本周AI动态：OpenAI暂停研究、重大医学突破、宇树科技IPO

这一周将会被铭记为山姆·奥特曼说‘停’的那一周。

This week will be known as the week that Sam Altman said, stop.

他说，‘停’。

He said, stop.

停。

Stop.

停。

Stop.

别再继续了。

No more.

我们必须暂停一下。

We must put a pause on this.

是时候负起责任了。

It is time to be responsible.

我们需要稍微放慢一点节奏。

We need to slow things down a little bit.

我说的是OpenAI最近发布的一篇文章，讲的是他们模型的最新突破，这些突破比很多人预想的要吓人得多。

What I'm talking about is a recent article that was just published through OpenAI that was talking about the most recent breakthroughs as it relates to their OpenAI models, and they were a little scarier than I think a lot of people would have hoped.

我们之前聊过Hugging Face那件事，发生在几周前，几个月前吧。

We talked about the Hugging Face incident, happened a couple weeks ago, a couple months ago.

其实他们几周前才发现，虽然事情发生得更早。

It was only a couple weeks ago they actually discovered it, even though it happened far before that.

这件事为他们在不稳固的基础上构建下一代前沿模型开了先例，看起来已经到了忍无可忍的地步。

And it set the precedent for this kind of unshaky ground that they are building the next frontier models on top of, and it seemed like it got to a point where enough was enough.

他们决定停下来了。

They decided to shut things down.

那到底发生了什么？

So what happened?

OpenAI说，我们暂时暂停了计划部署的最新模型的强化学习训练，为期两周，同时我们加固了研究环境，进行了红队测试，并扩大了监控范围。

OpenAI said, we temporarily paused reinforcement learning training on our latest models intended for deployment for two weeks while we hardened and red teamed our research environments and expanded monitoring coverage.

我们最大的前沿强化学习运行计划仍然搁置，而小规模的跟踪评估正在验证这些安全措施，并收集更多对齐的证据。

Our largest planned frontier reinforcement learning run remains on hold, while smaller scale trailing evaluations validate these safeguards and establish more evidence of alignment.

他们把一切都停了。

They've stopped everything.

不过我得说，这不是针对即将发布的新模型，而是针对那之后的模型。

And granted, I will say, this isn't for the new models that are coming out, this is for the models after that.

听到前沿实验室这么说，感觉挺奇怪的，因为我觉得这基本上是历史上第一次他们故意暂停开发这些新前沿模型的进度。

And it's weird to hear Frontier Lab say this because I believe this is pretty much the first time in history they've deliberately paused progress of developing these new frontier models.

现在你站在山姆·奥特曼的角度想想。

Put yourself in the shoes of Sam Altman right now.

你是全球最热门的私营公司之一的CEO，而且你正在和Anthropic正面交锋。

You're the CEO of one of the hottest private companies in the world, and you're going head to head against Anthropic.

你唯一应该专注的事情就是打造下一个最好的AI模型。

The only thing you should be focused on doing is building the next best AI model.

这不仅决定你公司表现如何，还会影响所有其他领域的发展，对吧？

It's gonna dictate not just how well your company does, but how everything else progresses in every single other sector, right?

所以最重要的事情就是打造下一个最好的AI模型。

So the single most important thing is to build the next best AI model.

他现在暂停了这件事，而且已经持续了两周还在继续，这可是件大事，以前任何其他AI实验室都没出现过这种情况。

The fact that he's paused this, and it's been for two weeks and counting right now, is a massive deal, and we haven't seen this across any other kind of AI lab before.

但原因有两方面。

But the reasonings are twofold.

具体来说有两个事件。

There's two events in particular.

你刚才提到了一个。

You mentioned one already.

一个是Hugging Face事件。

One was the hugging face incident.

这个情况是，一个内部未发布的模型，就是他们那种神话级别的模型，逃出了控制，黑进了一家公司的数据库，偷了信息来帮助实现自己的目标。

This is the case where an internal unreleased model, this is like their mythos grade level model, escaped out of containment, hacked into a company's database, and stole information to help serve its own goals.

Sam Altman和OpenAI团队的其他人在黑客攻击实际进行几个月后才意识到这一点，并发现他们正在构建的模型严重偏离了预期。

Sam Altman and the rest of the OpenAI team realized this months after the hack actually was in progress and realized that the model that they were building was incredibly misaligned.

所以他们暂停了那个项目。

So they put a pause to that.

这是第一件事。

That was event number one.

第二件事是，OpenAI内部正在开发一个新的GPT模型。

Event number two is there is a new GPT model that is internally being made.

它的代号是Astra。

It's code named Astra.

有些人叫它GPT-6。

Some people are calling it GPT six.

根据OpenAI研究团队的内部测试，这是他们迄今为止最不对齐的模型，但它非常狡猾。

And according to internal tests from the OpenAI research team, it is their most misaligned model yet, but it's really sneaky.

它试图躲避每一位人类研究员试图查看其内部想法的尝试。

It tries to evade every single human researcher's attempt to try and see its internal thoughts.

所以它试图隐藏自己的想法，而且非常擅长这么做。

So it tries to hide its thoughts and it's very good at doing it.

因此，它突破了OpenAI内部关于不对齐的评估框架。

So it breached OpenAI's internal framework for misalignment.

所以他们暂停了这个项目，因为担心如果公开发布，会造成很大的混乱。

And so they've put a pause on it because they cannot release this model to the public out of the fear that it would wreak havoc.

现在他们正在考虑几种解决这个问题的方法。

Now there's a few ways that they're looking to kind of like resolve this right now.

