🔬 未来的实验室应该像数据中心一样——安迪·比姆与拉法·戈麦斯-邦巴雷利，Lila Sciences

互联网上。

the Internet.

对。

Yeah.

对。

Yeah.

互联网上的某个地方。

Somewhere in the Internet.

互联网就在里面。

The Internet's in there.

这个我忘不了。

I can't forget that.

所以我们看到了很多类似的有趣的性格怪癖，这些都是强化学习带来的结果。

So we've seen lots of, like, funny, like, personality quirks like that as a function of RL.

有一些明显的强化学习问题，我不会称之为奖励黑客，但确实有病理现象，比如重复。

There's, like, obvious RL I wouldn't call them reward hacking, but pathologies, like repetition.

所以，思维链会崩溃，然后它就会一遍又一遍地重复它的最终答案。

So, like, the chain of thought will collapse, and it will just, like, repeat its final answer over and over and over again.

出于某种原因，这有时确实能带来更高的奖励。

For some reason, that reliably sometimes leads to higher rewards.

我们也不确定为什么病态的思维链或不可读的思维链在某些情况下会带来更高的奖励。

Like, we're not sure exactly why pathological chain of thoughts or nonlegible chain of thoughts in some cases lead to higher rewards.

抱歉。

So sorry.

我想打断一下。

I wanna interrupt.

我们刚才在讨论强化学习。

So we're talking about RL.

对。

Yeah.

你刚才说的听起来就像是在思维链上做强化学习，就像大家都在做的那样。

We're talking about the way way you're talking about it sounds like just RL on chain of thought, just like everybody's doing.

对。

Yeah.

但你的强化学习实际上有一个实验室步骤。

But your RL actually has a lab step.

对。

Yeah.

如果陷入病态循环，那是不是意味着实验室只是在反复做同一个实验？

If you're in a pathological loop, does that mean the lab is just, like, doing the same experiment over and over again?

不只是思维链，退一步说，思维链是模型用来解决问题的令牌。

Just not so a chain of thought, maybe just to to step back, is tokens that the model uses to solve a problem.

比如解数学题，你会用定理一、定理二、推论、引理。

So if you were solving a math problem, you would do theorem one, theorem two, corollary lemma.

你知道，你会把问题分解开。

You know, you you decompose the problem.

在科学领域，思维链也有类似的部分。

In science, the chain of thought, there's some of that too.

所以这里面是有推理过程的。

So there's reasoning that happens.

比如说，我想针对某个靶点做一个抗体。

You know, I'm trying to make a a an antibody for this target.

我对这个靶点了解多少？

What do I know about this target?

已知的表位有哪些？

What are the known epitopes?

我的进攻计划是什么？

Like, what's my plan of attack?

在思维链里，也会有工具调用。

In the chain of thought are also tool calls.

所以，我可能会用结构预测模型，来看看这个序列在三维空间里是怎么折叠的。

So I'm maybe I'm gonna use a structured prediction model in this case to get some read of how the sequence folds in three-dimensional space.

所以，工具调用是思维链的一部分。

So tool calls are part of the chain of thought.

不过在Lila，有趣的是，实验室仪器也是工具调用，或者是一系列工具调用组成的工作流，但这一切都是人类可读的。

At Lila, though, the the fun thing is that the lab instruments are also tool calls or a series of tool calls for a workflow or a but it's all human legible.

全部都是英文。

It's all in English.

对。

Right.

所以我们看到的一些问题就是，它直接跳过了中间部分，我们认为这部分对解决问题很重要，它直接给出答案，说这种情况下不需要做实验。

And so what some of the pathologies we've seen is it just skips all the middle part, which we would think is important for solving a problem and just goes right to the answer and says, don't need to do an experiment in this case.

我不需要调用工具。

I don't need to call a tool.

而且，在某些情况下，比如我们已经做过实验或者类似的情况，我们能判断一些事情，有时候这其实并不是一个糟糕的策略。

And, you know, in some cases where we can judge something because maybe we've already done the experiment or something like that, for some reason, it is actually not a bad strategy in some cases.

所以这里有些神秘之处。

And so there's some some mystery there.

它是个理论家。

It's a theorist.

对。

Yeah.

对。

Yeah.

它已经完成了计算。

It's done the calculation.

这可能是题外话了，但，它实际上是在潜在空间中思考，它发出的是令牌。

And this is probably too much of a tangent, but, like, it actually thinks in latent space, it emits tokens.

所以，思维链往往是一个不可靠的叙述者，不能真实反映模型实际的计算过程。

So, like, the chain of thought is often an unreliable narrator for what the model the computation of the model is actually doing.

所以，我们正在思考的一个大问题是，当我们开始处理一个问题时，比如拉法提到的电催化剂，我们实际上不知道什么是对的，什么是错的。

And so one of the big things we're trying to think about is when we're moving into working on a problem, you know, like Rafa said, for electrocatalysts, that we actually don't know what right and wrong looks like.

我们应该在多大程度上依赖思维链，而不是直接相信实验、验证器或模拟器作为最终的基准真相？

How much should we rely on the chain of thought versus just trusting the experiment, trusting the verifier, trusting the simulator as the the ultimate ground truth?

所以，你知道，Lila 不是一家生物技术公司。

So, you know, Lila is not a biotech company.

我觉得Lila在这方面其实挺特别的。

Lila is actually fairly unique, I think, in this way.

你知道，我参与过不少生物科技公司。

You know, I've I've been involved with biotechs.

我也帮忙创办过一些生物科技公司。

I've helped start biotechs.

通常，目标就是赶紧推进到临床试验。

Often, the goal is to sprint to a clinical trial.

对。

Yeah.

所以你要开发一个资产。

So you want to develop an asset.

你开发一个平台，是为了让你有选择进入哪个领域的灵活性。

You develop a platform in service of having optionality of what space you move into.

但一旦你有了临床资产，就把所有东西都进入医学上的昏迷状态，然后你通过临床试验，如果顺利的话，其他事情才能继续。所以我们把这个选项拿掉了。

But once you do the you have the clinical asset, put everything into a medically induced coma, and you you get through the clinical trial, and if it goes well, then other things get to so we are taking that option off the table.

在Lila，模型本身就是最有价值的东西。

The model itself is the thing of value at Lila.

所以从这个角度看，我们更像是一个新型实验室，试图找到新的方法来推动核心推理大语言模型的能力。

So in that sense, we're much more of like a neo lab trying to think of a new way to push forward capabilities of a core reasoning LLM based model.

连实验室平台都不算吗？

Not even the lab platform?

实验室平台就是生成数据的工具。

So the lab platform is the token generator.

好的。

Okay.

这个平台就是数据生成机制，最终也是Lila的护城河。随着它不断扩展，我们能在单位时间和单位面积内生成的数据量都会增加，这些数据反馈给模型，让它变得更聪明，然后模型再建议下一步该做什么实验。

That is the data generation mechanism that ultimately is the moat for Lila, is that once that continues to scale, the amount of data that we can generate both per unit time, but per unit square foot will go up, and that feeds back into the model to make it smarter, that then suggests the next experiment to do.

所以我们真正专注的是让这个核心模型尽可能高效和聪明。

And so we really are focused on making this core model as performant and smart as possible.

我们可以聊聊这如何适用于不同的商业策略，但归根结底，

And we can talk about how that lends itself to different commercial strategies, but ultimately,

我们感兴趣的是创造这种新型的AI模型。

we're interested in creating this new type of AI model.

所以我想引用Octant Bio的Sri Kosuri，他发了一条很棒的推文。

So I wanna quote Sri Kosuri from Octant Bio, who had a great tweet.

几周前我真的很喜欢那条推文。

I really loved, you know, a few weeks ago.

