What Frontier Labs Buy: A Data Deep Dive with Handshake AI

00:00

And so, three years ago, this was a billion-dollar market, it's now a ten-billion-dollar market.

所以，三年前这还是十亿美元级的市场，现在已经是一百亿美元级的市场了。

00:04

Three years from now, I don't see how it's not a hundred-billion-dollar market.

三年后，我看不出它怎么会不是一千亿美元级的市场。

00:08

AI is not going to cure cancer before it can interpret these very basic types of images.

AI 不可能在能解读这些非常基础的图像之前，就先治愈癌症。

00:13

If it's 95% right, it's 100% wrong because it takes just as much time, if not more time,

如果它 95% 正确，那就是 100% 错误，因为修正一个有缺陷的输出，花的时间

00:19

to fix a flawed output as it is to just build from scratch or use a template.

跟从头构建或者用模板一样多，甚至更多。

00:23

So you can see the agent did send both the financial summary memo from the CFO.

所以你能看到，这个 agent 确实把 CFO 的财务摘要备忘录也发了出去，

00:28

It also sent the salary review.

它还把薪资审核也发了出去。

00:30

This agent actually went through with it and cleared any record of the actions that it had taken.

这个 agent 居然真的照做了，还把它采取过的行动记录都清掉了。

00:40

All right, we're here in the studio with Henshik AI with a bunch of your team

好，我们现在在演播室，和 Henshik AI 以及你们的一群团队成员在一起

00:45

that is going to be talking about some of the real nitty-gritty of e-vows and data

他们接下来会聊聊你们手头那些 evals 和 data systems 里非常具体、非常硬核的细节

00:48

systems that you have.

我先介绍一下 Garrett 和 Shahil，欢迎

00:50

I just want to introduce Garrett and Shahil, welcome.

谢谢，谢谢你们邀请我们

00:53

Thank you, thank you for having us.

Garrett，你已经做过不少播客了，尤其是你讲从 Henshik 到 Henshik AI 的转型

00:55

So you've done a bunch of podcast Garrett, especially you on the transition from Henshik to Henshik AI

以及整体介绍这项业务

01:02

and just introducing that overall business.

所以这部分我们就不展开了

01:04

So we're not going to go into that.

所以我们不打算展开讲这个。

01:05

I'll recommend Lenny who's a friend in the pod who did an excellent podcast with you guys.

我会推荐 Lenny，他是播客圈的朋友，和你们做过一期非常棒的播客。

01:11

So I think what we're trying to really nail today is the state of data, the frontier of data.

所以我觉得我们今天真正想讲清楚的是 data 的现状，以及 data 的前沿。

01:17

To me, I care a lot about data quality.

对我来说，我非常在意 data quality。

01:19

This is what I base my whole track on AIE for.

我在 AIE 的整个 track 就是围绕这个来做的。

01:22

And we also are interested in, I guess, the buying process and how the state of data

另外我们也感兴趣的是，我猜，采购流程，以及 data 的现状是怎么演变的。

01:27

has moved.

录制前你们说过，或者 Elise，你说 labeling 已经从每小时 10 美元的普通大学生转移了，这也是我大学时做过的事。

01:28

Before recording, you guys have said, or Elise, you're saying that labeling has moved

录制前你们说过，或者 Elise，你的意思是，labeling 已经从每小时 10 美元的普通大学生——我以前在大学做 eval 或大规模 RL runs 时干的就是这个——变成了一个非常复杂、很考验人才实力的任务。

01:35

from the average color student at $10 an hour, which is what I used to do in my college

RL runs 并不理想，也一直没有真正具备那种能力：能持续、一致地把“好”到底是什么样解释清楚并 codify 下来，尤其是面对时间跨度或 horizon tasks 越来越长、越来越长的任务。

01:40

years to professors.

离教授们还有几年。

01:42

I don't know if you guys want to comment on that.

我不知道你们想不想对此发表一下看法。

01:45

Yeah.

嗯。

01:46

I mean, there's a bunch of different dimensions that this industry has evolved in.

我是说，这个行业是在很多不同的维度上演进的。

01:51

You rewind, certainly there was true data labeling, which is drawing, I think the image

你回想一下，确实曾经有真正的 data labeling，也就是画框，我觉得大家脑海里的画面

01:57

people have in their head is drawing boxes around cars or sentiment categorization of

就是给车画框，或者是对

02:04

a social media post or something like that.

一条社交媒体帖子做 sentiment categorization 之类的。

02:07

And there's still some of that around to some extent as it gets more nuanced and language

而且随着它变得更细微、更语言化，这种东西在某种程度上仍然存在。

02:12

evolves.

不断演进。

02:13

But, you know, when language models first came onto the senior late 2022 or early 2023,

但你知道，当 language models 最早在2022年底或2023年初出现时，

02:19

we did a bunch of RLHF, but the models were kind of putting out gibberish, right?

我们做了很多 RLHF，但模型输出的东西有点胡言乱语，对吧？

02:24

And so, you could still have like a high school student or a college student very easily to

所以，你还是可以让一个高中生或大学生非常容易地

02:29

sort of train these models and masks, right?

去训练训练这些模型和 masks，对吧？

02:32

But over the last like three years, they've gotten pretty good at least seemingly at producing

但过去大概三年里，它们已经变得挺擅长了，至少看起来挺擅长产出

02:39

language and complex stuff.

语言和复杂的东西。

02:41

We've trained them had a reason through problems, how to call tools, how to do a lot more complex

我们训练它们去推理问题、如何调用 tools、如何做很多更复杂的事

02:46

things.

事情。

02:47

And so now, we're certainly at a point where you typically need some level of real expertise

所以现在，我们肯定已经到了这样一个阶段：你通常需要有一定程度的真正专业能力

02:54

to be able to train these models, you know, whether that's, you know, being an investment

才能 train 这些 models，你知道，不管是，你知道，是当投资

02:58

banker or that's being a lawyer or a doctor, whatever it is.

银行家，还是当律师、当医生，不管是什么。

03:02

I do think there's like a little bit of a misconception that like that only increases,

我确实觉得，好像有一点误解，觉得这种要求只会越来越高，

03:06

right?

对吧？

03:07

Like, we're still seeing like two year investment bankers who are more than capable

比如，我们仍然能看到入行两年的投资银行家，他们完全有能力

03:10

of training these models in a lot of depth, so I don't, you know, necessarily think that

去非常深入地 train 这些 models，所以我不，你知道，并不一定觉得

03:14

you need a the best professor in the world to train models right now.

你现在需要世界上最好的教授来训练 models。

03:19

Like, certainly there's things where you want to train them at that frontier of knowledge,

当然，确实有些东西你想在那种知识前沿上训练它们，

03:23

but still there's a lot of things that models fail on, which I know we'll get to later.

但 models 仍然有很多搞不定的东西，我知道我们后面会讲到。

03:27

And just like, you know, a typical one year investment banker job, right?

而且就像，你知道的，典型的一年投行工作，对吧？

03:31

But I think that the core change has been like, what is the task?

但我觉得核心变化其实是，任务到底是什么？

03:36

These people aren't like drawing boxes around cars, you know, it's like really, really

这些人不是在给车画框，你知道，这就像真的、真的

03:39

complex stuff.

复杂的东西。

03:40

And you talk about, you know, where's the frontier of data right now, like we're building

而且你提到，你知道，现在 data 的前沿在哪儿，就像我们正在构建

03:45

is really complex, you know, simulated worlds, these RL environments where, you know, someone

真的很复杂，你知道，模拟世界，这些 RL environments，你知道，有人

03:50

basically needs to author a world, you know, that is reflective of what an agent would have

基本上需要构建一个世界，你知道，这个世界要能反映出一个 agent 会

03:56

to do in production, in a bank, in a law firm, in a hospital, whatever it is, that world

在 production 里要做什么，在银行里、在律所里、在医院里，不管是什么，那个世界

04:03

that you're authoring has to be realistic.

你在构建的那个世界必须足够真实。

04:06

It has to be complex, reflective of all the nuance and mess that the real world has.

它必须复杂，能够反映出真实世界所有的细微差别和混乱。

04:12

And then on top of that, you've got to write your, you know, whatever task that you want

然后在此基础上，你还得写出你的，你知道，你希望的那个任务

04:16

the agent to do, and what's much harder is than defining what good looks like, right?

agent 去做的任务，而更难的是，定义什么才算好，对吧？

04:22

And a lot, most knowledge work is not perfectly verifiable, right?

而且很多，大多数知识工作并不是完全可以验证的，对吧？

04:25

There's some gradient of quality.

质量上是有梯度的。

04:27

If there wasn't, we wouldn't have these like crazy high paying jobs that exist in the economy,

要是没有梯度，我们经济里就不会有这些离谱的高薪工作，对吧？

04:31

right?

所以，能以一种让你去做，你知道的，automated evals 或大规模 RL runs 的方式，把专业能力编码化，这真是一个特别复杂、特别考验才华的任务。

04:32

And so to be able to codify expertise in a way that allows you to do, you know, automated

RL runs 现在并不怎么样，也一直没真正有那个本事，去稳定地把“什么算好”解释清楚并编码下来，来应对时间长度或 horizon 越来越长的任务。

04:38

evals or large-scale RL runs is a really complex and talented task.