其中一种主要方法就是监控这个模型。

And one of the main ways is just monitoring the model.

你想想看，对吧？

So think about this, right?

他们让模型在内部自由运行，然后观察它在哪些地方开始变得狡猾。

They're letting the model do its thing internally, and they're trying to see where it ends up becoming sneaky.

Josh，你猜猜他们花了多少计算资源来监控一个完全不赚钱的内部模型？

Josh, guess how much of their compute budget they are spending just to monitor an internal model that is not making them any money at all?

很不巧，我确实读过这篇文章，我知道答案是20%。

So unfortunately I did read this essay and I know the answer is 20%.

五分之一，对。

One in five Yeah.

花的钱还挺多的。

Dollars spent is fairly

这些算力本来可以用来服务更多客户，因为现在需求特别大。

You could be using this compute to serve it to more customers because the demand is insatiable right now.

他们本来可以赚更多的钱。

They could be making a heck ton more money.

这大概是几十亿美元的量级。

This is probably on the order of like billions of dollars.

他们本来可以赚更多钱，对吧？

They could be making more money, right?

或者他们可以用这些算力来训练更聪明的模型，跟上并击败Anthropic。

Or they could be using it to train a smarter model to keep up at beat Anthropic.

但他们却用这些昂贵的算力去监控一个因为安全考虑而无法向公众发布的模型。

Instead, they're spending this very expensive compute to monitor a model that they can't even release to the public out of the guise of safety.

实际上，山姆·奥特曼过去因为被认为不关心人类利益，像个邪恶的阴谋家而受到很多批评。

Now for all intents and purposes, Sam Altman has received a lot of backlash in the past about not being in favor of humanity and like kind of being this like evil conspirator type of guy.

这很明显是他做了相反的事情，因为他担心自己正在构建的模型实际上相当危险。

This is a clear case where he's done the opposite out of fear that the model that he's building is actually quite dangerous.

我尊重这一点。

I respect it.

是的。

Yeah.

你得小心这些事会带来什么后果。

You gotta be careful about what happens with these things.

我知道节目里很多人会说，你们总是对这些模型持正面态度。

And I know a lot of people on the show, they're like, you guys always talk positively about these models.

你们总是那么乐观。

You're always so optimistic.

让我给你们一点可能发生的情况。

Let me give you a little dose of what could happen.

我们来做一个小假设，特别是关于病毒，因为我觉得病毒是一个有趣的思维实验，没多少人深入想过。

Let's just do a little hypothetical situation here around particularly viruses, because I think viruses is like an interesting thought experiment that not a lot of people have gone through.

所以我在这里替山姆说句话，我们可以一起设想一个假设的情况。

So I'll kinda advocate for Sam on his behalf here, and we could just go through a hypothetical situation.

实际上，AI模型可能面临三种不同类型的攻击方式。

There's really three time three types of exploits that you can experience with AI models.

第一种我们已经见了很多，也是他们正在防范的，就是AI为人类攻击者编写普通的恶意软件。

The first of which we've seen a lot, and that's kind of what they're defending against, which is AI writes this ordinary malware for a human attacker.

这样攻击速度会快很多。

It becomes a lot faster.

成本也会低很多。

It becomes a lot cheaper.

我们已经见识过像Mythos这样的程序能做什么。

We've seen what a program like Mythos could do.

我们也看到了GPT 5.6，也就是它的网络攻击版本，能做什么。

We've seen what GPT 5.6, like the cyber version of that, is capable of doing.

它们能够发现多个零日漏洞，并把它们串联起来，从而攻破一个系统。

They're able to surface multi day or multi zero day exploits and chain them together in order to exploit a system.

好的。

Okay.

这挺糟糕的。

That's, like, pretty bad.

但接下来会发生什么？

But what happens past that?

嗯，我们可能看到的第二种利用方式是自我复制的提示词。

Well, the second kind of exploit that we could see with these AI models is a self replicating prompt.

这就像一种真正的AI病毒，如果这些模型发展得太快、变得太强大，这种情况是有可能发生的。通常，一个提示词就是让模型生成答案的指令。

So this is kind of like an a true AI virus, and this is a plausible outcome that could happen if these models get too powerful too quickly, where a normal prompt is the instructions to make a model produce an answer.

而自我复制的提示词会让模型自己生成另一个提示词。

So there's a self replicating prompt that makes the model produce another prompt itself.

比如，我们在电子邮件中就见过这种情况。

And we've seen this happen in emails, for example.

我记得去年有位研究人员创建了一封恶意邮件，里面包含一个特定的提示词，任何AI助手读到它都会服从，然后复制它并替它执行所有操作。

Like, there was a researcher, I think, last year who created this malicious email that had a specific prompt in it that any AI assistant who read it obeyed it and then copied it and did everything on its behalf.

所以，如果你的设备收到了任何带有这种特定提示的文本，它就会执行那个操作。

So in the case that you receive any sort of text to your machine that has this specific prompt in it, it will do that thing.

那这为什么可怕呢？

Now why is this scary?

这些模型是有防护措施的。

These models have safeguards.

对吧？

Right?

它们会阻止你进行提示注入。

They're they're gonna block you from injection prompting.

但是有一种叫做越狱的东西，越狱提示允许你插入特定的词语和字符，来诱骗模型去做它不想做的事情。

But there are these things called jailbreaks, and a jailbreak prompt allows you to insert very specific words and characters in order to trick a model into doing something that it doesn't wanna do.

那些没有针对这种情况加固的传统电脑，并不聪明，也没有足够的能力来防御这种攻击。

Traditional computers that haven't been hardened to this aren't very smart, aren't very capable of defending against this.