内容是：机器学习在药物发现中的商业模式是什么？

It was, what is the business model in ML for drug discovery?

因为如果你需要数据来训练模型，但如果你已经有了数据，那还需要模型做什么？

Because if you need the data to train the model, but if you have the data, what do you need the model for?

当范围很窄的时候，确实是这样。

That is true when you are narrowly scoped.

在某个特定的科学领域内，确实如此。

So that is true within a given vertical of science.

我举个例子，如果你回到十年前，想创建一个编程助手模型，你只会得到编程数据。

The analogy that I would use is, like, if you went back ten years and you tried to create, like, a coding assistant model, you would just get coding data.

你也不会得到莎士比亚的诗歌、卡尼塔的食谱。

You wouldn't also get Shakespeare poetry, Carnita's recipes.

只是事实证明，当模型能够在更广泛的数据和更深层次的数据上进行训练时，会有溢出效应。

It it just it turns out that there is spillover as the model is able to train on a broader swath of data and a a deeper cut of data.

所以，再说一次，我们打的核心赌注是，这对科学领域也是成立的。

And so, again, the core bet that we're making is that is true for science.

如果模型在越来越广泛的数据集上训练，那么在特定领域所需的数据量就会减少。

That if the model is trained increasingly broad set of data, the amount of data that you need in a given domain, that data requirement is reduced.

在某些情况下，如果数据与模型之前见过的数据相近，需求甚至会降到零。

In some cases, will be reduced to zero if it's adjacent to what the model has already seen before.

所以有一个数据效率的论点，表明拥有一个能够生成广泛科学数据的通用平台是有价值的。

And so there's a data efficiency argument that would suggest that, again, having a general platform that can create a broad swath of scientific data.

我还想提一下，显然，我们使用的是已经商品化的东西。

I'll also just mention that, like, obviously, we are using things that are already commodities.

所以我们使用公共数据集、模拟器，而实验平台是对这些现有商品化资源的补充。

So so public datasets we use, simulators we use, and experimental platform is a complement to these existing commodity resources.

这让我想到一个问题，就是关于适用域的概念。不同的尺度会产生完全不同的信息类型，以及实体之间完全不同的关系。

This brings up a question in my mind about there's a concept of applicability domain where you have different scales, and they result in different types of completely different types of information and relationships between entities.

对吧？

Right?

所以有量子领域、化学领域，还有各种生物领域。

So you have the the quantum realm, you have chemical realm, you have, you know, sort of different bio realms.

我对跨领域方法的一个担忧是，这些领域之间到底有没有迁移。

One concern I would have with cross cutting approaches is there domain transfer between these at all.

而像 carnitas 食谱和胸痛问题，它们的共同点在于都是用语言写成的。

Whereas, you know, carnitas recipes and, you know, chest problems have the commonality that they're written in language.

但这里你几乎有一个完全独立的，甚至不是语言的东西。

Whereas you almost have a completely separate, not even language.

对吧？

Right?

这些领域之间是完全不同的模型。

It's a completely separate model between these domains.

人类科学家的工作

Human scientists work

涵盖了所有这些领域。

on all of those domains.

没错。

Correct.

而且他们之间大多通过书面语言交流，使用

And they mostly communicate with each other in written language using

好的。

Okay.

使用工具。

Using tools.

所以可以说，存在一个共同的推理过程，让人能够在每个领域中解决问题。

So would say there's a there's a common reasoning process that allows someone to solve problems in each one of those domains.

所以我认为这种逻辑也适用于我们正在训练的推理模型，它同样使用工具，能做数学，能写代码，但所有知识都存储在一个地方。

And so I think that that logic carries over to a reasoning model that we're training that, again, uses tools, can do math, can do code, but it's it's it's having all of that knowledge stored in one place.

对我来说，一个经典的跨领域例子就是，嗯，

One classic example for me of domain transfer Mhmm.

在复杂性理论和量子引力之间。

Is between complexity theory and quantum gravity.

嗯。

Mhmm.

对吧？

Right?

现在很多量子引力理论基本上都认识到，这两者背后的数学是完全一样的。

Where now a lot of the quantum gravity theories are are basically recognizing the the the identical math behind the two of them.

对吧？

Right?

你有没有遇到过那种，哎呀，这个领域居然能应用到那个领域的情况？

Do you have examples of this kind of sort of, oh, man, this domain actually applies to this domain?

所以我们构建了一个包含十万亿科学标记的推理数据集，这些推理轨迹在生命科学、化学和材料科学领域都经过了实验验证。

So we have assembled this reasoning dataset of 10,000,000,000,000 scientific tokens, reasoning traces that are experimentally verified across life sciences, chemistry, and material sciences.

我们发现这个通用模型往往比特定领域的模型表现更好。

And we have seen that this general model often beats the domain specific models.

所以很难指出模型里到底是什么让它做到这一点，它发现了哪些联系。

And so it's it's hard to point to what's in the model that is making it that it what connections it has realized.

但很明显，看过所有科学领域的更多数据，在样本对样本的比较中，确实胜过特定领域的推理模型。

But clearly, having seen more data across all of science beats sort of in a sample for sample kind of way domain specific reasoning models.

科学的未来就是语言。

And the future of science is language.

对。

Yeah.

所以我觉得未来

So I Future

化学的未来是

of chemistry is

是的。

the Yeah.

我觉得没必要。

I don't think it's necessary.

我觉得那不是科学超级智能的必要条件。

I don't think that's a necessary condition for scientific superintelligence.

我是说，上周有个说法，对吧？

I mean, there was this quote from, what, last week?

就是说，可能不值得把所有蛋白质或三维结构之类的数据形式都蒸馏成语言，因为有些数据模态和语言差别太大了。安迪总跟我说，拉法，英语是图灵完备的，所以你可以表达一切。

That it might not be worth distilling all the ways that live in sort of, know, protein, Or teralpha form, there are data modalities that are so different from language, and Andy always tells me, Well, Rafa, English is Turing complete, so you could express everything

它确实有好几种图灵完备的语言，但没错，是的。

It had in been several Turing complete languages, but yes, yes, is.

我同意。

I agree with that.

有些地方可能更高效，因为数据本身的特性，我是说，你做过几何深度学习。

There might be places where it's more efficient because of the nature of the, I mean, you've done geometric

深度学习。

deep I learning,

想想几何学，也许，你知道，我给我的同事泰德·史密斯打电话，我觉得几何学就是那种人们觉得问题本身的性质可能更适合其他架构的领域之一。

think for geometry, and maybe, you know, I call my colleague, Ted Smith, I think geometry is one of those places where people feel that there might be sort of, just the nature of the problem is more amenable to other architectures.

所以如果我们需要调用蛋白质折叠模型，或者需要调用等变扩散模型来生成晶体结构，那都是可以的。

So if we need to call a protein folding model, or we need to call an equivariant diffusion model to make crystal structures, that's fair game.

所以我会说，未来科学与我们交流的方式，肯定是语言。

So I would say, the future of the way science talks with us, for sure, is through language.

模型是否需要一直用英语思考化学，也许需要，也许不需要。

That the model needs to be thinking in English about chemistry all the time, maybe yes, maybe not.

化学家并不是用英语思考化学的。

Chemists don't think about chemistry in English.

好的，对。

Okay, yeah.

有道理。

Fair.

他们是用英语讨论化学的。

They talk about it in English.

对，我完全同意拉法说的所有内容。

Yeah, And I I agree with everything Rafa said.

比如，基于令牌的推理加上工具使用非常强大，我觉得我们想说的是，在科学领域，我们才刚刚开始探索这方面的潜力。

Like, token based reasoning with tool use is very powerful, and I think the claim that we're making is we have barely scratched the surface for that in science.