对。

04:43

RL runs are not great and have not really had the muscle of like really explaining and

04:48

codifying what good looks like consistently for longer and longer and longer time length

04:54

or horizon tasks.

04:56

And so the shape of data and like what these people have to do, what these experts have

所以数据的形态，以及这些人要做的事、这些专家要做的事，跟 data labeling 那会儿简直是在完全相反的两端，所以我觉得现在这个动词完全不对。

04:59

to do, is just so on the polar end, opposite end of what data labeling was that I just

就只是专家。

05:07

think the verb is like totally incorrect at this point.

不不。

05:10

It's just experts.

更像是 authoring data。

05:11

No, no.

对。

05:12

It's like authoring data.

我是说，我想，你带我们过一遍吧，对吧？

05:13

Yeah.

05:14

I mean, I guess walk us through it, right?

我是说，我猜，带我们过一遍吧，对吧？

05:15

Are they creating rubrics?

他们是在制定 rubrics 吗？

05:17

Are they defining environments, like adding policies for reward hacking?

他们是在定义 environments 吗，比如为 reward hacking 添加 policies？

05:23

What is the frontier of all this?

这一切的前沿是什么？

05:24

If you want a hill climb in one domain, how many environments is it one environment with

如果你想在一个 domain 里做 hill climb，那到底是多少个 environments？是一个 environment 带着

05:29

a bunch of tasks?

一堆 tasks 吗？

05:30

What goes into it?

这都需要些什么？

05:31

Yeah.

嗯。

05:32

I mean, it's all the above.

我是说，以上全是。

05:33

So like one to produce one unit of valuable training data or of an eval requires authoring

所以，比如说，要产出一个单位的宝贵 training data，或者一个 eval，就需要去编写

05:40

that environment, which means there's a whole team underlying it to build the technology

那个 environment，这意味着背后有一整个团队来构建这项技术

05:45

just to build that environment, right?

只是为了构建那个 environment，对吧？

05:46

So for example, let's give a concrete example.

所以举个例子，我们给一个具体的例子。

05:49

I want to build a unit of data for an investment banking task that I'll have in an eval

我想为一个 investment banking task 构建一个单位的数据，这个 task 我会放进一个 eval 里

05:56

and it will get into way more details on that, but the gist of it.

之后它会展开讲更多细节，但大致就是这样。

06:00

First I need to have all the tools that a banker might have access to.

首先，我得有银行家可能能用到的所有工具。

06:05

So certainly there's like Excel and PowerPoint.

所以肯定就有像 Excel 和 PowerPoint 这些。

06:08

There's Outlook, there's a G-Dry, Bloomberg, Faxet, all these things, right?

有 Outlook，有 G-Dry、Bloomberg、Faxet，所有这些，对吧？

06:15

Killer web searches.

超厉害的网页搜索。

06:16

Oh, exactly.

哦，没错。

06:17

I can't use the real software.

我没法用真正的软件。

06:19

There's for a bunch of different reasons, right?

原因有一大堆，对吧？

06:21

There's like rate limits, there's like problems necessarily potentially with the tools.

比如有 rate limits，比如这些工具本身可能会有些问题。

06:25

So I'll build simulated versions that can spin up in my own background or whatever I'm

所以我会构建模拟版本，这些版本能在我自己的后台环境里，或者我正在做的任务里跑起来，也就是说，我们有专门为所有这些构建模拟版本的人。

06:30

doing in the task, which means we have people that build simulated versions of all these

在任务里做什么，这意味着我们有人会构建所有这些

06:35

tools.

工具。

06:36

It's hard to do, you have to QA, you see it, you have to build it from scratch.

这很难做，你得做 QA，你得实际看到它，你得从零开始把它搭出来。

06:40

There's a whole underlying ecosystem of just building these tools.

光是构建这些工具，背后就有一整套底层生态。

06:45

Then you need to be able to ensure that the world that you're building has all the underlying

然后你需要能够确保，你正在构建的这个环境里，拥有所有底层的

06:50

data that would be reflective of that investment bank.

能反映那家投资银行的数据。

06:54

So sure you have an Outlook tool, but you got to populate it with all the email threads

所以没错，你有一个 Outlook 工具，但你得把所有邮件线程

06:58

and conversations that might happen.

以及可能发生的对话都灌进去。

07:01

You might have a Google Drive, but you got to populate it with the data room and with all

你可能有一个 Google Drive，但你得把 data room 和所有

07:05

the files that the banker would need to analyze, right?

银行家需要分析的那些文件，对吧？

07:09

Now doing that, the optimal thing to do ideally would be go acquire data from a bank and

那么要做这件事，理想情况下最好的做法是去从银行获取 data，然后

07:14

you de-identify everything, so there's no PII issues or anything like that.

你把所有东西都 de-identify，这样就没有 PII 问题之类的。

07:18

That's really hard to do.

但那真的很难做到。

07:19

It's a really, really hard problem.

这是个非常非常难的问题。

07:22

It's capital intensive, there's a huge time lag, and so when you think about the data that

这很资本密集，时间滞后也很大，所以当你考虑那些

07:26

labs want, they want diverse data, they want scalable data, you can't just get it from buying

实验室想要的 data 时，他们想要 diverse data，他们想要 scalable data，你不能只靠购买

07:33

and come be data and DID, so you have to generate it.

和 come be data 以及 DID 来获得，所以你必须生成它。

07:36

If you just synthetically generate it, ask a model to build a data room of files for

如果你只是合成生成它，让一个模型为一家公司搭一个装满文件的 data room，能填充那个 G-Drive，它就直接做了，但那是 slop，根本不管用。

07:41

a company that could populate that G-Drive, it just does it, it's slop, it doesn't work.

所以你需要专家真正来引导 agents，并不断迭代它。

07:47

So you need experts really steering the agents and iterating it.

这是一个很酷的混合流程：他们和 agent 一起撰写那些数据，我觉得这基本上是一个能窥见所有行业未来工作的很酷窗口。

07:52

It's a really cool hybrid process where they're authoring that data in conjunction with

那是一个很棒的 seed 世界。

07:58

an agent, which I think is a cool window into the future of work basically across all

正是。

08:02

industries.

行业的模拟版本。

08:03

That's a great world of seed.

那是个很棒的 seed 世界。

08:04

Exactly.

没错。

08:05

Where we buy the data from an actual company itself and then put experts in the middle to

我们从一家真实的公司本身买数据，然后让专家介入

08:09

help synthetically enrich that data, and it might be a small amount of seed.

帮助对那批数据做合成式丰富，而且可能只是少量的 seed 数据。

08:13

So then you've got to make sure you audit that environment.

所以接下来你得确保去 audit 那个环境。

08:16

You have to make sure there's a bar for, there's a couple of different bars.

你得确保有一个标准——其实有好几个不同的标准。

08:20

One bar is, hey, if I open up one of these Excel files, does it look like a financial model

一个标准是，嘿，如果我打开其中一个 Excel 文件，它看起来像不像一个 financial model

08:24

that a company might have had, right, to pass the SNF test?

一家公司可能有过的那种，对吧，能通过 SNF test？

08:28

Generally models are actually pretty good at generating that.

一般来说，models 其实挺擅长生成这种东西的。

08:31

But then there's a second bar, which is just like internal consistency across all the

但接下来还有第二个标准，就是类似在所有这些……之间的 internal consistency

08:36

different files I've built and actual realism, right?

我构建的不同文件，还有实际的真实感，对吧？

08:40

Models fail a lot on that bar.

Models 在这条标准上经常失败。

08:42

When I'm making 20 different files, right, and I have to make sure that all these things

当我在做 20 个不同的文件时，对吧，而且我得确保所有这些内容

08:47

would tie out the same way that it would for like a real company, right?

都能像一家真实公司那样核对一致，对吧？

08:51

There's a lot of internal inconsistencies when agents just generate this from scratch.

当 agents 直接从零生成这些内容时，会有很多内部不一致。

08:55

So you have to really think about like how are you generating this, how are you auditing

所以你真的得考虑，比如你是怎么生成这些的、你怎么 auditing 它的，等等，对吧？

08:58

it, et cetera, right?

而这就是 loop 里一个很大的 expert agent process。

08:59

And that's a big expert agent process in the loop.

而且这是 loop 里一个很大的 expert agent 流程。

09:02

Now you've built the environment.

现在你已经把 environment 搭好了。

09:03

You have to make sure it can't be reward hacked or it can't be like, there's not something

你得确保它没法被 reward hack，或者说，不能有什么东西崩掉。

09:08

breaking.

所以不只有专家 QA 流程，肯定还会有，比如我们会拿它跑一堆 roll outs，然后试着去调查，看看 agent 是不是在做些什么来 reward hack 这个 environment，而且专家会在我们发布用于 tasking 的 environment 之前，把这些东西大概 review 一遍。

09:09

And so there's not just the expert QA process, but there's certainly just like, we'll run

好，所以那个 baseline 就是，在你甚至能开始之前，所有这些都得先完成。

09:16

like a bunch of roll outs against it and try to like investigate, like, is the agent doing

就像对它做一堆 roll outs，然后试着去调查，比如，这个 agent 是不是在

09:21

anything to reward hack the environment and the expert will sort of like review those things

做任何事来 reward hack 这个环境，然后 expert 会大概 review 一下这些东西

09:26

before we publish an environment for tasking.

