听，都请给我们评分。

Wherever you're listening to, give us a rating.

这些对我们帮助很大。

We are helped by these so much.

这能大大提高我们的知名度。

It is bringing so much awareness.

算法现在对我们很有利。

The algorithm is favoring us.

我们获得了所有这些精彩的观看量和新观众。

We're getting all these wonderful views and new incomers.

上周我们新增了一千名订阅者，这简直太疯狂了。

We've got a thousand of you from last week, which is just insane.

大家好。

Hello.

欢迎来到本频道。

Welcome to the channel.

希望你们喜欢我们的内容，下期节目再见。

We hope you enjoy the content, and we will see you on the next one.

是的。

Yeah.

在我放他们一马之前，我得先确认一下。

Before I let them off the hook, I'm I'm checking.

我正在做数据统计更新。

I'm doing the stat update.

上周观看视频的观众中有83%没有订阅。

83% of the people that watched last week were not subscribed.

如果你是在YouTube上看的，伙计们别来了。

If you're watching this on YouTube, don't come guys.

或者去Spotify，这是我最推荐的播客收听平台。

Or go on Spotify, my preferred place to find this podcast.

它是最好的。

It's the best.

说真的，我实在找不到更好的词来形容这些人了。

I'm telling you, I I don't know how to describe this people any better.

Spotify太棒了。

Spotify is so good.

你可以看视频、听音频，还能锁屏播放，而且不需要会员。

You have the video, you have the audio, you could turn it off and lock your phone without needing a premium membership.

请去那边。

Please go go over there.

去那边留言吧，因为评论区也挺热闹的。

Go leave a comment over there because the also the comment section's kinda popping too.

是啊。

So yeah.

总之，嗯。

Anyway Yeah.

感谢你的

Thank you for

所有支持。

all the support.

我们不会挑剔你在哪里收听。

We do not pick and choose wherever you listen.

尽管去吧。

Go for it.

就这样吧。

There you go.

好的。

Alright.

我们下期再见。

We will see you guys in the next one.

一如既往地感谢观看。

Thank you for watching as always.

非常感谢。

Much appreciated.

祝安好。

Peace.