我们并不是想把特定领域的模型压缩成一个推理模型。

We're not trying to distill domain specific models into a reasoning model.

它可以有效地利用这些工具。

It can use those tools productively.

所以，把经常用英语进行的推理，还有用Python等进行的推理，和工具使用结合起来，确实非常强大，而在科学领域，我们才刚刚起步，还不知道能把它推进到多远。

And so it's just the combination of reasoning often in English, but also in Python and things like that combined with tool use is very powerful, and we're very early in science and understanding how far we can push that forward.

明白了。

I see.

那你能举一些你们正在进行的、有代表性的活动例子吗？

So can you give some examples of campaigns that you are running that are are representative?

实际上，在你回答之前，你能先确认一下吗？

And actually, before you do that, can you please check back?

我意识到我们还没解释清楚，你不只是做生物技术。

I realize we still haven't explained that you don't just do bio.

对。

Right.

只是你的技术背景。

Just your tech bio.

所以我觉得这正好是个很好的引子。

So I think this is a Yeah.

很好的引子。

Great great lead into this.

对。

Yeah.

所以不只是技术背景，你在科学方面做什么？

So not just tech bio, what do you do in terms of science?

科学？

Science?

不是。

No.

好的。

Okay.

所以也许是这样。

So maybe yeah.

我们训练模型的方式，其实涵盖了整个生命科学领域，包括DNA、RNA、蛋白质、细胞、小分子、各种化学物质，以及不同类型的材料。

So the the way that we train the model is is I mean, it's across life sciences, so DNA, RNA, protein cells, small molecules, different kinds of chemistries, and different types of materials.

所以这就是我们目前的范围，确实挺广的。

So that is, like, where we're scoped now, which is admittedly a large

材料本身的范围也不小，跟我们在生物领域看到的那些差不多广。

Materials itself is also not just that's also as largely scoped as everything we've seen in the bio credit.

我们可以举一些例子。

We can give some examples.

比如现在，我们可以制造薄膜。

So today, we can make thin films.

我们可以做粉末。

We can make powders.

我们可以做量子点，我们有高精度的量子点。

We can make quantum dots, we have acute quantum dots.

你们听说过吗？

Are you folks familiar?

量子点是某些电视里用的发光技术，你需要精确控制，让它们的大小完全一致，达到纳米级别，而纳米尺寸决定了它们发什么颜色，你需要最纯的红、最纯的蓝和最纯的绿，才能让电视的色彩又锐利又丰富。

Quantum dots are the luminescent technology in some TVs, and you need to control to make them of exactly the same nanometer size, and the nanometer size you make them controls what color they're going to be, and you need the purest red and the purest blue and the purest green to make really sharp and rich color palettes for your TV.

而且你需要让它们非常均匀，全都一样，不然颜色又会混在一起。

And you need to make them as homogeneous, they all need to be the same, otherwise the color gets again, blended.

我们有个很酷的演示，我们的自动驾驶实验室，会请访客选一个波长，就是他们希望量子点是什么颜色，当他们来办公室的时候。

So we have a cute demo, where our self driving lab, we ask our visitors to pick a wavelength, what color do want the quantum dot to be when they come into the office?

然后我们启动机器，模型会推理，有时候我们甚至会加入模型从未见过的新化学品，就为了看看它会怎么反应。

And then we fire off the machine, the model reasons, even sometimes we even throw in new chemicals that the model had never seen, just to see how it moves.

机器在运行，等到参观结束，大概一个小时到一个半小时，机器可能已经做出了几代量子点，而且通常都能达到目标，不然我们也不会这么做，对吧？

The machine is running, and by the end of this sort of hour, hour and a half tour, the machine has made maybe one, maybe more generations of quantum dots, that tend to hit, otherwise we wouldn't do it, right?

它们往往能达到人们建议的颜色。

They tend to hit the color that people suggested.

所以我们有能力制造很多种材料。

So we have the ability to make lots of materials.

我们可以配制液体、聚合物和软物质。

We can formulate liquids and polymers and soft matter.

我们非常关注能源和可持续性，所以我们在电化学方面有很强的能力，涉及化学转化与作为可再生能源的电之间的相互作用。

We care about energy and sustainability a lot, So we have a good chunk of electrochemistry capabilities, about the interplay of chemical transformations, and electricity as a renewable energy source.

我们关注传统催化，也关注材料的力学性能。

We care about traditional catalysis, and we care about mechanical properties of materials.

所有这些都汇聚在项目中，我们制造催化剂，制造用于腐蚀防护的高性能涂层，或者用于航空航天高性能机械应用的材料。

And all this comes together in programs where we make catalysts, we make high performance coatings for corrosion, or aerospace high performance mechanical applications.

在过去的几周里，我们与外部合作伙伴一起，启动了多个我们以前没有做过的项目，涉及从粘合剂到冷却液等多个领域。

And over the last few weeks with an external partner, we started multiple sprints of things we weren't doing before, that touch, from adhesives to cooling fluids.

所以我们越来越能够在开放式的化学和材料科学领域快速产生令人兴奋的发现。

So we've been able to sort of more and more spin up just exciting discoveries in sort of open ended chemistry and material science spaces.

量子点和蛋白质设计之间有什么联系吗？

Do you have any connection between quantum dots and, let's say, protein design?

它们是同一个平台，

It's the same platform that

都能做这个。

does that.

它们用的是同一套能力。

It's the same set of capabilities.

所以有一个共享的基础设施，让我们能在同一个屋檐下做所有这些事情。

And so there's a shared infrastructure that lets us do all of those things under the same roof.

如果没有这种联系，我们就没有能力做所有这些事情。

If there were no connective tissue, then our ability to do we just would not have the ability to do all those things.

我们有没有做过像MEKINTERP那样的分析，就是深入模型内部，看看电催化剂方面的见解是否对其他领域有启发？我们还没有对MEKINTERP做深入分析。

Have we done, like, the MEKINTERP thing where we look inside the model and see, does this insight from electrocatalyst inform we haven't done a deep dive on on the MEKINTERP thing.

我们确实看到，我们执行这些项目的能力变得更快了，是的。

We have seen that our ability to to do these programs has gotten faster Mhmm.

随着平台越来越成熟。

As the platform has become more mature.

那么，LMP是不是就是你们工具包里的一个很常见的东西？是不是因为有了它，你刚才提到的那些想法里，很大一部分都能实现？

So is LMP just like a common thing along like, is this something really common in your toolkit that because of this, this enables, like, a large fraction of these ideas that you've just mentioned?

在生物方面当然说得通，但我对生物的了解比材料多。

It certainly makes sense on the bio side, but I know bio much more than materials.

这是不是你们实验室工具包里的一个共同主题？

Is like a common theme amongst your lab toolkit?

AI科学工厂的能力越强，我们就能越快地去追求新的目标产品特性，以及新的令人兴奋的机会。

The AI science factory, the more capabilities it has, the faster we've been able to go after new target product profiles and about new exciting opportunities.

因为模型能做的事情更多了，实验室能做的事情也更多了，我们的科学家在整合新能力方面也变得更灵活、更迅速。

Because the model is prepared to do more things, the lab can do more things, our scientists are more flexible and faster in order to incorporate new capabilities.

我们拥有的仪器越多，添加新仪器的速度就越快。

Adding new instruments has become faster, the more instruments we have.

所以，这可能跟我们是哪种类型的公司有关，这里有一种超大规模循环，软件上的这种扩展，是由硬件上的扩展来支撑的。

So, maybe it goes with sort of what type of company we are, sort of the, there's cycles of hyperscaling here, of this sort of scaling in software, is backed by scaling in hardware.