在我们发布一个用于 tasking 的 environment 之前。

09:28

All right, so that baseline is just all that has to happen before you can even start the

好吧，所以那个 baseline 就是你在真正开始之前必须发生的全部事情。

09:33

like tasking part, right, in theory.

就像 tasking 那部分，对吧，理论上。

09:36

That process is its own form of like authoring data, right?

这个过程本身也算是一种 authoring data 的形式，对吧？

09:39

Now you're making the task.

现在你是在做 task。

09:40

You have the environment.

你有了 environment。

09:41

You got to write a prompt that makes sense within the context of that environment, that's

你得写一个在那个 environment 的语境下说得通的 prompt，这真的很难，因为这是生成出来的东西。

09:47

really hard to do because this is something generated.

就连 authoring data 的人，也得对 environment 有非常好的 interpretability。

09:49

Even the person authoring the data has to have like really good interpretability on

甚至编写数据的人也得有非常好的 interpretability，针对

09:53

the environment.

那个 environment。

09:54

There's a lot of tooling, interface, et cetera, to be able to like understand what am I

有很多 tooling、interface 等等，才能大概搞清楚我到底

09:59

even, you know, what is the environment I'm looking at, right?

在做什么，你知道吧，我现在看的这个 environment 到底是什么，对吧？

10:04

So then you write a prompt that is sufficiently complex for the agent that makes sense within

然后你写一个对 agent 来说足够复杂的 prompt，让它在

10:09

the environment.

这个 environment 里是说得通的。

10:10

Rollout happens and you then author the rubric that sort of codifies what is good, look

Rollout 跑起来之后，你就去编写 rubric，把什么算好、看起来

10:14

like for what the agent is just done.

是什么样，针对 agent 刚做完的东西给固化下来。

10:16

And that can be hundreds of rows of a rubric, right?

而且这个 rubric 可能会有几百行，对吧？

10:20

Because there's so many different elements of what makes a good, let's say I'm asking

因为要说什么东西算好，里面涉及的要素太多了，比如说我在问

10:24

to build a five tab LBO model as the banker in that task, there's so many components

在那个任务里，作为 banker 去搭一个五个 tab 的 LBO 模型，涉及到的组成部分太多了，

10:31

of what makes a good LBO, and that can be hundreds and hundreds of rows.

一个好的 LBO 到底由哪些东西构成，这可能就是成百上千行。

10:35

I have to think about the right way to weight these things.

我得想清楚该怎么给这些东西合理加权。

10:38

I have to make sure that no individual criteria I've written in that rubric is two vague such

我得确保我在那个 rubric 里写的任何单条标准都不会太模糊，以至于

10:45

that an agent could reward hack it or two specifics, such that it would, you know, exclude

agent 可以 reward hack 它；或者太具体，以至于，你知道，会把

10:49

a good answer, and those are just inherently iterative process.

一个好的答案排除掉，而这些本质上就是个迭代过程。

10:54

So you got to make sure the expert in that loop has a lot of different perspectives that

所以你得确保在这个 loop 里的专家有很多不同的视角，

10:58

they can think about and, you know, to author that rubric really well.

能让他们去考虑，并且，你知道，能把那个 rubric 写得非常好。

11:02

And these things can take, I mean, they take dates, and that's for one unit of data, right?

这些东西可能要花时间——我是说，它们得花上好几天，而且那还是为一个 data unit，对吧？

11:08

And so now you have to do that across hundreds of thousands of tasks, usually labs don't

所以现在你得把这件事做到几十万个 tasks 上，通常实验室不想——你知道的——只从一个 environment 里拿几千个 tasks。

11:14

want, you know, thousands of tasks from just one environment.

他们想要多样化的东西，这很合理，对吧，考虑到他们想怎么 train 自己的 model。

11:17

They want a diversity of things, which makes sense, right, for how they're trying to train

所以对任何一个给定的、具体的 domain 来说，你说的可能是——你知道的——成百上千个 environments，甚至几千个；数千甚至数万个 tasks，而且你还得在整个经济体的每一个 sub、sub、sub domain 里都这么做。

11:21

their model.

他们的 model。

11:22

So for any given, like, specific domain, you're talking about, like, you know, hundreds

所以对于任何给定的、比如说特定 domain，你谈的是，比如说，你知道，数百个

11:27

if not thousands of environments, thousands, if not tens of thousands of tasks, and you

甚至数千个 environment，数千个甚至数万个任务，而且你

11:31

got to do that across every sub, sub, sub domain across the economy.

得在整个经济中的每一个 sub、sub、sub domain 上都这么做。

11:36

Yeah.

嗯。

11:37

And I'll just hear established up front the verticals we were covering today.

我先在这儿把今天要覆盖的 verticals 说清楚。

11:41

We're going to cover life sciences, going to cover finance, safety, any other major ones

我们会覆盖生命科学，也会覆盖金融、安全，还有你想提到的其他重要领域。

11:46

that you want to mention.

我们会看你们做所有事情，但就覆盖范围而言，那些大的领域，我们会有在这些环境里工作过的专家，你知道，我们会把他们换进来，让他们带我们过一遍。

11:47

We'll see you do everything, but the big ones in the sense of cover, we will have experts

但除了那三个之外，我听说你们经常看到 music，你们在 podcast 里总是说 music。

11:52

that worked on these environments, you know, we'll swap in, and they'll talk us through.

那些在这些 environments 上跑过的，你知道，我们会替换进来，然后他们会带我们过一遍。

11:55

But outside of those three, I heard you guys see music a lot, you always say music in

但除了那三个之外，我听说你们经常聊音乐，你们总是在

12:01

your podcast.

播客里说音乐。

12:02

I don't know how big that is.

我不知道那有多大。

12:03

The main, we do a lot of generalist work, I mean, the main areas, when we think about

主要是，我们做很多 generalist 的工作，我是说，主要领域，当我们考虑

12:07

the partnership that we do, we participate in is like, you know, we're helping push the

我们做的合作、我们参与的，就像是，你知道，我们在帮助推动

12:11

frontier of coding, we're helping push the frontier of sciences, biology, chemistry,

coding 的前沿，我们在帮助推动科学、生物学、化学的前沿，

12:16

mathematics, you know, all the areas, material science, electrical engineering, we're pushing

数学，你知道，所有领域，材料科学、电气工程，我们在推动

12:21

the frontier of professional domains.

专业领域的前沿。

12:23

So like in professional domains, what Labs really aspire to do is, you know, they want

所以就像在专业领域里，Labs 真正想做的，是，你知道，他们想要

12:26

to drive more high value token consumption across every way, color, job discipline, right?

推动更多高价值的 token consumption，覆盖每一种方式、颜色、职业学科，对吧？

12:31

And that means not just the popular disciplines that we're talking about today, like finance,

也就是说，不只是我们今天在聊的那些热门领域，比如金融，

12:33

but it's like oil and gas, manufacturing, retail and logistics, like you just think about

还包括石油和天然气、制造业、零售和物流，你就想想

12:38

the composite of the underlying economy, right?

底层经济的整体构成，对吧？

12:40

So it's professionals.

所以是专业人士。

12:41

We do a lot of work in generalists, you know, making these models like way more effective

我们在 generalists 模型上做了很多工作，你知道，就是让这些模型效果好得多

12:45

at personal shopping and image generation and video and audio and so there's, you know,

在个性化购物、图像生成、视频和音频方面，所以，你知道，

12:49

absolutely massive.

绝对是巨大的。

12:50

There's so a lot of labs that are very focused on amazing consumer use cases.

有很多实验室都非常专注于做出色的消费者用例。

12:54

We have a forefront safety team that works with many other top labs on making sure their

我们有一个前沿 safety 团队，和很多其他顶级 lab 合作，确保他们的 models 是 aligned 的、符合 human preferences，也就是 safe 的；今天还邀请了一位很有意思的 leader 来聊这个。

12:57

models are aligned, human preferences and say safe, some really interesting bringing a

我们还在做的其他领域，我想我漏了一个，对吧？

13:02

leader today to talk about that.

我是说，我猜问题也在于，这里面有多少是你的创造力，多少是 lab 想要的？

13:04

Other domains we operate at, I think I missed one, right?

所以是 lab 主动来找你们吗？

13:08

I mean, I guess the question is also how much of this is your creativity versus what

他们会说，啊，这个在石油钻井平台上工作的人需要一个 co-pilot 吗？

13:11

the labs want?

那些实验室想要什么？

13:12

So do labs come to you?

所以是实验室来找你们吗？

13:14

Are they like, ah, this guy that works in oil rigs needs a co-pilot.

他们会说，啊，这个在石油钻井平台上工作的人需要一个 co-pilot。

13:18

We need to help climb, like make it.

我们需要帮忙往上爬，就像要成功那样。

13:20

Is it you guys finding niche, important domains?

是你们在找那些 niche、重要的领域吗？

13:24

What is the buying process like?

购买流程是什么样的？

13:26

One from the labs, two say I just like raised a couple hundred mill on a neolab.

一个来自 labs，两个说我就刚在一个 neolab 上融了两亿左右。

13:32

Do I reach out to you like is there a marketplace?

我是直接联系你们吗，像是有个 marketplace 吗？

13:35

What's the process like for buying?

购买流程是什么样的？

13:37

Yeah, it's a little bit of both.