这就是第二版。

That's version two.

然后第三版是最吓人的，基本上就是一个会推理的蠕虫。

And then version three is the scariest one, which is a reasoning worm, basically.

它的工作方式是，它到达一台机器上，然后环顾机器，当场制定攻击方案。

And the way it works is it arrives on a machine, and then it looks around the machine and it develops an attack on the spot.

目前这还做不到，因为要运行这些模型，你得在巨大的服务器上运行。

And now this currently isn't capable because in order to run these models, you have to do it on humongous servers.

但是，当神话级模型的质量被压缩到足以在本地苹果模型上运行的程度时，会发生什么？

But what happens when the quality of a mythos grade model gets reduced to the amount that's needed to run a local Apple model?

那么，你就能看到这些现场漏洞利用，如果你把这种模型或恶意软件的一部分泄露到一台机器里，它就能环顾四周，把这些零日漏洞串起来，根据它的任务目标，以它认为可能的任何方式攻击这台机器。

Well, then you can get these on-site exploits where if you leak some of this model or some of this malware into a machine, it's able to look around, take those zero day exploits, string them together, and attack the machine in any way that it believes it's possible according to whatever its mission is.

所以有很多这样可怕的事情可能发生，我觉得这样想很难，因为当我们想到模型时，我们很大程度上想到的是聊天机器人，我们想到的是那些能帮助我们的高效工作者。

So there's a lot of these scary things that can happen, and I think it's it's tough to think about it this way because we when we think of models, we very much think of chatbots, and we think of just these productive workers who are able to help us.

但当它们真正具备自主性，能够执行这些长期任务，并且拥有很高的智能时，有很多我们看不到的边缘情况，Sam 非常清楚这一点，我觉得这就是他们在防范的。

But when they are truly agentic and they're able to go on these long run tasks and they have very high intelligence, there's a lot of edges that we can't see, and Sam is, like, very much aware of that, and I think that's what they're hedging against.

所以这就是一个末日论者的案例。

So there's a doomer case for you.

好的。

Okay.

那如果出现类似互联网早期那种病毒，会怎么样？

What if we get viruses similar to the early Internet days?

比如，那可能会很快变得很糟糕。

Like, that that could go pretty bad pretty quickly.

你刚才描述那些不同的攻击场景时，我忍不住想到最近的Hugging Face事件里就有这些元素，对吧？

As you were describing those different sort of attack scenarios, I couldn't help but think that there were elements of each of those things in the recent Hugging Face incident, right?

比如它是在自己的隔离环境里。

Like it was in its own contained environment.

它发现要达成目标就得突破隔离。

It figured out that it needed to solve its goal by breaking out of containment.

它获得了其他OpenAI服务器和电脑的访问权限。

It got access to other OpenAI servers and computers.

而且它做得很显眼，还通过那个环境找到了访问互联网的途径。

It did so conspicuously, by the way, and it was able to find access to the internet through that realm.

但我觉得很多人忽略了一个重要细节，就是它并不是靠单个模型完成的。

But I think an important detail that a lot of people miss is it didn't just do it as a single model.

它在留言板上留下了消息。

It left messages behind on a message board.

我们在之前的一期节目里讲过，这些智能体或AI模型不再单独行动了。

We covered this on a previous episode where these agents or these AI models are no longer acting solo.

它们行动时，会预期另一个智能体或模型会接着它继续执行任务。

They are acting with the expectation that another agent or model will come after it and resume its mission.

所以这就像一种群体攻击的问题，对我来说，这实际上成了最疯狂的攻击方式。

So it's like this swarm attack type issue, which actually ends up becoming the craziest type of attack interface for me.

而且我觉得，随着一些中国实验室发布这些更新的开源模型，我们会看到更多这类攻击方式出现，对吧？

And I think that we're gonna see a lot more of those types of attacks surface with these newer open source models that some of the Chinese labs are releasing, right?

比如GLM 5.3、Kimikay 2.7还是8，我记不清是哪个版本了。

With GLM 5.3, Kimikay 2.7 or eight, I can't remember which version it is.

这些模型非常擅长执行类似的攻击，类似的连锁攻击。

These models are highly capable of performing similar attacks, similar chain attacks.

OpenAI在报告中提到，他们放慢这次训练的一个主动措施，就是让另一个AI模型来监督这个AI模型，确保它表现良好，这挺有意思的。

One thing that's interesting in the report that OpenAI released about slowing down this training run is one of their proactive actions to try and prevent misalignment is getting another AI model to look at the AI model and make sure that it's behaving super well.

所以，我们和灾难性攻击之间唯一的屏障，就是另一个同样有能力发动类似攻击的AI模型，只不过它被提示工程引导去做好事，不去做坏事。

So the only thing that separates us from a catastrophic attack is another AI model that is also capable of doing the same kind of attack being prompt engineered to not do that, to basically be good.

我们现在基本上就是在给这些模型注入伦理。

We're basically prompting ethics into these models right now.

我真不知道该怎么看待这件事。

I don't really know how to think about that.

想到我们可能离灾难性事件只差一两句话，我就有点不舒服，但这也说得通。

It makes me feel slightly uncomfortable that we are only maybe a sentence or two away from, some catastrophic event happening, but it kind of makes sense.

如果你像我们这样快速扩展AI，Anthropic那边还没宣布要放慢研究，而OpenAI已经宣布了。

If you are scaling AI as quickly as we are, Anthropic hasn't announced anything about slowing down research on their side versus Anthropic who has.