而且我们有数万平方英尺的实验室即将投入使用，配备几十到上百台仪器，这给了我们快速行动的广度。

And the fact that we have sort of tens of thousands of square feet of lab coming online with sort of dozens to hundreds of instruments, is giving us this breadth to move fast.

有些领域，比如我同事Heather Kulik最近在播客里提到的，我们正在看到的一个领域是脱附，我想听众对这些材料应该很熟悉，我就不多介绍了。

There are places, folks had my colleague Heather Kulik here in the podcast recently, one of the areas we're seeing, desorption, so I think the audience will be familiar with these materials, I don't need to spend a lot of time introducing them.

这些材料是由分子与金属的相互作用构成的。

These materials are made of the interaction of a molecule with a metal.

结果发现，我们的模型是在小分子药物发现上训练的，它们学到的所有关于药物发现的化学知识，都迁移到了这些金属有机框架材料上，我们可以用这些材料从空气中捕获二氧化碳或过滤氨气。

And it turns out our models had been trained on small molecule drug discovery, and all of the chemistry that they had learned thinking about drug discovery, carried over to start reasoning over this metal organic framework materials that we can use to take CO2 out of the air or to filter ammonia.

我觉得这太迷人了。

I find that fascinating.

我经常看到人们做机器学习，他们在大数据集上训练，然后转向新的领域。

Like, when I think so many times I've seen people work on machine learning where they train some big dataset and then they move to some new domain.

而且很多时候，迁移的程度很小。

And oftentimes, the amount of transfer you see is small.

是的。

Yes.

是不是有一些，怎么说呢，你们常用的基础组合，就像三原色那样，经常用来做实验？

Are there, like, a group of, I don't know how to say this, primary colors that you have that you combine together that oftentimes result in your experiments?

在生物方面，那些就是很明显的候选。

So on biology, they'd be the obvious candidates.

核酸能力、无细胞表达，还有下游的检测，都是我们关心的。

Nucleic acid competence, cell free expression, and then downstream assays are things that we care about.

所以这些核心能力，我们可以用来组合出很多不同的事情。

So there are core competencies that we can then, you know, sort of give rise to a factorial number of different things that you can do.

在材料方面呢。

And on the material side.

我觉得是配方。

I think formulation.

对。

Yeah.

而且这其实不是，我不想说它很普通，但它太常见、太重要了，所以一开始那些特别聪明的地方，想的都是更炫酷的东西，结果发现很多工业界和世界各地的人其实都很关心配方，也就是把液体和黏糊糊的东西混在一起，做出其他黏糊糊的东西。

And it wasn't even it's so I don't I don't want to say mundane, but it's so common, it's so important that it wasn't one of the first sort of superintelligent places, thought of flashier things back at the beginning, and it turns out a lot of people, you know, in industry and in the rest of the world care about formulation, meaning mixing liquids and gooey things to make other gooey things.

但那就是润滑剂，那就是睡眠纳米颗粒，那就是除臭剂。

But that's lubricant, that's sleeping nanoparticles, that's deodorant.

就像消费品、工业产品和医药里都有这些东西，你知道，从凝胶到皮肤移植，所有这些都源于这种混合黏糊糊材料的过程。

Like there is all these things in consumer products and industrial products and in medicine, you know, gels to skin grafts, all those things emerge from this sort of mixing gooey materials.

这是我们正在打造的一项核心能力。

And that's a muscle that we're building.

这是一个非常常见的平台，我们跟人聊得越多，它就越频繁地出现。

That's a very common platform that is showing up all the time the more we talk with people.

有意思。

Interesting.

我有一个经验法则，在考虑规模化时经常用到，那就是每当你把系统扩大一个数量级，你面临的问题就会完全改变。

I have a rule of thumb that I often use when I'm thinking about scaling, which is that every time you scale an order of magnitude in a system, that your set of problems completely changes.

你们选了最难的两个问题，对吧？

You guys pick the two hardest problems, right?

材料和生物，你们也做其他事情，但材料和生物是出了名的难以推向市场。

Materials and bio, you do other stuff, but materials and bio are notoriously difficult to get to market.

对吧？

Right?

就是那种十年、十五年的长期规划。

The it you know, ten year, fifteen year time horizons.

尤其是材料规模化，原因就在这。

And the reasons are especially for materials scaling.

那你们是怎么考虑这个问题的？

So how are you thinking about that?

你们是只做发现阶段，还是说会一直做到产业化？

Are you just saying we're discovery or are you saying that we'll get to it?

就是，你们是怎么想的？

Like, what how are you thinking about it?

我最后一次做学术讲座前准备的讲稿，题目是《材料与化学规模化的苦乐参半的教训》，因为情况就是这样。

The last academic lecture I prepared before I stopped giving academic lectures called the bitter sweet lesson of scaling in materials and chemistry, because it's this.

事实证明，在AI领域，规模化是好事，因为它给你指明了前进的方向。

It turns out, in AI scaling is a good thing because it gives you a roadmap of what you need to do.

而在化学和材料领域，规模化是个让人头疼的问题，因为事实证明，只有那些能规模化的东西才重要。

And in chemistry and materials, scaling is a spooky thing because it turns out only the things that you can scale matter.

所以我们非常清楚这一点，对吧？

So, we're extremely cognizant, right?

我们的产品团队、实验室团队，大家都明白。

Our product team, our lab team, we all know.

比如在量子点这个例子里，我们一直能用同一个配方，从几毫升一直做到差不多一升。

For instance, in the quantum dot example, we've been able to use the same recipe from single digit milliliters to a 100, almost a liter.

所以，有些地方我们目前的能力已经开始在规模化和技术成熟度上发挥作用了。

So, are places where our capability today takes bites into scaling and into technology readiness level.

然后我们也在构建系统，让它们在现在做实验的时候就能思考以后什么会变得重要。

Then we are making the system such that they can reason about what's going to matter later as they're doing the experiments now.

这也会用在我们那些不含稀有金属，或者说不含铂族金属的催化剂上，对吧？

And this will be in our rarer free, or sort of platinum group free catalysts, right?

问题的本质恰恰在于，我们需要能够把这些东西规模化，对吧？

Precisely the nature of the question is that we need to be able to scale these, right?

所以这是有供应链意识的，当我们启动第一个实验时，我们已经读完了所有论文，我们已经有一个技术经济分析代理在角落里待命，准备对我们做的任何事情进行技术经济分析。

So it's supply chain conscious, as we're firing off the first experiment, we've already read every paper, we've already have a techno economic analysis agent sitting on the corner, ready to do the techno economics of anything we do.

归根结底，我们不会做临床试验，我们不会为一个特定的工艺建立中试工厂，然后把它放到你的炼油厂里。

At the end of the day, we're not going to do clinical trials, we're not going to make pilot plants for one particular process that you will put in your refinery.

在那个时候，这些地方我们会和客户合作，或者如果我们发现了一些非常惊人的东西，甚至不需要任何指导，我们就直接去卖。

At that point, these are places where we will work with our customers, or if we find something so amazing that we don't even need any instruction, and we just go sell it.

但通常，我们会交接，就像我们要支持客户的疗法发现一样，我们要在客户的痛点处支持材料创新。

But typically, we will hand off, just like we're going to support therapeutic discoveries for our customers, we're going to support materials innovations at the pain points that our customers have.

那些也跟规模化有关。

And those also have to do with scaling.

你们进展到哪一步了？

How far have you gotten

到目前为止？

so far?

在生命科学方面，我想我同意拉法说的所有观点。

So on the life sciences, and I think I agree with everything Rafa said there.

要理解人们会怎么用这个平台，或者说我们到底在做什么，其实更像是给科学领域做一个类似Claude Code的东西。

The way to think about how people would use the platform is, or just kind of like what we're building is much more of like a Claude Kodish kind of thing for science.