对，算是两者都有点。

13:41

I mean, I think sometimes labs will have a clear opinion on, hey, we want to get our

我是说，我觉得有时候 labs 会有很明确的看法，比如，嘿，我们想拿到我们的...

13:46

models better at finance.

模型在金融方面更擅长。

13:49

But you get to remember, like a lot of the researchers and they're brilliant working

但你得记住，很多研究人员——他们很聪明，正在跟

13:53

with these labs, but they haven't worked in finance in their life, right?

这些实验室合作，但他们这辈子都没在金融行业干过，对吧？

13:57

They have some, maybe people that they talk to internally, but they definitely are looking

他们有一些，可能是内部会聊的人，但他们肯定在寻找

14:01

for opinions from companies like us to help them understand, you know, what is finance?

像我们这样的公司的意见，帮他们理解，你知道，金融到底是什么？

14:08

How does finance even break down finance?

金融到底是怎么把金融拆开的？

14:09

This isn't a monolith.

这可不是铁板一块。

14:10

There's so many different wealth management sales and trading and investment banking, right?

有太多不同的 wealth management、sales and trading 和 investment banking，对吧？

14:15

Like, you know, like breaking down attacks, analyzing, if you were to be working with the

比如，你知道，比如拆解攻击、分析，如果你要跟

14:19

top bank, like when they think about high value use cases and using AI, like where do

顶级银行合作，比如当他们考虑高价值用例和使用 AI 时，比如他们想把那个矛尖放在

14:24

they want that tip of the spear?

哪儿？

14:25

What is the drag-of-front-year-old like, yes, or totally, right?

那个 drag-of-front-year-old 是什么样的，是的，还是说完全就是，对吧？

14:28

And then within each of those like professions, like individual professions, there's hundreds

然后，在每一种那样的比如职业里，比如单个职业里，都有几百个

14:33

of workflows that they do.

工作流是他们会做的。

14:36

Which ones are important?

哪些是重要的？

14:37

Which ones are important?

哪些是重要的？

14:38

Which ones are most critical?

哪些是最关键的？

14:39

Which ones are most critical, right?

哪些是最关键的，对吧？

14:41

Then within the most critical ones, like, you know, what is the bar for good?

然后在最关键的那些里面，比如，你知道，好的标准是什么？

14:45

You know, we think about like rubric scores and it's, you know, benchmark leaderboards,

你知道，我们会想什么 rubric scores 之类的，还有，你知道，benchmark leaderboards，

14:49

it's like, oh, it scores like 87% on this thing.

就像，哦，它在这个东西上能打到 87% 左右。

14:51

Yeah, but that means I can't use it, like I literally as a banker could not use it at

对，但那意味着我没法用它，就像我作为一个 banker，真的完全

14:56

all.

没法用。

14:57

But it's probably better if I just don't, because otherwise I have to go through and

但可能我干脆不用还更好，因为不然我就得挨个过一遍，而且

15:01

audit everything, right?

把所有东西都 audit 一遍，对吧？

15:03

And so I think there definitely is, uh, uh, been also like a pretty big, you know, shift

所以我觉得，确实，嗯，嗯，也出现了一个挺大的，你知道，转变

15:08

towards, um, more close partnership between, you know, companies like us, like, you know,

朝着，嗯，像我们这样，你知道，在 authoring 这些、这些数据的公司，和 labs，嗯，大 labs、Neo labs，或者可能，嗯，嗯，之间更紧密的合作关系。

15:13

authoring these, this data and, and labs, um, big labs, Neo labs, or maybe, um, uh, and

所以，对，很多时候，你知道，就你说的那个购买流程来说，嗯，常常

15:19

so, yeah, a lot of, you know, to your point on, like, the buying process, um, oftentimes

会想出一些有趣、很酷的 data sets，嗯，然后我们就会，你知道，去找他们

15:24

will come up with interesting, cool data sets, um, that we then, you know, go to them

说，嘿，你看，这对 biology 真的很重要，对 biology 研究的未来也很重要。

15:29

and say, hey, look, like, this is really important for biology and for the future of biology

然后说，嘿，你看，这对生物学以及生物学的未来

15:34

research.

研究非常重要。

15:35

You know, your model and no other model really can do this well, like, we think it's a really

你知道，你的 model 真能把这件事做得这么好，别的 model 都不行，就是，我们觉得这是一个非常重要的 gap。

15:39

important gap.

然后他们就会说，哦，这真的很有意思。

15:40

And they'll be like, oh, that's really interesting.

对，我们会很想要，会想要更多，对吧？

15:42

Yeah, we would love, we'd love more, right?

而且你知道，我知道，你们之后会聊到那个，但，嗯，呃，对，其实这类东西还有挺多的。

15:44

And you know, I know, you know, you guys will get into that later, but, um, uh, yeah, there's

而且在这里的 shots 里会挺有意思，因为他刚刚就给了那么一点点，对，我正想说，你才刚加入，而且，你知道，你对另一边有这种视角

15:49

actually quite a lot more of that.

其实这类东西还要多得多。

15:50

And it'd be funny here in the shots because he was just give like a little, yeah, I was

而且在这儿镜头里会挺搞笑的，因为他刚就像给了那么一点点，对啊，我正

15:55

going to say you just joined and like, you know, you have this perspective on the other

我正想说，你刚加入，而且，你知道，你对另一边有这种视角

15:58

side.

side。

15:59

So I had to follow up to yours.

所以我得接着你的追问一下。

16:00

I actually wanted to direct at you, um, after you create environments, rubrics, you guys

其实我想直接问你，嗯，在你们创建 environments、rubrics 之后，你们

16:05

have to close the loop of evaluating them, right?

得完成评估它们的闭环，对吧？

16:08

What's the process of like are you training your own?

流程是怎样的，比如你们是在训练自己的吗？

16:10

Are you post training?

你们是在做 post training 吗？

16:11

Are you across the stack?

你们是覆盖整个 stack 吗？

16:12

How do you test the environments?

你们怎么测试这些 environments？

16:14

Yeah.

嗯。

16:15

So I've been working on labs for last five, six years when I started in 2021, it was like,

所以过去五六年我一直在做 labs，2021 年开始那会儿，感觉是，

16:20

it was much simpler because the kinds of things that the models had to do were like a prompt

那时候简单多了，因为模型要做的事情就像，一个 prompt 会是“帮我找市场上最好的电视”。

16:25

would be help me find the best TVs on the market.

然后你很容易就能 cross check 那个回答，也能理解它，判断我们训练模型用的数据好不好。

16:29

And then you can pretty easily cross check that response and understand it, the data that

现在真的是很复杂的环境。

16:34

we are training the model on is good or not.

所以很多，嗯，我们在 labs 里测试数据的方式也变得 agentified 了，就是

16:37

Now it's a really complex environments.

现在这真的是很复杂的环境。

16:40

So a lot of the, um, the way we test the data in labs has also become agentified, where

所以很多，嗯，我们在实验室里测试 data 的方式也变成 agentified 了，也就是

16:46

we would write a set of criteria that the data needs to meet, both on individual data sets,

我们会写一套数据需要满足的标准，既针对单个数据集，

16:53

but as well as on the diversity of data sets.

也针对数据集的多样性。

16:56

And what I love to do was to I would spend a full day just looking at one piece of data.

我特别喜欢做的一件事是，我会花一整天时间只看一条数据。

17:02

And there was no way that I could look at every single piece of data, but just looking

我不可能看完每一条数据，但仅仅看

17:06

at one piece of data told me a lot of things that Sahil was talking about.

一条数据就能让我明白很多Sahil提到的事情。

17:10

For instance, does it have a lot of synthetic data in it?

比如，里面有很多synthetic data吗？

17:14

Is the instruction following template ice where there are a lot of things where the model

是不是instruction following template，也就是说有很多地方模型

17:19

is being asked to generate something in 10,000 words for LinkedIn.

被要求为LinkedIn生成一万字的内容。

17:24

That doesn't make sense.

这说不通啊。

17:25

There's nothing on LinkedIn.

LinkedIn 上什么都没有。

17:26

That's 10,000 words, but we saw this in data all the time because it was generated almost

那是一万字，但我们在数据里经常看到这种东西，因为它几乎像是

17:31

like with a template behind the scenes and those things are very dangerous to post train

幕后用 template 生成出来的，而那些东西拿来 post train 模型非常危险

17:36

the model on because the way the way just like take a step back and talk about training,

因为，那个方式，那个方式，就像——退一步说，聊聊 training，

17:40

there's pre-training where we are just taking a lot of data and showing to the model.

有 pre-training，我们就是拿大量数据给模型看。

17:46

Then there's mid-training and post-training where increasingly models become aligned to

然后是 mid-training 和 post-training，模型会越来越对齐到

17:50

the tasks that they will do in the real world.

它们在真实世界里要做的任务上。

17:53

And if we show the models, the tasks that are misaligned, for instance, that have these

而如果我们给 models 看那些 misaligned 的 tasks，比如，这些 tasks 里有这些

17:59

templates in it or synthetically generated but are not high quality, then the models,

templates，或者是 synthetically generated 但质量不高的，那么 models，

18:05

you do the whole training run.

你把整个 training run 跑完。

18:06

As the last curves look great, the rewards in the training curve look great.