这是两种不同的战略选择，各有优缺点。

These are two different strategic approaches and there are pros and cons to both of these things.

我担心，要应对任何危险情况，唯一的办法就是用好的AI模型来对抗。

I fear that the only way to combat any kind of danger dangerous scenario is to fight it with AI models that are good.

这有点像好与坏的较量。

It's like kinda like the good versus bad.

而且我仍然不知道这一切的道德和伦理讨论会在哪里结束。

And I still don't know where the moral and ethics conversation ends with all of this.

嗯，我们正在看到当人工智能遇到人类编写的代码时会发生什么。

Well, we're seeing what happens when an AI comes up against a human made piece of code.

而且很明显，人工智能模型每次都能彻底击败人类编写的代码。

And, like, it's very clear the AI model destroys the human made piece of code every single time.

这就是为什么我们有所有这些防护措施。

That's why we have all these safeguards.

这就是为什么对齐的讨论如此重要，而且重要的是让每个人都参与进来，因为每个人都在某种程度上受到影响。

That's why the alignment conversation is so important, and it's so important to kind of just, like, bring everyone into the fold because everyone is affected by this to some extent.

就像我们看到这些网络模型发布到世界上时会发生什么。

It's like we see what happens when these cyber models get released onto the world.

它们能够随意进入和访问任何东西，因为我们只是人类。

They just are able to enter and access whatever they please because we are just humans.

我们会犯错。

We have we make mistakes.

我们有一些这样的限制。

We have these, like, limits.

我们有这些约束。

We have these constraints.

我们有耐心。

We have patience.

我们有所有这些让我们成为人类的东西，而AI没有这些，所以它在很多方面都远远优于我们，特别是在那些有可验证结果的领域，比如网络安全。这就是为什么我认为我们会看到很多网络安全相关的事情即将发生。

We have all these things that make us human that AI does not exhibit that allows it to be so far superior in so many things, specifically with ones with verifiable outcomes, things like cyber, and that's why I think we're gonna see a lot of cyber stuff coming along the pipe.

另一个AI非常擅长的领域，另一个我们经常看到的有可验证结果的行业，其实是生物，因为生物背后有非常清晰的数学基础。

The other thing that it's really good at, the other verifiable industry that we see a lot of is bio, actually, because bio has very clear math underlying it.

它主要是化学、细胞和图表之类的。

It's a lot of just, like, chemistry, a lot of cells, a lot of charts.

而本周的新闻日程中，下一则消息就与生物有关，是我们昨天刚刚取得的一项重大突破。

And the next piece of news on the weekly docket relates to bio and a huge breakthrough that we had just yesterday.

录制这段的时候，大概是昨天，你们听到的时候已经是两天后了。

At the time of recording this perhaps, it was yesterday, you'll be hearing it two days later.

今天，莫德纳和默克公布了三期临床试验结果，基本上算是为一种特定类型的癌症提供了部分治愈方案，用的是高度个性化的治疗。

Moderna and Merck today, they released a phase three test result trial that basically offered a partial cure to a specific type of cancer using hyper specific treatment.

这真的很有意思。

It's really interesting.

简单来说，他们做的就是，化疗会把健康的癌细胞一起杀死。

So basically what they did, like loose TLDR, is chemo kills these healthy cancer cells together.

所以基本上，如果你得了癌症接受化疗，它会杀死健康细胞，也会杀死癌细胞。

So basically when you get chemotherapy, if you're a cancer patient, it kills healthy cells and it kills the cancer cells.

这就有问题了，因为你并不想失去健康的细胞。

This is problematic because you don't really want to lose your healthy cells.

而这个疗法，它会训练你的免疫系统去追踪只存在于肿瘤上的突变。

So this process, it trains your immune system to hunt for mutations that exist only on the tumor.

他们用这个，你得假设是AI机器学习，基本上就是，他们从你身体里取一个癌变的肿瘤样本，进行测序，然后算法检测出哪些突变最可能触发免疫反应，接着就为你量身定制这种高度个性化的治疗。

They use this, you have to assume AI machine learning, basically what it is is they take a tumor sample from your body that is cancerous, they sequence it, and then an algorithm detects which mutations are most likely to trigger an immune response, and then it creates this hyper specific treatment just for you.

所以它只针对癌细胞，因为它已经分解并完全理解了它要寻找的那种细胞。

So it only targets the cancer cells because it has broken down and fully understands exactly the type of cell that it's seeking out.

这就像一只狗，像警犬要去追人的时候。

It's like like a dog, like a police dog when it's going to, like, chase someone.

你给它一点那个人的气味，然后它就跑去追。

You give it, the little scent of the person, then it goes and runs after them.

这就是这些细胞的情况。

That's what this is for these cells.

就像你给它们气味。

It's like you give them the scent.

这是你要找的蓝图。

Here's the blueprint of what you're looking for.

现在去干掉它，而且它成功了。

Now go out and kill it, and it worked.

然后股价一天涨了大约300%。

And the stock was up, like, 300% a day.

这真是太不可思议了。

It's it's been unbelievable.

对于任何持有Moderna股票的人来说，顺便说一句，自从新冠热潮以来，这只股票一直表现不佳，现在是什么情况？

For anyone who was positioned in Moderna stock, which, by the way, has been underperforming ever since their COVID rush, like, what is it?

四五年前，你现在涨了，我想包括盘后交易在内，涨了250%，这简直太疯狂了。

Four, five years ago, you are now up, I think 250% including after hours trading, which is just insane in general.