所以，吸引早期客户的一个原因就是，我们不是一家做体内CAR T的公司。

So one of the things that has drawn early customers to us is so we're we're not an in vivo CAR T company.

现在做体内CAR T的公司很多。

There's lots of in vivo CAR T companies.

这个领域现在特别火。

Super hot right now.

我们确实看到，六个月前Capstan被收购，价格超过20亿美元。

We did see, you know, six months ago with the Capstan acquisition for, like, 2 plus billion dollars.

如果大家不太了解这个，体内CAR T是一种很新的治疗方式，以前主要针对血液癌症，现在越来越多用于自身免疫疾病。

If folks aren't familiar with that, like, in vivo CAR T is this very new heart therapeutic modality previously for blood cancers, but now increasingly for autoimmune disease.

我们内部确实具备做体内CAR T所需的三项核心能力。

We did have in the internal, like, sort of triumvirate of capabilities that you would need to do in vivo CAR T.

也就是结合剂设计，这个我们当然能做，还有LNP配方，以及mRNA设计。

So binder design, obviously, we we can do that, LNP formulation, and then mRNA design.

让大家知道CAR T是什么

Just so people know what CAR T

因为它真的很酷。不，

is, because it's really freaking No.

我喜欢这个。

I love it.

太酷了。

It's so cool.

我能说说CAR T吗？

Can I can I talk about CAR T?

可以。

Yeah.

对。

Yeah.

我们来聊聊CAR T吧。

Let's talk about CAR T.

CAR T 从八十年代末或九十年代就开始研究了。

So CAR T has been worked on since, like, the late eighties or nineties.

大概在2010年左右，它在癌症治疗上真正火了起来。

Really caught fire around 2010 or so for for cancers.

以前的做法是，先提取一些T细胞。

The way it used to work is you'd extract some of those T cells.

然后你会改造出一种叫做嵌合抗原受体的东西，放在T细胞表面，告诉它该去攻击哪种细胞。

You would engineer what's called a chimeric antigen receptor that goes on top of that that tells the T cell to what go in to go what kind of cell to go and kill.

所以你基本上是在改造患者的T细胞。

So you're you're basically modifying people's T cells.

对。

Mhmm.

你把它们

You take

取出来。

them out.

你把它改造一下，让它的表面带上这种特殊的抗原受体。

You modify it so that it has this weird antigen receptor on its surface.

就是去搜寻并消灭它们，对。

To seek and destroy Yeah.

标记。

Tag.

通常他们会用一种叫CD19的蛋白质，这种蛋白质主要在B细胞上表达。

Usually, they use a protein called c d 19, which is preferentially expressed on B cells.

当B细胞发生癌变时，就会引发血液癌症。

When B cells get malignant, they create blood cancers.

还会引发自身免疫疾病。

They create autoimmune diseases.

这样做几乎会清除掉一个人所有的B细胞。

You you wipe out someone's almost their entire B cell repertoire when you do this.

这会造成很大的附带损伤。

There's a lot of collateral damage.

但本质上，你是在告诉T细胞该去杀什么。

But, essentially, you're telling the c T cell what to go and kill.

所以这项技术大概在2015年左右开始火起来。

So this really started to catch fire around 2015.

它既贵又慢。

It was expensive and slow.

你得提取病人的T细胞。

You have to extract someone's T cells.

还得对它们进行改造。

You have to engineer them.

每次输注大约要40万美元，而且它仍然是治疗多种癌症的奇迹疗法。

It was like a $400,000 per infusion and just like it's still a miracle cure for lots of different types of cancer.

这有点跑题了，但有个叫艾米丽·怀特黑德的孩子，在费城儿童医院接受过治疗。

Too much of a tangent for this, but there's this child named Emily Whitehead who was treated at the Children's Hospital of Pennsylvania CHOP.

她是CAR-T疗法在儿童癌症中最早治愈的病例之一。

She was one of the first cures in pediatric cancer by CAR T.

她本来要被送去临终关怀的，结果接受了CAR T治疗。

She was gonna be referred to hospice care, got CAR T.

又跑题了，聊到心理方面去了。

Another psych psych tangent.

她最后还是因为CAR T治疗初期的发烧去世了。

She always died of a fever from this initial CAR T treatment.

那确实很难。

That's difficult.

她能活下来，完全是因为她的主治医生有个女儿得了儿童关节炎，知道那种特定的抗体能抑制她对CAR T的IL-6反应。

The only reason she survived is because the doctor who was treating her had a daughter with pediatric arthritis and knew that the this specific antibody would blunt her IL-six response to CAR T.

所以，这里面有很多值得深思的地方，关于AI在科学中的应用，以及那个案例中所有的巧合，才能让一切顺利进行。

So, like, there's a lot to unpack there in terms of AI for science, all the serendipity that had to happen in that specific case for all of that to go right.

而且，如果你再重复一千次，可能都不会再遇到那样一个医生，在那一刻恰好知道该给她用什么抗体，让治疗从致命变成治愈。

And probably if you roll that dice a thousand more times, you probably don't get that that doctor at that moment who knew exactly what antibody to give her to make that treat treatment curative instead of lethal.

所以，我们真正想自动化的，就是这类充满巧合的事情。

So, again, like, those are the types of, like, serendipity things that we'd we'd actually like to automate.

所以，总之，这样就很便宜。

So, anyway, so inexpensive.

后来人们发现，其实只要通过一次输液，如果你拿一个编码嵌合抗原受体的mRNA，把它放进一个叫脂质纳米颗粒的脂肪球里，再在这个脂肪球外面加上一个CD8靶向部分，然后通过输液给病人。

People then realize that actually through just an infusion, if you take an mRNA that encodes for the chimeric antigen receptor, you put it in a ball of fat called a lipid nanoparticle, you put a CD eight targeting moiety on the outside of this ball of fat, It will then you give them in the infusion.

它会去结合T细胞，然后被吞进去。

It will go bind to the T cell, get ingested.

脂肪球溶解了。

Ball of fat dissolves.

mRNA释放出来。

MRNA comes out.

嵌合抗原受体表达出来，并呈现在T细胞表面。

Chimeric antigen receptor gets expressed and presents on the top of the T cell.

所以你就是让T细胞重新编程，去表达这些奇怪的抗原。

So you're just telling you're reprogramming the T cells to express these weird antigens.

真的是在编程生物学。

Literally programming biology.

对。

Yep.

然后T细胞就会去执行任务，把带有CD19的细胞都清除掉。

And then the T cell goes and does its thing and wipes out whatever has c d 19 in this case.

所以恶性B细胞能解释很多血液癌症。

So malignant B cells are explain a lot of blood cancer.

它们也能解释很多自身免疫性疾病。

They also explain a lot of autoimmune diseases.

B细胞常常会产生针对自身抗原的抗体之类的。

B cells often make antibodies in response to autoantigens and things like that.

所以最近，六个月前，经过大约六年的研究，从一位诺贝尔奖得主的实验室孵化出来，投入了大约一亿美元的研发资金。

So recently, six months ago, as the result of about six years' worth of work spun out of a Nobel Prize winner's lab and about a $100,000,000 worth of r and b r and d.

我们看到了——这音乐真不错，兄弟。

We saw That's some good music, man.

是的。

Yeah.

对。

Yeah.

对。

Yeah.

R&B。

R and b.

对。

Yeah.

我们看到了治疗自身免疫疾病的体内CAR T疗法中一些最有说服力的临床前数据。

We saw, like, some of the most compelling preclinical data for in vivo CAR T treatment of autoimmune diseases.

那是一家叫Capstan的公司。

It was by a company called Capstan.