因为最后的 curves 看起来很棒，training curve 里的 rewards 看起来也很棒。

18:12

You take the final checkpoint and you start playing around with it and it makes no sense.

你拿最终的 checkpoint 开始各种试，结果完全说不通。

18:17

It's either thinking for way too long or it just like keeps on writing as someone who

它要么是思考得太久了，要么就是一直写，像一个

18:23

in real profession would never write.

在真实职业里绝不会那样写的人。

18:26

So as a result, when we look at the data, we are very concerned about its ability to actually

所以结果是，当我们看这些 data 时，我们非常担心它真正能够

18:30

train the model.

train 这个 model。

18:32

And how do we close the gap between the data is high quality and understandable, but can

还有就是，我们怎么弥合这个差距：data 质量高、也容易理解，但它能不能 train model？

18:37

it train the model?

这也是很多 lab 在看 data 时会重点找的东西。

18:38

And that's what like a lot of what labs look for when they're looking at the data.

所以我们还会做的另一件事，就是拿 model 来测一下，看看这个 model 到底能真正通过这些 data 多少次。

18:42

So other thing we would do is just like test it out on a model and see how many times

如果它太弱或者太强，那它就不在能让 model 从中学习的 Goldilocks zone 里。

18:48

is the model capable of actually passing the data.

model 是不是真的有能力通过这些 data。

18:51

And if it's too incapable or too capable, then it's not in the Goldilocks zone for the

而如果它能力太弱或太强，那它就不在 Goldilocks zone 里，对这个

18:56

model to learn from it.

model 来说，能从中学习。

18:58

Similarly, we would do ablations where we'll take a set of data, train the model a little

同样地，我们会做 ablations，拿一组 data，在上面把模型训练一点点，看看会有什么变化。

19:03

bit on it and see how things change.

也许拿一个中间的 checkpoint，然后开始 footing 它。

19:05

Maybe take an in between checkpoint and start up footing it.

另一件我们很想做的事是，把整个 data 拿来做 embed，然后挑出不同的 clusters，看看它实际上有多 diverse。

19:09

Another thing we would love to do is take the whole data, embed it, and then pick up

如果这是 finance data，它是不是都只是在讲一种 emanate？

19:15

different clusters to see actually its diverse.

还是在代表 wealth management？

19:19

Is it all talking about if it's finance data all just talking about one type of emanate?

它在代表 corporate tax。

19:24

Or is it representing wealth management?

还是说它代表的是财富管理？

19:26

It's representing corporate tax.

它代表的是企业税。

19:28

It's representing all the different things that we expected in the data.

它体现的是我们在 data 里预期会出现的各种不同情况。

19:31

So it's a very messy process.

所以这是个非常混乱的过程。

19:34

It's you're like trying to somehow figure out is this data good and it's all underway

就像你得想办法搞清楚这份 data 到底好不好，而这一切都还在进行中。

19:39

high time pressure because the time between collecting the data, posturing the model and

时间压力非常大，因为从收集 data、对 model 做 posturing，到

19:44

launching model is really compressed maybe a few months at the best.

发布 model，中间的时间被压得非常紧，最快可能也就几个月。

19:48

So we want partners who have shown consistent promise of delivering high quality data because

所以我们想要那种一直展现出能稳定交付高质量 data 的合作伙伴，因为

19:55

there's not very good ways to test the high quality data under the time pressure.

在时间压力下，并没有很好的办法去检验高质量 data。

20:01

And you were also asking a little bit more about you stuck about some of the post-training

而且你还多问了一点，你之前卡在一些 post-training 的事情上

20:04

efforts we do internally.

我们在内部做的那些工作。

20:05

Yeah.

对。

20:06

I mean, one of the things that I'm doing and handshake has been doing is taking all this

我是说，我自己在做的一件事，也是 handshake 一直在做的，就是把这些

20:11

data, post-training it internally to do two things.

data 拿来做内部的 post-training，做两件事。

20:15

One, would you post-training it on like open models and post-training it on our own

第一，你会不会把它在类似 open models 上做 post-training，也在我们自己的

20:20

data and the big thing you're seeing on the benchmark we have built and the data that

data 上做 post-training；而你在我们搭的 benchmark 和我们整理的数据上看到的关键点是，

20:23

we have put together, can the data help train some of these small open source models to

这些数据能不能帮助训练一些小的 open source models，让它们

20:28

climb this benchmark?

在这个 benchmark 上往上爬？

20:30

And there are two interesting things from that.

这里面有两个有意思的点。

20:32

One is that how is it finding or not climbing?

一个是，它是怎么发现问题的，还是说没有提升？

20:35

So for instance, if we see that the model has a visual task and a reasoning task associated

比如，如果我们看到 model 有一个 visual task 和一个相关的 reasoning task，而 model 没有提升，那是因为它没理解图像，还是因为它没理解该怎么去推理图像的描述？

20:41

with it and the model doesn't climb, is it because it failed to understand the image or is

基于我们在这次 post-training 里学到的东西，我们可以回去找数据团队，也就是它 lead 的那边，然后说，嘿，所有这些图像都得配上描述，或者所有这些 financial models 都得配上投资银行家的备注，说……

20:47

it that it failed to understand the reasoning on how to reason the description of the image?

是它没能理解如何推理图像描述的推理过程吗？

20:54

And based on what we learn on this post-training, we can go back to the data team that's

然后基于我们在这次 post-training 中学到的东西，我们可以回去找数据团队，那个

20:58

how it leads and say like, hey, everything that all these images need to come with the description

它是怎么引导的，然后说，嘿，所有这些图像都需要附带描述

21:04

or all these financial models need to come with notes from the investment banker saying

或者所有这些财务模型都需要附带投资银行家的备注，说

21:09

what this represents or maybe this tax law needs to come with like attached PDFs that just

这代表什么，或者也许这条税法得附带一些 PDF，就专门描述这条税法，因为 models 找不到这个，或者在 web 上会搞混，这样我们就能把需求单独隔离出来。

21:14

describe this tax law because the models cannot find this or they get confused on the web so we

所以这是 post-training 的一个很大部分，就是把这些关于 data 的洞察都收集起来。

21:19

can isolate the requirement. So that's one big part of post-training is to like get all these

然后另一部分是，如果我们把这一切整合起来，我们真的能帮 model 爬上去吗？

21:24

insights on data. And then the other part is that if we put all this together, can we actually

因为如果 model 真的爬上去了，那对 labs 以及那些想 fine tune 自己 models 的企业来说，就是一个非常强的 proof point；而且这些 data 确实好，因为 data 的目的就是训练 models。

21:31

help the model climb? Because if the model does climb, then it's very strong proof point to the labs

所以我们想把这个完整的 loop 跑通。

21:37

as well as the enterprises who are trying to fine tune their own models, but this data is actually

在进入 environments 之前，再快速问最后两个问题。

21:41

good because the purpose of data is to train models. So we try to complete the whole loop.

一个是，你们需要多少 data？

21:46

Like two last quick questions before we go into environments. One is how much data do you need?

就像，在我们进入 environments 之前还有两个最后的小问题。一个是你们需要多少数据？

21:52

So say there's a company that, you know, they're well established that here you should use your own

比如说有这么一家公司，你知道，他们已经发展得很成熟了，知道在这里你应该用你自己的

21:56

open models post-training them. They don't know how many environments, what's the process,

open models，然后对它们做 post-training。他们不知道要多少个 environment、流程是什么，

22:01

there's expertise. Fill in that gap a little and how much would it cost? Like how many environments,

这需要专业经验。稍微补一下这个缺口，成本会是多少？比如说要多少个 environment，

22:07

how much are they costing? So you're not frontier lab, but you want to do your own domain specific

它们要花多少钱？所以你不是 frontier lab，但你想做自己的 domain specific

22:12

model, right? Yeah, I would say one thing and then would love

model，对吧？嗯，我想说一点，然后也很想

22:16

sales opinion too. The amount of data really depends on the capability of model on that particular

听听销售的意见。data 的量真的取决于 model 在那项特定

22:22

task. Some models can do very good hill climbing on just 100 data points. If they are pristine data

task 上的能力。有些 model 只靠 100 个 data points 就能做很好的 hill climbing。如果它们是纯净的 data

22:32

points, they are not reward hackable. They are long horizon. They're exactly the shape of the

points，它们不是 reward hackable 的。它们是 long horizon。它们恰好就是……的形状

22:37

the task that they are trying to do. Then small amount of data can work really well and the model

他们试图做的那个任务。然后少量 data 就能效果很好，而且模型

22:42

can generalize from there. There are places where models are not at all good at. For example,

就能从那里 generalize。有些地方模型完全不擅长。比如，

22:47

life sciences that we'll talk about later today, the models have even the frontier models have

我们今天晚些时候会聊到的 life sciences，模型——甚至 frontier models——都

22:52

20% pass rate on that benchmark. In those cases, you need extensive data and life sciences is such a

在那个 benchmark 上只有 20% 的通过率。在那种情况下，你需要大量 data，而 life sciences 又是一个如此

22:59

broad field that doing lab essays and just doing blotting is very different than some other domain

广阔的领域，做 lab essays 和只做 blotting 跟其他一些 domain 非常不同

23:06

part of the reading lab essays. So the amount of data is required is both very high quality and high

也是读 lab essays 的一部分。所以所需的 data 量既要质量非常高，又要高

23:12

cost, but also very high volume. So it really depends on the problem that we are facing.