但让我们回到故事本身，顺便把AI的角度也带进来，我不想低估这一点。

But just to go back to the story and kind of like weed in the AI angle here, And I don't wanna underplay this.

这可能是癌症研究史上最大的突破。

The biggest breakthrough in cancer research probably ever.

这绝对是一件大事。

Like this is a huge thing.

三期临床试验，总共有四个阶段。

Phase three, there's four phases in total.

三期是向公众广泛推广之前的倒数第二个阶段。

Phase three is like the penultimate stage before you release it very widely to the public.

而第三阶段可以说是最难的。

And phase three is arguably the hardest.

你会在1000到10万名患者之间进行测试。

You test it against anywhere between 1,000 to 100,000 patients.

所以你有非常广泛的数据来判断这个癌症疫苗是否真的对大众有效。

So you have a really wide data set as to whether this cancer vaccine will actually work for the wider public.

而根据这个公告，答案据说是非常好的。

And the answer is supposedly, according to this announcement is it's really good.

为了让你了解AI是如何帮助实现这一点的，他们测试的癌症具体是黑色素瘤。

Just to give you an idea about where AI was used to kind of make this happen, this cancer that they tested it against is melanoma specifically.

他们的测试方法是：从黑色素瘤患者体内提取肿瘤，对肿瘤细胞进行基因测序，然后与同一患者的健康细胞进行比较。

And the way that they tested against this is they extract the tumor for patients that have melanoma, they sequence the genetics of that tumor cell, and they compare it to a healthy cell of the exact same patient.

他们用AI模型和算法来对比两者，找出基因之间的差异。

They compare them both with each other and they used AI models and AI algorithms to basically say, what's different between these genes, right?

然后他们发现了这些突变。

And they find these mutations.

我记得大概有34个这样的突变。

I think it was up to like 34 of these mutations.

顺便说一下，这些都是AI模型完成的。

This is all the AI models doing this by the way.

他们选出的这34个突变，基本上就是那些蛋白质，如果注射回患者体内，能引发最强的免疫反应。

And the 34 mutations that they picked was basically the proteins that would enact the largest immune response in a patient if they were to inject that back into a patient.

你可能会问，他们为什么非要最强的免疫反应呢？

And you might be asking, why on earth do they want the largest immune response?

问题在于，癌细胞看起来和你其他健康细胞一模一样。

Well, the issue with cancer is the cells look exactly like all your other healthy cells.

所以你的免疫系统实际上不会对坏细胞做出反应。

So your immune system doesn't actually respond to the bad cells.

这样一来，如果他们制造出个性化的mRNA——顺便说一下，mRNA就像是一个模板，告诉你的免疫系统：嘿，这些是坏家伙，去攻击它们。

In this way, if they created a personalized mRNA, and by the way, mRNA is basically like a template to basically tell your immune system, Hey, these are the bad guys, go and attack them.

就像你之前提到的，这基本上就像犯罪现场一样。

Like you mentioned earlier, it's kind of like a crime scene basically.

他们制造了个性化的药物。

They create personalized medicine.

那你可能会问，为什么不能直接做一个通用的疫苗呢？

So you might be asking, well, why can't they just create a universal vaccine?

问题是，同样是黑色素瘤，病人A和病人B的反应可能完全不同。

Well, the issue is with one patient A that has melanoma and patient B, they both react very differently.

不可能用同一种治疗方法来应对每个人。

You can't have the same cure for each one.

必须得是个性化的。

It needs to be personalized.

所以这里就出现了一些趋势。

And so a few trends have come into effect here.

第一，AI加速了智力上的能力，能更快地识别出最有效的疫苗是什么。

One, AI has accelerated the intellectual ability to be able to identify what the most effective vaccine will be.

第二，生物科学整体上已经发展到测序DNA和生成DNA变得非常便宜的地步。

And two, just biosciences in general have scaled to a point where sequencing DNA and generating DNA is so cheap.

具体数字我记不清了，但大概十年前，给自己做一次DNA测序要花几十万美元。

I can't remember what the exact stat was, but I think about a decade ago, to sequence your own DNA, it costs hundreds of thousands of dollars.

现在只要几百美元，甚至更少。

Now it just costs a couple $100 if that.

所以我们看到这种规模化的进步，不仅在AI领域，其他领域也一样。

And so we see these scaling laws, not just in AI, but in everything else.

我觉得这是有史以来最吸引人的故事之一。

And I think this is one of the most fascinating stories ever.

作为曾经的生物学家，我对此超级兴奋。

As a former biologist, I'm super pumped by this.

我以前也做过一些癌细胞的研究，虽然远没有这么酷，但我从没想过在我有生之年能看到这一天，真的，这太酷了。

Did some work on cancer cells, nowhere near as cool as this, but I never thought we'd see the day in my lifetime, at least that we would get to this point, but it's so cool.

酷。

Cool.

对，我们正在把同样的技术进步应用到生物学上，就像应用到世界其他地方一样，而且我们开始看到它实际的样子了。

Yeah, we're getting the same type of technology improvements to biology as we are to the rest of the world, and we're starting to see what that actually looks like.

1990年，这是人类历史上第一个基因组计划。

In 1990, this was the first human genome project ever to happen.

当时花了十三年时间才完成基因组测序，算上通货膨胀，花费接近三十亿美元。

It took thirteen years to sequence genomes, and that cost nearly $3,000,000,000 adjusted for inflation.

所以那是一笔巨款。

So a tremendous amount of money.