他们被艾伯维以大约21亿美元收购了。

They were bought by AbbVie for, like, $2,100,000,000.

所以在Lila，我们一直在分别研究这三件事。

So at Lila, we had been working on all three of those things in isolation.

大概六个月前，Lila 内部有两三个人组成的小组，想看看我们在体内 CAR T 方面能做些什么。

So about six months ago, a team of two or three people inside of Lila tried to see what we could do to in the in vivo CAR T.

我们一直在研究的是 mRNA 设计。

And what we had been working on was mRNA design.

大多数 RNA 药物，最关键的可调参数就是表达峰值和表达持久性。

So, like, most RNA medicines, the biggest knob that you can turn is expression peak and expression durability.

也就是说，当你给人注射疫苗或其他 mRNA 药物时，每单位 mRNA 能产生多少蛋白质？

So how many proteins do you get per unit of mRNA when you give someone a vaccine or some other mRNA medicine?

我们开发了一些非常强的 UTR，也就是非翻译区，它们位于蛋白质编码区的两侧，决定了这些表达特性。

So we have developed some monster UTRs, untranslated regions, which flank the the protein coding region, which dictate those expression properties.

大约是 Moderna 和 Pfizer 参考值的十倍。

Something like 10 x the references from Moderna and Pfizer.

在六个月的时间里，我们在非人灵长类动物中获得了体内数据，B 细胞清除效果明显优于 Capstan 的数据，而且持久性以及我们观察的所有特征都显著更好。

And over the course of six months, got to in vivo data in nonhuman primates, where B cell depletion was significantly better than what was shown in the Capstan data, and the sort of, like, durability of that was also all of the the characteristics that we looked at were significantly better.

增加 CAR 的表达可能是提升 CAR T 疗法效果最有效的方法之一。

Having more CAR expression is probably one of the most potent ways to improve a CAR T therapy.

受体表达的数量决定了T细胞一旦找到坏细胞，它有多大的可能性去结合它，而T细胞简直就是连环杀手，它们会杀死一个细胞，然后继续去杀下一个，再下一个。

The number of receptors that get expressed dictates how likely that T cell is to bind to the bad cell once it finds it, and T cells are literally serial killers, and that they will go they'll kill a cell, then they'll go to the next one, the next one.

所以它们能持续多久，取决于CAR表达的持久性。

And so how long they can do that is dictated by the how durable the expression of the the CAR is.

再说一遍，我们不是做CAR T的公司。

Again, we're not a CAR T company.

我们是科学迷。

We're science nerds.

我们喜欢做酷的事情。

We like to do cool stuff.

所以我们在大约六个月内达到了那个验证点，甚至到了你可能考虑为新临床资产提交IND申请的地步。

So we got to that proof point in about six months where, again, all the way up to where you might think about filing an IND for a new clinical asset.

我们不会那么做。

We're not gonna do that.

我们不会进行临床试验。

We're not gonna do a clinical trial.

再说一次，那会涵盖所有方面。

Again, that would be all all encompassing.

但有些在Lila待了很久的人，把这看作一种基本上两到三人的全职创业方式，就是几个有领域知识的科学家，加上模型和平台的组合，能在六个月内完成五年的生物技术工作，而总投资只需10%。

But some folks who had been around Lila for a long time saw that as a way to do essentially like a two to three person FTE startup, where there's a couple scientists who have domain knowledge and a combination of the model plus platform can do five years worth of biotech work over a six month period for 10% of the total investment.

所以我们现在考虑的很多商业合作，基本上就是零全职员工的创业模式，就是有人带着一个想法来。

And so a lot of the, like, commercial relationships we're thinking about now are essentially like the zero FTE startup model, where someone comes with an idea.

他们说，你知道，如果市场上有一个CAR T能结合两种东西，如果是双特异性的，或者有这些其他特性，我知道这个产品能填补的市场空白。

They say, you know, if I if there was a CAR T in the market that could bind to two things, if it was a bispecific or if it had these other properties, I know the hole in the market that that thing would plug into.

所以我们很多商业合作实际上就是虚拟创业公司跑在Lila上，有人带着一个非常明确的问题来，但他们不知道怎么实现。

And so a lot of our commercial engagements are effectively virtual startups running on Lila now, where someone comes with a very well specified problem, they don't know how to get there.

可能还涉及靶点识别之类的事情，但他们基本上可以在更短的时间内，以更低的成本完成整个项目。

There may be some things related to target identification and things like that too, but they can effectively run that entire program over a much shorter amount of time with at a fraction of the cost.

所以这些就像是，一个合作伙伴来找你说，我有一个想法。

And so those are those are like, a partner comes to you says, I have this idea.

我不想建一个

I don't wanna build a

我不想建实验室。

I don't wanna build a lab.

我也不想招团队。

I don't wanna hire a team.

我就想直接

I I wanna get

把活儿干了。

their jobs.

就想把它搞定。

Wanna get it done.

对。

Yep.

所以可能是大学里的某个学者，说，我有个想法。

So it could be some academic at a university that's like, I I have this idea.

我稍微做了一点验证。

I kinda did a little bit of validation.

我觉得这个能成。

I think it'll work.

对。

Yep.

我能做吗？

Can I do it?

我能跟你们一起待六个月，把它做出来吗？

Can I sit with you guys for six months and make it work?

可以。

Yeah.

我觉得，这么想是对的。

I mean, that that's the right way to think about it.

对。

Yeah.

从合同上来讲，实际操作是有一个平台接入费的。

The way that, like, contractually it plays out is there's, a a platform access fee.

比如说，我们得支付试剂费用、系统运行费用，还有一些杂项开销之类的。

There's, like, we have to pay for reagents and running the system and then some overhead and stuff like that.

然后还有一些收益分成。

And then there's like some upside sharing.

这是一个可扩展的模式，随着平台越来越好，我们可以提供服务。

That is a scalable model where we can service as the platform gets better.

不用只做几十个，我们可以同时开发几百个甚至上千个虚拟初创公司，这样我们近期有收入来支付账单，同时也能和那些选择和我们一起建设的人建立收益共享的伙伴关系。

Instead of doing dozens of those, we can do hundreds and then thousands of simultaneous, you know, kind of virtual startups being developed on the platform where we we have revenue that helps pay the bills in the near near term, but then we also have this upside partnership with folks who decide to to to build with us.

这太棒了，因为我们看到的是，人们越来越倾向于去掉所有多余的设施，利用自动化，专注于创意本身。

That's amazing because this is what we're seeing is that people are more and more pushing towards getting rid of all the extraneous infrastructure and using automation and focusing on the idea.

我觉得，我们大多数人进入科学领域是因为好奇，想解答问题。

I mean, the way that I think about it is, like, most of us got into science because we're curious and wanna answer questions.

你知道，我本身是学计算机的，我喜欢通过软件来解答问题。

You know, I'm I'm a computer scientist by training, and, like, I like to answer questions through software.

但如果我得用二进制编程，那我可就没那么喜欢了。

However, if I had to program in binary, I would enjoy that significantly less.

现在有越来越高级的抽象层了。

There there are high level abstractions, increasingly high level abstractions.

以前就是Python和Java。

You know, it used to just be Python and Java.

现在都变成云代码了。

Now it's like Cloud Code.

对。

Yep.

它们帮我更快地找到答案。

They help me answer questions faster.

打个比方，科学家们还在用二进制编程。

You know, the analogy is that, like, scientists are still programming in binary.

他们有问题想解答。

They have a question that they wanna answer.

他们得把问题直接转化成实验方案。

They have to just compile that down to an experimental protocol.

然后他们还得去做体力活，像移液那样，结果得了关节炎，这就像是用二进制做科学编程一样。

Then they have to go and do the manual labor and get arthritis by, like, moving liquids from one that's the equivalent of scientific programming in binary.