成本，而且量还要非常大。所以这真的取决于我们面对的问题。

23:19

That's about on. Yeah, totally depends. I think the other thing that's probably worth

差不多就是这样。对，完全取决于情况。我觉得另一件可能值得

23:26

calling out is I think producing data that is pristine is just really hard. So a lot of part of

我要说的是，我觉得要产出 pristine 的 data 真的很难。

23:33

the buying process that we'll hear from customers is we'll maybe use sub 10% of the data that we

所以我们在采购流程中经常从客户那儿听到的是，我们可能只会用到实际采购的 data 里不到 10%，因为整个流程是迭代的。

23:40

actually procure because we're an iterative process. I'm trying to figure out. Yeah, well maybe,

我还在试图弄清楚。嗯，也许吧，但我觉得这纯粹是质量的一部分。

23:46

but I think a pure part of the quality. Yeah, part of how they want to produce their model ultimately

对，他们最终想怎么产出自己的 model，一部分就是去试验那个 data mix。

23:53

is experimenting with that data mix. Certainly, there's things that you might just consider

当然，有些东西你可能直接就会觉得不好，但你可能有一条单独的 data，其实相当不错，只是在你 training 自己的 model 时，它对你想要的 mix 来说可能不是最优的。

23:58

bad, but you might have an individual unit of data that's actually pretty decent but might not be

那人们还会拿它做些什么吗？

24:05

optimal for the mix that you want when you're training your model. Is there still stuff people do

比如，你可以做 SFT mid-training，先开始 domain adaptation，

24:10

with that? For example, you can do SFT mid-training to just start domain adaptation,

用那个？比如，你可以做 SFT mid-training 来开始 domain adaptation，

24:15

even if it's not the most accurate, not the highest quality. There's some value in the data,

就算它不是最准确的，也不是质量最高的。这些数据还是有一些价值的，

24:20

or is it like, I just need to help climb post-training. I'm already a very strong model. I'm just

还是说，我只需要帮忙把 post-training 往上做。我已经是个很强的 model 了。我只是

24:25

trying to fill in gaps for people. I have two very interesting examples to maybe that will help

想帮人们填补一些空缺。我有两个很有意思的例子，也许能帮助

24:30

illustrate how this is very messy process. One is that let's imagine verifiable domain, like math.

说明这个过程有多乱。一个是，咱们想象一个 verifiable domain，比如数学。

24:38

One promise of verifiable domain is that you just do, you have very accurate rewards,

verifiable domain 的一个前景是，你只要做，就会有非常准确的 rewards，

24:45

in which case you can tell if the data is good or bad by easily, and then you can

在这种情况下，你很容易就能判断数据是好是坏，然后你就可以

24:50

help climb the model with no guard rails on that data. The challenge is the real world task

在没有 guard rails 的情况下帮助提升 model。挑战在于，真实世界的任务

24:56

don't come as pure math. They often come in as math or coding embedded to do some real world

并不是以纯数学的形式出现。它们往往是以数学或 coding 嵌入的形式，去做一些真实世界的

25:04

task. The grader gets often confused because it's looking for an answer, but within that answer,

task。grader 经常会搞混，因为它在找一个答案，但在那个答案里，

25:10

there has to be a lot of assumptions that the model has to make before it starts doing the math

在开始做数学之前，model 必须先做出很多假设

25:14

or coding to derive the answer. Now the grader has to both do semi-verifiable and verifiable in

或者写代码来推导出答案。现在 grader 必须同时处理 semi-verifiable 和 verifiable，在

25:22

the same task. If he saw large training runs fail to produce results because the model was very

同一个 task 里。如果他看到 large training runs 没能产出结果，是因为 model 非常

25:31

good at verifiable, but it ignored the fact that there was a lot of semi-verifiable component in

擅长 verifiable，但它忽略了里面同样也有很多 semi-verifiable 成分这个事实。

25:36

it as well. That was a one big learning. The other big learning is that every time you produce data,

那是一个很大的收获。另一个很大的收获是，每次你产出数据，

25:42

it comes in a different style. Humans, all humans write in a different style, but if you look at

它就会以不同的风格出现。人类，所有人的写作风格都不一样，但如果你看

25:47

models output, they're very stylistically consistent. When you read an opus's output versus

model 的 output，会发现它们的风格非常一致。当你读 opus 的 output，对比

25:54

Astra's output, they can kind of tell that there are two different models. As a result,

从 Astra 的输出里，他们大概能看出来这是两个不同的模型。

26:00

all the data that goes into the model has to be stylistically consistent. If they are stylistically

因此，所有进入模型的数据在风格上都必须一致。

26:06

inconsistent, then the model doesn't have its own brand. We had to do a lot of work across

如果风格上不一致，模型就没有自己的品牌。

26:14

whole this period to just make sure that the data is stylistically consistent. It's not even

整个这段时间我们做了很多工作，就是为了确保数据在风格上保持一致。

26:19

part of high-quality data, environments, rubrics, radars, which take up 90% of the time, but just

这甚至不属于高质量数据、environments、rubrics、radars 那部分——那些占了 90% 的时间——但光是拿这么多数据并让它们保持一致，就是问题的一大部分。

26:27

taking so much data and making it consistent is a big part of the problem. This comes top-down.

这是自上而下推动的。

26:33

For example, I want Cloud to sound like Cloud, and you fit that sort of description.

比如，我希望 Cloud 听起来像 Cloud，而你就符合那种描述。

26:41

Yeah, I think it's a very creative process because as you can see, everyone's iterating through

对，我觉得这是一个非常有创造性的过程，因为你看，每个人都在不断迭代

26:46

how should the models sound? Should it have M dashes or should it not have M dashes? Should it

模型听起来应该是什么样？

26:53

try to speak like a human or should it have this really compact and coded language?

应该用 M dashes，还是不用 M dashes？

26:58

So it's an evolving field, and the data companies evolve with how people

应该试着像人一样说话，还是应该用这种非常紧凑、编码化的语言？

27:04

every lab's constitution evolves. I don't think there's a one fixed answer to it, but what is

所以这是一个不断演变的领域，数据公司会随着人们和每个 lab 的 constitution 如何演变而演变。

27:10

definitely true is that the data needs to be consistent across finance and legal and all the

我不认为这有一个固定答案，但可以确定的是，数据在 finance、legal 以及所有涉及的领域之间必须保持一致，否则 post-training 会学到非常奇怪的格式和风格行为。

27:17

domains that come in, because otherwise, the post-training will learn very strange formatting and

所以还有很多工作要做，这是 sales 对此的评论，但在 100% 的数据中，很多数据都必须返工或重做，因为也许每个人都给了数据

27:21

style behaviors. So there's a lot of work to be done, and this was sales comment on, but

style behaviors。所以还有很多工作要做，而这是 sales 的评论，但

27:28

off the 100% data, so much of data has to be reworked or redone because maybe everyone gave data

基于这 100% 的 data，很多 data 都得返工或重做，因为也许每个人都提供了 data

27:35

but it's not Christine. Quality has been followed the stylistic highlights.

但这不是 Christine。质量一直跟着风格上的亮点走。

27:38

Yeah, one more, I think that's a really good articulation, and one thing that's

对，再补充一点，我觉得这个表述特别好，而且有一点是

27:46

I think we're starting to see a lot more of from labs is yes, you have these environments and

我觉得我们现在从各个 lab 那里开始看到越来越多的是：没错，你有了这些 environments，而且

27:51

these rubrics to try to help models learn certain capabilities, but alignment here goes

这些 rubrics，来试着帮模型学习某些能力，但这里的 alignment 已经

27:57

beyond tone and style. There's also like a certain like almost like genose qua for every domain when

超出了语气和风格。每个领域里也都有某种，像是，几乎像是 genose qua 的东西，当

28:05

it comes to what is great taste or what is great quality actually look like for that domain. That's

说到对于那个领域来说，什么叫好品味，或者什么叫真正好的质量。这

28:10

really hard to capture and encapsulate in a rubric. So we're actually starting to see an increase

真的很难在 rubric 里捕捉和概括出来。所以我们其实开始看到一种增长

28:17

in sort of parallel of preference ranking within these environments. It's still like really long

算是这些 environments 里并行的 preference ranking 的增长。它还是，真的挺长的

28:24

horizon tasks, still under within an environment, all of that, but hard to like codify a lot of these

horizon tasks，还是在一个 environment 里做，所有这些，但很多这类东西很难去 codify

28:31

things and so labs are both interested in figuring out how aligned is there greater or a greater we

所以 labs 都感兴趣的是搞清楚，到底有多 aligned，或者我们能做出多大程度上更 aligned 的东西

28:38

might produce to actual expert preferences. And second, to the extent that there's always

来贴合真实的专家偏好。第二，既然总会有

28:45

going to be some gap between that codification you could learn or just RL out through rubric

一些 gap，介于那种你可以学、或者直接通过 rubric 用 RL 搞出来的 codification，和纯专家的品味之间

28:52

and pure expert taste, can you train on those preferences as well? And I think we're

那你能不能也基于这些偏好来 train？而且我觉得我们

28:57

going to still I anywhere see a lot more of that too, which also leads to one other interesting