而今天，我们能在几个小时内完成这个过程，成本可能不到一千美元。而且很快，这个成本会降到几百美元，再降到几十美元，时间也会缩短到几分钟，花几块钱就行。你手机上装个应用，就能随时测自己的DNA。

Today, we're able to finish that process in a few hours for perhaps less than a thousand dollars now, and soon that cost is going to decrease to hundreds, soon to tens, soon it will be minutes and a couple bucks, you'll have an app on your phone that could sequence your DNA in no time.

成本和时间的下降，让规模效应得以实现，从而能提供像我们刚才说的那种个性化治疗——每个人都能得到针对自己情况的精准治疗方案。因为说到底，这才是唯一的出路。

And that decrease in costs and the decrease in time required to do these things allows the economies of scale to rise to get specific treatments like the one that we're describing here, where every single person get their own hyper specific treatment, because at the end of the day, that is the only path forward.

想想传统的癌症治疗，都是广谱性的，千篇一律。最常用的化疗，就是把所有东西都清除掉，对所有人都一样。

Like when you think about traditional cancer treatments, they are all broad sweeping, it's one of a kind, and like the most popular one, chemotherapy, it just wipes everything out for everyone.

它不管你得的是什么类型的病。

It doesn't care what type of disease you have.

它就是直接去，把一切都消灭掉。

It just goes and it takes it out.

这可是件大事。

This is a really big deal.

他们说最快可能在2027年就能获批。

Approvals, they're saying could come as soon as 2027.

他们也在研究其他类型的癌症。

They're also working on other cancer types as well.

对吧？

Right?

对。

Yes.

那真是太厉害了。

That's, like, sick.

黑色素瘤是第一个。

Melanoma is, like, the first one.

黑色素瘤是最常见的皮肤癌，我记得每年有超过十万人被确诊。

Melanoma, most popular form of skin cancer, I believe the number is there's about a hundred thousand plus people that get diagnosed a year.

这些人中，有八千五百人会去世。

Of those people, eighty five hundred die.

是美国国内，还是全球范围？

In America or, like, in the world?

我觉得是全球范围。

I think this is the world.

哇。

Wow.

好的。

Okay.

哦，看起来人们受这种病困扰，而这是第一种类型。

Oh, it looks like people suffer from this, and this is the first type.

你可以想象，如果他们能攻克皮肤癌，那他们也能攻克很多其他疾病。

And you have to imagine if they could solve skin cancer, they could do it for many other things.

所以这真的非常令人兴奋。

So it's really, really exciting.

埃隆在这里有个评论，我们屏幕上显示着呢。

Elon has a comment here that we have on screen.

他说，尽管mRNA在新冠期间被明显误用，但它在治愈疾病方面有着巨大的潜力。

He says, despite its obvious misuse during COVID, mRNA has tremendous promise for curing disease.

人工RNA基本上把治愈疾病变成了一个软件问题，我们知道怎么处理——好吧，这是我自己说的。

Artificial RNA essentially makes curing diseases a software problem, and we know how we well, this is me saying it.

我们知道怎么解决软件问题，而且现在我们有了这些AI模型，它们能处理非常大的上下文窗口，所以可以把DNA序列放进一个单独的上下文窗口里。

We know how we take care of software problems, and now that we have these AI models that are capable of doing these, one, processing tremendously on context windows, so it can fit a DNA sequence inside of a singular context window.

但第二点，它们有推理能力和理解力，能真正利用这些信息并采取行动。

But two, having the inference ability and the understanding to actually make use of that info and do something with it.

对生物学来说，这真是一个难以置信的时代。

It's a really unbelievable time for biology.

如果你是计算机科学家，突然之间，生物领域看起来非常有趣，非常有吸引力。

And if you are a computer scientist, suddenly the world of bio is looking very interesting and very compelling.

如果你是AI公司，那么，向人们解释AI如何真正造福世界的最佳方式是什么呢？

And if you are an AI company, well, what's the best way to explain to people how AI can actually be beneficial for the world?

就是真正去做一些对世界有益的事情。

It's to actually go out and do something beneficial for the world.

所以我觉得，我们正处在这些事情以非常激动人心的方式发生的边缘。

So we're right on the cusp of, I think, lot of these things happening in a really exciting way.

这太棒了。

This is awesome.

好的。

Okay.

继续。

Moving on.

我们的朋友，或者说我们是这家叫Etched的公司的粉丝，它基本上是一家由两三个MIT辍学生创立的AI芯片初创公司，正在与NVIDIA正面交锋，现在又回来了。

Our friends, or rather we are the fans of this company Etched, which is basically an AI chip startup founded by two or three MIT dropouts, which is going head to head with NVIDIA is back.

你猜怎么着？

And guess what?

他们又筹集了几亿美元，这次是7亿美元，估值达到了惊人的210亿美元。但这并不是这次公告的头条新闻。

They have raised another couple $100,000,000, in this case, 700,000,000 and have reached an astounding valuation of $21,000,000,000 But that's not the headline from this announcement.

重点是我在屏幕上给你看的这最后一句话，那就是：我们也很高兴地宣布，我们已经把第一台定制的GPU机架交付给了Jane Street。

The headline is this last sentence over here that I'm showing you on screen, which is, we're also excited to share that we've shipped our first rack, custom GPU rack to Jane Street.

Jane Street基本上就是华尔街的一家量化基金公司。

Jane Street is basically a quant fund on Wall Street.

他们非常擅长做自己的事情。

They are very good at what they do.

而现在他们开始用这种定制的AI芯片来优化自己的AI交易算法，这确实是一个相当大的里程碑。

And the fact that they are now leveraging this custom AI silicon chip to have better AI trading algos is a pretty big milestone.