所以我们想帮科学家们提升抽象层次，比如，也许你的想法行不通。

And so we're trying to help scientists move up the abstraction ladder where what's the like, you know, maybe your idea isn't gonna work.

大多数临床试验都会失败，但如果你不用同时做体力活和一部分脑力活，把所有环节都安排妥当，你至少可以快速失败。

Most clinical trials fail, but you can at least get to failing fast if you don't have to do both the physical labor and also some of the intellectual labor to, you know, get all the the pieces in the right place.

你知道，大多数临床试验都会失败，大概在58%左右。嗯。

You know, most clinical trials fail, you know, somewhere between 58% Mhmm.

实际上，从IND到嗯，临床试验的成功率。

Of clinical trials actually get from IND to Mhmm.

批准。

Approval.

所以，发现并不是真正的瓶颈。

So the the discovery is not actually the constraint.

嗯。

Mhmm.

我挺感兴趣的。

I was interested.

你刚才提到的那种经济建模代理。

You were talking about the sort of economic modeling agent.

我记不清你具体怎么称呼它了。

I can't remember what exactly you called it.

但我觉得，那才是真正需要解决的问题。

But, I mean, that seems like the problem to solve.

对。

Mhmm.

你怎么看这个问题？

How do you think about this?

就是临床试验的成功率？

The pro the the success rate of of clinical trials?

嗯，那背后的经济模型，关于整体扩展的，是怎样的？

Well The economic model underlying what scaling in general?

不管是生物还是材料，很多时候都有一个特别特别长的流程。

For both bio and materials, like, oftentimes, there's this huge huge process.

比如，一旦你有了一个你认为最终的东西，像是一个IND或者开发候选物。

Like, once you have something which is you consider final, like a I n d or d development candidate

对。

Yeah.

对于材料来说，通常还有十年的临床试验，或者材料科学领域的认证，才能把它变成产品。

For for material, like, there's still usually, like, ten years of clinical trials or, you know, qualification in the material science world to just get that into a product.

对。

And Yeah.

很多时候，瓶颈在于规模化生产、监管方面。

Oftentimes, the bottlenecks there are things about scale manufacturing, about regulatory Yep.

还有安全性，以及那些往往很难提前回答的问题。

About safety, and things that are oftentimes just very hard to answer upfront.

所以每次做这个，你只能碰运气。

So every time you do this, you just have to three, you roll a die.

而且，如果你要处理这个问题，基本上就是一个投资组合的模式。

And there is the typical, if you will deal with this, is, you know, essentially a portfolio model.

对。

Yep.

你知道，从融资的角度来看，这很大程度上是——你唯一能赚钱的方式就是通过某种程度的风险校准来扩大规模。

You know, financing wise, it's it's very much a you know, the only way you can make money is if you scale with some level of, like, you know, risk calibration.

对。

Yep.

你知道，听到你能具体做这些事情真的很令人兴奋，但这如何融入更大的图景呢？即使你立刻解决了这些问题，那也只是整个问题的10%。

You know, it's really exciting to hear that you can do these things specifically, but how does it feed into the larger thing where even if you solve these problems immediately, it's still only 10% of the problem.

美国生物技术输给中国生物技术的原因不是创新问题。

The reason why US biotech is losing to Chinese biotech is not because of an innovation problem.

还有一个监管框架需要支持快速临床试验。

There's a regulatory framework too that has to go to enabling, like, fast clinical trials.

FDA最近在这方面采取了一些措施，既包括在某些情况下需要提交的临床前数据，也包括我们将如何运行和监控试验。

The FDA has made motions towards that recently, both for the preclinical data that you have to submit in some cases, but also how we will run and monitor trials.

对。

Mhmm.

所以，如果觉得某一家公司，甚至所有公司加在一起，就能凭一己之力改变这一切，那也太离谱了。

So it it'd be crazy to think that, like, one company could come or even, like, any company combined could could change could change that on their own.

所以必须和监管机构一起合作才行。

So it has to be done in tandem with the regulators.

不过，临床前成功率哪怕只是小幅提升，影响也很大。

However, the minor moves in preclinical probability of success matter a lot.

从投资组合理论的角度来看，这会让投资变得更有吸引力。

You know, from a portfolio theory perspective, it makes an investment much more attractive.

也就是说，如果失败的项目更少，预期上药物就能更快地送到患者手中。

It means, you know, in expectation, medicines get to patients faster if fewer of them fail.

所以我想说，这就是我们现在专注的方向——通过一个已经受益于上百万种独特mRNA设计的系统，来最大化那些已知能转化为治疗益处的因素，这会实实在在地推动那些临床前……你知道，我觉得，与其用一枚公平的骰子，不如用一枚灌了铅的骰子。

And so I I would say, like, you know, that is the area that we're focusing on now is that medicine created by a system that has had the benefit of, in this case, a million unique mRNA designs to maximize things that are known to translate to therapeutic benefits will will meaningfully move those pre it's still you know, it's I guess, would say it's better to throw a loaded die than it is a fair die.

是的。

Yeah.

所以我们就是想尽量让这个骰子更有偏向性。

And so we're just trying to, like, make the die as loaded

我觉得我一直在想的就是，你怎么把转化这件事，对吧，以及在材料领域里对应的东西，真正做起来。

I as guess my thinking this is the thing that I think about constantly is how do you bring, you know, basically translation, right, and whatever the equivalent is in materials.

不管你怎么称呼它，我真的希望看到一个模型，它能思考这些因素，并且很擅长说，我在筛选我的设计，只留下那些我认为能通过三期临床的。

So however you name that, that I really want to see a model that thinks about these are the factors that reasons about, and is very good at saying, I'm filtering my designs to the ones that I think are going to make it through phase three.

对。

Right.

我妻子就是做转化科学的。

My wife is a translational scientist.

她是做生物技术的。

She's a biotech.

她经常提醒我，你们应该用AI做转化科学。

She reminds me very often, you guys should be doing AI for translational science.

从某种意义上说，我觉得这确实有一些呼应，尤其是在材料和化学领域，我们的工具可以调用过程工程模拟器，去算出应该用多大的管道直径和什么样的热交换器，才能把工艺放大到经济上划算的程度。

In a sense, I think that's sort of some of the echoes, especially in the materials and chemistry, our tools can call process engineering simulators, and go figure out what pipe diameters and what heat exchangers you should be using in order to scale up the process for the economics to be worth it.

所以，还是有可能，我不知道我们会不会在测量之前就加一个过滤器。

So, still, maybe there is, I don't know if we're gonna gank up a filter until we go measure it.

但现在能够提前思考下游会发生什么，这有点像转化科学和人工智能会做的事情，就是提前思考什么才是重要的，因为你之前说过，当你拿到IND时，你就被锁定了，在化学方面确实如此，你选择的分子和序列。

But the ability to reason now about the things that will come downstream, which is sort of a little bit what the translation and sort of AI would do, is reason now about sort of what's gonna matter, because when you said earlier, when you have the IND, you're locked in, and it's true on the chemistry side, the molecule, sequence you've chosen.

当然，你要给哪个群体用药，以及如何衡量成功，这些是你之后做的选择。

Of course, which population you're gonna give it to, and how you're gonna measure success, those choices you make afterwards.

我想说，我不做临床前的工作，但在化学和材料方面，这正是我们现在试图灌输的行为，通过验证器和我们能访问的数据源，要么是因为有人考虑过它们，要么是因为物理规律允许，要么是因为我们现在能测量足够好的代理指标，告诉我们以后会发生什么。

And I would say, I don't work on the preclinical stuff, but on the chemistry and materials, that is precisely the type of behaviors we're trying to instill now, with the verifiers and the data sources that we can access, either because somebody has thought about them, either because the physics allows it, or because we can measure good enough proxies now that tells us what's gonna happen later.