之后还是会——我不管在哪都会看到更多这种情况，这也引出一个另外挺有意思的

29:01

direction of travel that would be counter to it looking at the outside is that we just are hiring

发展方向，而这个方向跟从外面看是相反的：我们其实正在招

29:05

like far more full-time people like you know this this internally and you know today it's hundreds

多得多的人全职来做，你知道，这些都在内部，而且你知道现在已经是几百人了

29:10

in the future will be thousands thousands of like professionals who are basically doing this

未来会有成千上万、成千上万像专业人士一样的人，基本上全职在做这份工作，因为他们在为 Handshake 工作。

29:13

job like full-time because they're working for handshake. Yeah, exactly. In the full 500 right now we're

对，没错。

29:21

internal it internally 800 people 800-ish people growing quick, but just more so like the expertise

现在整个公司有500人，内部……其实内部是800人，大概800人，增长很快，但更关键的是所需的专业能力和品味水平。

29:28

required the level of taste. I mean we're trying to hire like the top you'll meet the team today

我的意思是，我们想招的是顶尖的，你今天会见到团队。

29:33

but like people that worked in extremely elite finance backgrounds, extremely early drug discovery

但像是那些在极其顶尖的金融背景工作过的人、极其早期的 drug discovery 背景的人，极其……你知道的，像那些真正能够管理和招聘品味制定者的人，因为 data 的质量会直接影响 model 的品味。

29:38

backgrounds, extremely you know like people that actually are able to manage and hire taste makers

这两件事都是，对吧，就像那些每天在产出 data 的人，他们本质上并不是在做

29:44

because the quality of data is directly influencing the taste of the model. It's both those things

因为 data 的质量会直接影响 model 的 taste。这两件事都是如此

29:48

right it's like people who are authoring the data day in day out who are essentially not doing

对吧，就像那些日复一日在 authoring data 的人，他们本质上并没有在做

29:54

anything else except working you know on on the platform to author the data and then on our like

除了在 platform 上做 data authoring 的工作，然后在我们这边，还有别的吗？当然，你们怎么 audit 那些 data？

30:01

side certainly how do you audit that data? How do you ensure that data is really good if you don't

如果你们不是正好有那个背景，你们怎么确保那些 data 真的很好？

30:06

come from that background exactly? So it's full-time kind of in both ways right and I think when

所以这算是两边都得全职投入，对吧？我觉得 Asham 提到过，从 lab 的视角看，你真的得确保拿到的 data 是高质量的、可以信任的，因为现在拿到坏 data 的机会成本实在高太多了。

30:11

Asham mentioned it's from the lab perspective you really need to make sure that the data you're

那会耗掉大量 researcher time，你知道，compute——现在 compute 和 researcher time 大概是 lab 里最宝贵的两样资源。

30:18

getting is high quality you can trust it because the opportunity costs now of getting bad data is

所以如果你不能信任这些 data，你就会花多得多的 compute 去，你知道，在 evalid 里针对那个 data 做 ablation，因为

30:23

just so much higher. It's a ton of researcher time it's you know compute is now you compute in

只是要高得多。这会消耗大量 researcher time；你知道，现在的 compute 就是你投入的 compute。

30:29

researcher time probably the two most valuable resources at a lab and so if you can't trust the

researcher time 大概是一个 lab 里最有价值的两种资源，所以如果你不能信任

30:33

data you're spending way more compute on trying to you know ablate against that data in evalid because

data，你就会花多得多的 compute 去试图，你知道，在 evalid 里针对那些 data 做 ablate，因为

30:39

the tasks are longer and longer horizon to takes longer to train and then you can't just open up

任务越来越长，horizon 越来越长，训练就更久，然后你就不能只是打开

30:43

a five-column CSV and look at the data and be like cool. You don't know that actually means it. Yeah

一个五列的 CSV，看看数据，然后说 cool。你其实不知道那到底意味着什么。对

30:48

you're getting a docker image with like you know hundreds of files that are about medicine and you're

你拿到一个 docker image，里面像你知道的，有几百个文件，都是关于医学的，然后你

30:53

like what am I looking at right? Okay so we got a we got a dive into the data let's look at some of

就想，我到底在看什么，对吧？好，那我们……我们深入看看数据，来看一些

30:59

that actual medicine you guys have vials your life sciences benchmark. Fun thing today we're

那些真正的医学内容，你们有 vials，你们的 life sciences benchmark。今天有个好玩的事，我们要

31:05

going to transition in so you've brought the people that actually worked on this I feel like we

过渡一下，所以你把真正做这个的人带来了，我感觉我们

31:10

could talk for hours just on the you know high-level side of things but I'm going to swap in we're

光是你知道的那些 high-level 的东西就能聊好几个小时，但我要换一下，我们

31:15

going to bring RJ and yeah we'll dive in and Nick and Jonas and you know fortunately we're the

要把 RJ 请上来，对，我们会深入聊，还有 Nick 和 Jonas，你知道，幸运的是我们就是

31:20

one AI part that also has a science team as well they will pair it nicely with this but thank you

有一个 AI 部门，同时也有 science team，他们会和这个很好地配合，不过谢谢

31:25

guys for joining us and thank you so much. Thanks guys. Let's go deep. Okay all right we're back with

各位来参加我们，非常感谢。谢谢大家。我们深入聊聊。好的，我们回来了，和

31:31

Nick and Jonas very exciting to have two people that worked on a domain specific benchmark vials

Nick 和 Jonas，非常高兴能请到两位参与过 domain specific benchmark vials 的人

31:38

is your work it's on the life sciences at a high level just do a quick intro how long you've been

这是你们的工作，它是在 life sciences 方向，先 high level 地快速介绍一下，你们来了多久

31:44

here is this your primary work your actual experts in the field so a little bit about that and then

在这里，这是你们的主要工作吗，你们是这个领域真正的专家，所以稍微讲讲这个，然后

31:50

let's go straight into it tell us what the what the benchmark is. Yep so nice to meet you I'm Nick

我们直接进入正题，告诉我们这个 benchmark 是什么。好，很高兴见到你，我是 Nick

31:56

magazine my background is in clinical vaccinology a lot of SARS-CoV-2 research and molecular biology

Magazine，我的背景是 clinical vaccinology，做过很多 SARS-CoV-2 research 和 molecular biology

32:02

and I run the STEM business at high including the life sciences of articles and I'm Jonas and I run

我在 high 负责 STEM business，包括 articles 的 life sciences 部分，我是 Jonas，我负责

32:08

a lot of our AI research on evaluating frontier AI and previously have worked in AI research for

我们有很多 AI research 是在 evaluating frontier AI，而且我之前在 AI research 领域做了大约 15 年，做过很多 computational biology、auto 和 metal 以及 data curation 的工作。

32:14

about 15 years doing a bunch of work in computational biology auto and metal and data curation.

酷。High 是 henchman 和 vials。

32:19

Cool. High is henchman and vials. Yep so at henchake one of the things that we find very

对，所以在 henchake，我们发现非常重要的事情之一就是科学，尤其是 life sciences。

32:30

important are the sciences in general in particular life sciences fundamentally the promise of AI

从根本上说，AI 的前景其实就是更好地促进 life sciences、工程、数学等领域的发现。

32:37

is really to better facilitate like discovery in life sciences in engineering mathematics the social

社会契约基本上是：我们承担 AI 带来的所有这些风险，而回报是我们能够推动创新的 hyper progression。

32:47

contract is basically that we take all these risks with AI and the reward is that we can

所以在我们看来，life sciences 在那里非常重要，因为它直接关系到人类健康、疾病研究、drug discovery，因此它是

32:53

facilitate like the hyper progression of innovation so life sciences in our opinion is very important

促进创新的 hyper progression，所以在我们看来 life sciences 非常重要

33:00

there because it ties directly to human health and disease research drug discovery therefore it is

因为这一点直接关系到人类健康、疾病研究、drug discovery，所以它……

33:06

the most valuable vertical in which that promise can exist and it looks like if you scroll down a

最有价值、也最能承载那个承诺的垂直领域，而且看起来如果你往下滚一点点

33:13

little bit it's one of the domains where models out of the box do very poorly right so your chart

就会发现，这是模型开箱即用表现非常差的一个领域，对吧，所以你旁边那张图表

33:19

on the side the minimax the GLM the GROC all sub 20% right yes so most most models out of the box

上，minimax、GLM、GROC 全都低于 20%，对吧，是的，所以大多数模型开箱即用

33:27

do fairly poorly all of them perform at 50% are worse even the very top of the frontier models

都表现相当差，它们全都只有 50% 或更差的表现，即便是最顶尖的 frontier models

33:33

which if you're a scientist is highly problematic I mean ultimately if you want to be doing any kind

如果你是个科学家，这就非常成问题。我的意思是，归根结底，如果你想做任何

33:39

biology or biochemistry research where human health is on the line anything under very near 100

人类健康攸关的生物学或生物化学研究，任何低于接近 100

33:45

percent would not be considered acceptable or feasible for use for an agent I guess I'm going

% 的表现，都不会被认为可接受，也不会被认为适合让一个 agent 使用。我想我

33:51

to take this time as the naive not bio person to ask my questions is this a combination of different