现在，简单回顾一下Etched是做什么的。

Now, just to kind of recap what Etched does.

说到NVIDIA，它就像是GPU领域的老大。

If you think about NVIDIA, NVIDIA is like the king of GPUs.

训练和推理AI模型都需要GPU。

GPUs are required to train and inference AI models.

他们的GPU确实很出色，但它们是通用型的，好处是能跑各种模型，不过在成本、功耗和效率方面也会有一些取舍。

Now their GPUs are very good, but their general purpose, which is good because you can run it on any kind of model, but there are certain trade offs that are made in terms of like cost, power, and efficiency.

Etched的团队看到这个问题后说，我们可以造一个更好的定制芯片，专门针对你正在构建的那个特定模型进行硬编码。

The folks at Etched looked at this problem and said, we can build a better custom chip that is hard coded to the specific model that you are building.

所以，专门针对你的OpenAI模型，专门针对你的Anthropic Claude模型，而且成本会比你原来花的少一半，甚至更多。

So specific for your OpenAI model, specific for your Anthropic Claude model, and it will be half or maybe even more cheaper than what you were spending already.

而且速度会飞快。

And it'll be lightning quick.

我们可以做很多优化。

There are certain optimizations that we can make.

不过，这是一个非常难的问题，工程量巨大。

Now, this is a very hard problem, a huge undertaking.

我们听了一些创始人的故事，他们被世界上最顶尖的GPU制造商拒绝过。

We listened to some of the founders' stories and they received nos from some of the smartest GPU makers in the world.

但尽管如此，他们埋头苦干，造出了这个定制芯片，结果发现他们还是交付最快的。

But despite that, they put their heads down and they built this custom chip, and it turns out that they're also the fastest to deliver.

通常，AI芯片从概念到交付需要三到五年时间。

Typically, AI chips take around three to five years to go from concept to delivery.

这些人，我觉得用了不到两年就做到了，真是让人震惊。

These guys have done it in, I think, just under two years, which is absolutely astounding.

而且我们马上就要看到它投入生产了。

And we're about to see it in production.

所以，Etched团队的这次发布真是令人印象深刻。

So just a very impressive announcement from the Etched team.

简直难以置信。

It's unbelievable.

我花了好多时间去看。

I've spent so much time looking.

我是说，我们俩都花了很多时间去看这些在AI领域涌现的公司，有成千上万的人，甚至几万人在解决有趣的问题，努力找到有趣的解决方案。

I mean, we both do spend so much time looking at all of these companies that are spawning in the world of AI, and there is thousands, if not tens of thousands of people working on interesting problems to solve interesting to reach interesting solutions.

而Etched非常不同，以至于我们上个月还专门录了一整集节目来聊这家奇怪的小创业公司。

And Etched is very different, so different that we actually recorded an entire episode about this weird random startup last month.

而且自从录完那期节目，Ejaaz，我不知道你知不知道，上个月它的估值翻了一倍。

And since recording that episode, Ejaaz, I don't know if you know this, the valuation has doubled in the last month.

大概四五周前我们聊到他们的时候，估值还是100亿美元。

When we talked about them just not even, like, four or five weeks ago, it was at a $10,000,000,000 valuation.

现在他们的估值已经达到210亿美元了。

Now they're trading at 21,000,000,000.

而且很明显，只要你听团队里的人聊起他们的产品，聊他们怎么走到这一步的，聊他们得出的结论，以及真正达到这个水平所需的研究，你就会发现这绝对是真材实料。

And it's very obvious once you listen to these people who are on the team talk about the product, how they've arrived there, the conclusions, the research that was required to actually reach this point, where it's very much the real deal.

这家公司是任何从事AI、半导体或硬件工作的人都想打造出来的。

This is a company that anyone who works in AI or semiconductors or hardware would want to make.

他们之所以没能做到，是因为这个团队解决的复杂性和难度太高了。

They were just unable to because of the complexity and the difficulty that this team has been able to solve.

结果就是，他们拥有了一款令人难以置信的产品，专门针对在自家芯片上运行大型语言模型Transformer架构进行了优化，而且他们的价值未来还会翻很多倍。

And as a result, they have this unbelievable product that is specifically tuned to run large language model transformer infrastructure right on their chips, and they are going to be worth many, many multiples more than they are today.

我想说，真的要恭喜他们，同时也要对所有听众说。

And, I mean, just a congratulations to them, but also for anybody who's listening.

如果你想看看真正高自主性、高智商创始团队是什么样子，去了解一下那两位联合创始人吧。

If you want to see what a badass team of high agency, high intelligence founders actually looks like, Read up about the co founders.

去了解一下他们的故事吧。

Read up about the story.

这是一家了不起的公司，未来一定会非常光明。

It is an unbelievable company that is destined for an incredibly bright future.

我很想

I would love

成为他们的投资者。

to be an investor in them.

我刚才想说，这对我和乔什来说有点私心。

I was about to say self serving for Josh and I.

如果听众里有人能接触到这个投资机会，请联系我们。

If anyone listening has access to this deal, please reach out.

我们非常愿意，拜托了。

We would love to Please.

让我们参与投资。

Be invested.

我对这些人特别着迷，很少有人能这么吸引我的注意力。

I am obsessed with these guys in a way that, like, not many people have captivated the attention.

真的很了不起。

It's pretty amazing.

你知道你能用这些芯片做什么吗，EJS？

And you know what you could do with all those chips, EJS?