要说明的是，这些我们内部经常讨论。

And to be clear, like, all those things are things that we talk about internally a lot.

我们已经在过度扩展的边缘了。

We're already on the verge of being pathologically over scoped.

你知道吗？

You know?

就是

Just

再多给我一点。

give me more.

再多一点。

Me more.

但绝对，我们相信的是，随着模型变得更聪明，随着它们消化临床试验数据库，随着我们与制药公司合作并接触到那个“饼干罐”，这些先验概率将会发生实质性的变化。

But absolutely, like, the the belief that we have is that as models get smarter, as they ingest clinicaltrials.gov, as we partner with pharma companies and get access to, you know, that cookie jar, these pre these probabilities will meaningfully change.

在生物制造方面，能够接触到良好的生产流程和放大工艺，我们认为模型也能在这方面做出贡献。

On the the biomanufacturing side, having access to, you know, buy good manufacturing processes, scale up processes too, we think the models will be able to contribute there.

我们只是选择把很多商业和合作活动集中在那些我们认为现在就能解决的科学前沿问题上，但目标是超越这些。

We've just chose to focus a lot of our commercial and collaborative activity on the sort of, like, frontier of science that we think we can address now, but the goal is is to push past push past that.

如果让我总结一下，这有点像是一个工具调用。

If I may summarize, it's kinda like it's a tool call.

对。

Yeah.

每个都是，全都是令牌。

Every it's it's all tokens.

全都是工具调用。

It's all tool call.

对。

Yeah.

对。

Yeah.

只要有令牌和工具调用就够了。

Tokens and tool calls are all you need.

但是还有推理机制，比如你提到的那个治疗IL-6抗体的医生，他是从生活经验和阅读文献中学到的。我们相信我们的理论，即广度带来深度，通过做更多的事情，我们会在这方面变得更好。而且，

But also the reasoning mechanisms that maybe you mentioned for that doctor that treated, know, the IL-six antibody that you mentioned, well, person had learned that from a combination of lived experience and reading the literature, and we get since we believe our thesis that the breath gives us depth, we will get better at those things by doing more things And with

在那么多可能的世界里，那个医生不是治疗Emily Whitehead的人，CAR T可能看起来又像是一个失败的药物，又一个墓碑。

so many counterfactual worlds, that doctor was not the one treating Emily Whitehead in that case, and CAR T may have looked like it may have been yet another gravestone in law, you know, for, you know, yet another failed drug.

所以我认为，那件事的成功概率从百分之二变成了百分之九十八，仅仅因为那个人在场。

So I do think that, like, that went from, like, a two percent success probability to a ninety eight percent just because that person happened to be in the room.

所以如果我们能把这个过程操作化，就能大幅改变很多概率。

And so if we could just operationalize that, again, like, the you're gonna move a lot of probabilities.

对。

Yeah.

所以这就是一个很好的例子，说明拥有非常广泛的科学知识有多重要。

So that's a that's a good example of where just having really broad knowledge of scientific information.

对。

Yeah.

所以这几乎就像谷歌一样。

So that's like that's almost like Google.

因为那药当时只用于治疗小儿关节炎。

Because that was only being used in pediatric arthritis.

就是另一个非常小众的医学领域。

Like another, like, very niche area of medicine.

明白了。

I see.

对。

Yeah.

所以你的团队里有肯·斯坦利。

So you have Ken Stanley on your team.

是的。

Yes.

他写了那本著名的书《为什么伟大不能被计划》。

Famously written wrote the book Why Greatness Cannot Be Planned.

对。

Yeah.

而且他在机器学习研究方面非常强调开放性和偶然性。

And is very big on open endedness and serendipity in ML research.

那么在Lila，开放性扮演什么角色？

So what is the role of open endedness at Lila?

哦，对。

Oh, yeah.

首先，肯非常棒。

One, like, Ken is awesome.

所以，对于那些不了解的人来说，肯开创了机器学习和人工智能的一个领域，叫做开放结局，我觉得这就像是机器的创造力。

So for those of you who don't know, Ken pioneered an area of machine learning and AI called open endedness, which I think of as machine creativity.

比如，我们怎么让模型进行开放式的探索，同时还要有品味，知道什么有趣，什么值得深入。

Like, how do we get models to do open ended exploration and also, like, have a sense of taste about what's interesting, what things that should go down.

所以，如果你只是个擅长考试的人，你不可能拥有科学超级智能。

So you can't have scientific superintelligence if you're just a good test taker.

所以，如果你想想强化学习在做什么，即使在大规模下，它也是在以一种冷酷的、瓦肯式的、像斯波克那样的方式回答问题。

So, like, if you think about what reinforcement learning is doing, even at scale, it's answering questions in kind of like a ruthlessly Vulcan esque, like, you know, Spock kind of way.

但你大概只会有限地认为那个模型是极其有创造力的。

But you probably only in, like, limited ways would think of that model as being supremely creative.

所以肯在Lila的团队里创建或构建了一个开放结局的机制，来承担那个推理挑战的外层循环或元部分。

And so Ken has created or built an open ended in his team at Lila to sort of take take the outer loop or the the meta part of that reasoning challenge on.

那么，我们怎么让我们的模型不仅能回答难题，还能首先提出有趣的问题呢？

So how can we get our models to not only be able to answer tough questions, but ask interesting questions in the first place?

所以，这就是肯的真正使命。

And so that's really Ken's mandate.

过去几个月，他一直在组建一支世界级的团队，他们现在正在厨房里忙活着呢。

He's been building like a world class team over the last, like, several months, and they're, you know, they're in the kitchen cooking now.

我觉得到今年年底，肯的团队会拿出一些很酷的东西来分享。

And I think by the end of this year, we'll have some cool stuff from Kin's group to share.

那我们来看一段实验室的视频吧。

So we're gonna hop into a video here of the lab.

这段视频会展示几个不同的东西。

This is gonna show a couple of different things.

好的。

Alright.

那可能是个剥皮器或者封口机。

So that's probably a peeler or a sealer.

当你把培养皿从一个仪器移到另一个仪器时，显然里面会有液体。

So when you move plates from instrument to instrument, obviously, there's liquid in it.

生物学大部分都是湿的。

Most of biology is wet.

所以你要把这些贴纸贴上去。

So you put these stickers on it.

那就是在给培养板封膜。

So that was a plate being sealed.

好的。

Alright.

那我们现在来看这里。

So here we go here.

它在拿起一个……对。

It's picking up a Yeah.

这是在液体处理设备里面。

So this is inside of a liquid handler.

我们等它切到更广的镜头，这样你就能看到PCI总线和一些机械臂。

Let me go we'll wait till it gets to a wider shot here so that you can see the PCI bus and see some of the the robotics.

液体处理设备是磁性的，对。

So the liquid handlers are magnetic yeah.

这就是平面电机系统，板子在这里磁悬浮。

So this is the planar motor system here where the the plate magnetically levitates.

这是PCI总线，传输层通过它连接所有仪器。

This is the PCI bus where the transport layer connects all the instruments.

你可以看到那些台子，所有仪器都放在上面。

You can see benches there where all the instruments sit.

机械臂把它拿起来。

A robot arm picks it up.

现在要把它转移到另一个板子上，继续下一步。

It's now gonna transfer it to a different plate to go on to the next.

这里需要做一些交通控制。

There's a little bit of a traffic control thing that you have to do here.

比如，它们实际上会停下来等一会儿，等轨道拥堵缓解。

Like, they actually will go and park for a while while track congestion clears.

对。

Yeah.