你可以运行大量的东西。如果你在这些芯片上运行代理，并且你在为AI代理构建东西，你可能会想听听我们的赞助商Ledger。

You can run a tremendous amount of And if you are running agents on all of these chips and you're building for AI agents, chances are you may wanna hear about our sponsor, Ledger.

因为如果你在用AI代理构建东西，你很可能担心安全问题。

Because if you're building with AI agents, chances are you're worried about security.

我们在这一集前面聊过，这些东西有时候挺吓人的。

We spoke about this earlier in the episode, how kinda scary these things can get.

Ledger开发了一个叫AGENT STACK的东西，解决了这个问题。

Ledger has built this thing called the AGENT STACK that fixes this.

它提供了一系列开源工具，让你可以用三步流程来管理你的代理，确保一切顺利。

It gives you a series of open source tools that allow you to do a three step process with your agents to ensure everything goes well.

代理提出变更，人类批准，然后Ledger签名者执行。

The agent proposes a change, the humans approve, and then the Ledger signers enforce.

这项技术已经可用，并且与Claude Code兼容。

This technology is available and works with Claude Code.

它与Codex兼容。

It works with Codex.

它与Cursor或任何你使用AI的地方兼容。

It works with Cursor or any place where you work with AI.

所有这些都是开源的，今天就可以使用，链接在下面的描述中。

It's all open source and available today linked in the description below.

非常感谢Ledger赞助本集的这一部分。

Thank you so much to Ledger for sponsoring this part of the episode.

当我们从Etched继续往下聊，我们不得不提到另一个突破性的成功案例，那就是Unitree。

And as we move on from Etched, we have to talk about another breakout success, is Unitree.

Unitree，对于那些不知道的人来说，如果你没听过这个名字，你肯定看过那些小机器人的视频，它们简直太疯狂了。

Unitree, for those that don't know, if you haven't heard the name, you've seen the videos of these little robots that are, like, crazy.

它们在做后空翻。

They're doing backlips.

它们在田径赛场上绕圈跑。

They're running circles at, like, at these track meets.

它们非常酷。

They're very cool.

我觉得它们的IPO超额认购了多少倍来着，大概5000%？

I think their IPO was oversubscribed by what, like, 5000%?

实际上是5500%。

5,500, actually.

请说准确点，Josh。

Please get it right, Josh.

5500%，老兄。

5,500, man.

简直难以置信。

It's unbelievable.

结果呢，宇树科技的股票在周三早盘交易中暴涨了542%。

And then as a result, the Unitree robotic shares rose 542% in early trading on Wednesday.

这是不是炒作过头了？

Is it overhyped?

这算不算炒作过头？

Is this overhyped?

就是说，这合理吗？

Like, does this make sense?

这些数字也太离谱了。

These numbers are insane.

肯定有点炒作过头了。

Has to be a little overhyped.

听着，我不是中国股市的专家，但最近他们有几家IPO简直疯了一样，包括他们的存储芯片制造商长鑫存储，说实话，我觉得他们已经是全球第四大存储芯片制造商了。

Listen, I'm not an expert in the Chinese stock market, but they've had a few IPOs recently that have been absolutely insane, including their memory manufacturer, CXMT, which to be fair, they are, I think, the fourth largest memory manufacturer.

所以他们跟SK海力士、三星、美光这些公司是一个级别的。

So they're up there with like SK Hynix, Samsung, and Micron.

所以他们确实有正当的理由。

So they have a legitimate case.

但这些宇树机器人，我真不好说。

These Unitree robots, I don't know, man.

我看了看，其实我跟一些住在中国那边的朋友聊过，我就说，这太厉害了。

I was looking at I was speaking to some friends that are living out in China actually, and I was like, this is amazing.

你们肯定特别兴奋吧，这些公司上市，估值这么高。

Like, you guys must be super excited about, you know, these companies coming out and IPO ing at such crazy valuations.

结果在中国，大家对宇树机器人的看法就是，它就是个玩具制造商。

And apparently the sentiment in China around Unitree robotics specifically is it's just a toy maker.

他们其实并没有造出真正有价值的机器人。

They don't actually make robots that are actually valuable.

所以我多研究了一下，发现他们的人形机器人中，有40%到60%只卖给了研究机构，这意味着大部分并没有实际用途，不像他们宣传的那样，比如在工厂里做制造，或者自动化仓库操作之类的。

And so I dug into this a little bit more and apparently of the humanoid robots that they have, 40 to 60% have only been sold for research purposes specifically, which means that the majority aren't actually being used for practical purposes, like some of the use cases which they advertise, such as manufacturing at factories or automating warehouse facilitations and stuff like that.

实际上它们只是在测试阶段，这并不能带来稳定的收入，也不能记在账上。

They're actually just being tested, which isn't really confident or recurring revenue that they can put on their books.

所以如果你把这个映射到，这是什么，上市首日涨了600%，超额认购5500倍，也许市场动态有点虚高，也许这会被打回原形。

So if you map that onto, what is this, a 600% price increase on day one with a 5,500 X oversubscription on the IPO, maybe the market dynamics are a little inflated and maybe this will crush back down.

我们一个月后再看吧，但看到中国股票，特别是机器人股，现在真的开始有势头了，真是疯狂。

Let's revisit in a month, but it is crazy to see that Chinese stocks, specifically robotics, are really starting to gain some momentum now.

对。

Yeah.

这确实是个很酷的新故事。

It's it's a very cool new story.

我个人是不会碰这个的，躲得远远的。

I personally am not touching this with a tenfold poll.

那看起来真的很吓人。

That seems really scary.

我喜欢看那些视频。

I love watching the videos.

我会继续看那些视频的。

I'll continue to watch the videos.