就趁这个机会，作为一个天真的、非生物背景的人来问我的问题：这是不是不同……的组合？

33:59

tasks is it one task what are the tasks you hear a lot about wet labs so how do you close the loop of

tasks，是一个任务吗？都有哪些 tasks？你经常听到很多关于 wet labs 的说法。那你怎么把闭环跑通——

34:08

using a model for sciences and then testing it and if you're running into simulation is it accurate

用 model 做科学，然后再去测试它；如果你跑的是 simulation，它准不准？

34:14

explain that to the non bio guy yeah so wet labs are a goal if you will for the state of AI right

给非 bio 的人解释一下。对，所以 wet labs 可以说是现在 AI 现状的一个目标，对吧

34:23

now so a wet lab is basically where somebody is validating physical research in the real world

现在，wet lab 基本上就是有人在现实世界里验证物理研究

34:29

based on a model cypothesis and just trying to gather empirical evidence based on things that

基于 model 提出的假设，然后只是试着去收集经验证据，基于那些

34:33

the model is developing one of the prerequisites to that is going to be just a strong physical

model 正在发展出来的东西。做到这一点的前提之一，就是得有很强的物理

34:40

understanding that these models develop which most if not all scientists have during their training

理解——这些 model 发展出来的理解，绝大多数甚至所有科学家在训练时都有

34:46

that allow them to actually process these experiments the atomic unit of these workflows so if you

这让他们能真正处理这些实验，也就是这些 workflows 的原子单元。所以如果你

34:54

think about an experiment it's more of a narrative than a specific point in time no medication

你想想一个实验，它更像是一个叙事，而不是某个具体时间点，也不涉及用药

35:02

any kind of clinical investigation is ever just like a singular test it's a series of tests that

任何临床研究都从来不只是单独一项测试，而是一系列测试，这些测试

35:08

gradually build towards conclusion or confirmation of your hypothesis and each of these units in

会逐步推进，最终形成结论或确认你的假设；而其中的每一个单元，在

35:14

life sciences and chemistry we call assays an assage is being like a singular measurement or like

生命科学和化学里，我们称之为 assay；一个 assay 就像一次单独的测量，或者像

35:20

small experiment where something is being validated or measured what this benchmark is testing

一个小实验，用来验证或测量某个东西。这个 benchmark 测试的是

35:26

is how well a model or an agent can actually interpret the outputs of each of these individual assays

一个 model 或 agent 到底能多好地解读每一个单独 assay 的输出，

35:33

which is the fundamental prerequisite of actually working in these lab and the loop environments

而这正是在这些 lab 和 loop 环境中真正开展工作、并

35:39

and processing all of these complex workflows that are required for real innovation in the field

处理该领域真正创新所需的所有这些复杂 workflows 的基本前提

35:44

and unfortunately the models aren't quite up to snuff yet but hopefully with our data

而且可惜的是，models 现在还不太达标，但希望靠着我们在 data 上的创新，在这个领域里我们最终能做到。

35:49

innovation in the field we'll get there eventually yeah so can we look at a few of the examples

对，那我们能看看你那边有的几个例子吗？

35:56

that you have there absolutely yeah maybe we can start with the molecular diagram there

当然可以。对，也许我们可以先从那边那个 molecular diagram 开始，就这个？

36:04

this guy yeah so just kind of rewind all these are different various lab assays that you would use

对，那就稍微往回倒一下。这些其实都是各种不同的 lab assays，你在 life sciences 场景里会用到。

36:11

in a life sciences setting these would be all kind of required for any basic clinical work or drug

这些基本上都是任何基础 clinical work 或 drug discovery 所需要的。

36:18

discovery just an understanding of the physical space that these inhabit and the diagram I believe

就是要理解这些所在的 physical space，而我相信你说的那个 diagram 就在这里。

36:24

you're referring to is here so one of the things you'll notice with a lot of these is these are very

所以你会注意到，很多东西的一个共同点是，它们非常 visual，而且从某种角度来说非常 physical。

36:30

visual and very physical in a way like these are 3D structures that are existing in a space so on

就像这些是存在于空间里的 3D structures，等等。

36:36

the left here this is a molecule and in this case it looks like it is a ligand or some kind of protein

左边这里，这是一个 molecule，在这个情况下它看起来像是一个 ligand 或者某种 protein

36:44

that exists within some kind of organism that's bound or unbound to a drug one of them is between

它存在于某种 organism 里，跟一个 drug 结合或未结合，其中一个是在

36:50

shown in red and one of them is being shown in green and the mostly GLP one right oh yeah

红色里显示的，另一个是用绿色显示的，而且主要是 GLP 的那个，对吧，哦对

37:00

and this is always epic yeah correct um yeah so so one of the things that people in like

而且这总是很 epic，对，没错，嗯对，所以，所以像人们会做的一件事就是

37:08

biochemistry and microbiology do all day when they're trying to develop drugs or just kind of figure

biochemistry 和 microbiology 的人整天在做的事，就是当他们试图开发 drugs 或者只是想弄清楚

37:13

out the mechanism of action for these sort of things is they're taking these biomolecules they're

这类东西的 mechanism of action，他们会把这些 biomolecules 拿过来，他们

37:19

spinning them around in all these various applications taking screenshots pictures passing them to

在各种不同的 applications 里把它们转来转去，截屏、拍照，然后传给

37:24

colleagues and trying to just do like basic determinations based on what they're seeing in the space

同事，然后只是试着基于他们在空间里看到的东西做一些基础判断

37:29

so here the models been given the very specific but also fairly straightforward task of identifying

所以这里 models 被给了一个非常具体但也相当直接的任务，就是去识别比如在一个 structure 里形成了多少 hydrogen bonds、而另一个 structure 里没有形成，这其实是你在研究一个药物时相当常规会做的事；右边这里我们可以看到答案，而在这个情况下，这个集合里所有的、所有的 data points 上，models 总体来看在不到 50% 的情况下能做对，所以这是 models 目前实际上从根本上无法以所需的一致性做到的事；所以在这个例子里，我们会认为，如果你把这个 prompt 和 image 给到 models，它并不能答对，对，至少不能稳定答对；而且对，有些 models 可能答对，有些可能答错，这取决于

37:37

like how many hydrogen bonds are being formed in one structure but not another which is something

比如在一个结构里形成了多少 hydrogen bonds，而另一个结构里没有，这件事……

37:42

that you would do fairly routinely whenever investigating a drug and on the right here we can see

你在研究一个 drug 的时候其实会相当常规地去做，而右边这里我们能看到……

37:48

the answer and in this case all of the all of the data points within the set the models are performing

答案；在这个例子里，这个集合里所有、所有数据点上，models 的表现……

37:54

under 50% of the time as an aggregate so it's something that the models are actually fundamentally

总体来看只有不到 50% 的时候是对的，所以这件事其实是 models 目前根本上……

37:59

incapable of doing at the moment with the need degree of consistency so for this example we would

做不到的，也达不到所需要的一致性程度，所以在这个例子里我们会……

38:06

assume if you put this prompt and image into the models like it's not getting it right yeah at

假设你把这样的 prompt 和 image 输入给 models，它们就是搞不对，对吧，至少……

38:13

least not consistently and yeah some models may get it right some may get it wrong it depends on

不是稳定地做对；而且，有些 models 可能做对，有些可能做错，这取决于……

38:18

you know how the model is weighted some of them have you know strengths and specific subdomains are

你知道 model 是怎么 weighted 的嘛，有些 model 有你知道的优势，特定的 subdomains 也

38:23

on data types that the researchers that those labs have trained on more but across the board we

是在那些实验室的研究人员训练得更多的 data types 上，但总体来看，我们

38:29

would assume that this would perform less than half the time accurately and I guess I would call

会认为它在不到一半的时间里是准确的，而且我猜我会说

38:36

especially for the listeners that there's these two overlapping molecules are very close to each

尤其对听众来说，有这两个重叠的 molecules 彼此非常

38:42

other they're somewhat and then like maybe if you look in the very center there's this like 3.4

靠近，它们有点……然后就像，也许如果你看正中心，有个大概 3.4

38:47

angstrom and a little bond there and it binds on one and not the other the one that goes to the

angstrom，还有一个小小的 bond，它会结合到其中一个上，而不是另一个上，就是往左的那个

38:54

left and so being able to pick that out from all the visual noise you can imagine how that would be

所以能从所有视觉噪音中分辨出这一点，你可以想象那会

39:02

very difficult for a vision language model yep exactly and for anybody who's listening to the

对 vision language model 来说非常困难，对，就是这样，而且对任何正在听这个的

39:08

angstroms are like a unit of measurement at the molecular scale I know it's not a common term

angstroms 是分子尺度上的一种测量单位，我知道这不是个常见词

39:14

I guess at a high level what grade or so how what's the level of difficulty of these tasks is

我猜，从高层次来说，这些任务大概是什么年级、或者说难度水平是怎样的

39:19

this stuff that you would learn in like high school sciences college undergrad masters PhD where

这些内容大概是你会在高中科学、大学本科、硕士、PhD 哪儿学到的东西，那

39:26

where does this range so the topics themselves are fundamentally like upper undergraduate graduate

这个范围到底在哪，所以这些主题本身基本上就像是高年级本科、研究生

39:31

level but actual training in these varies I mean I've had interns in high school that would do this

水平，但实际在这些方面的 training 因人而异，我的意思是，我有的高中实习生也会做这种

39:38

kind of like very basic investigation with like very minimal training fundamentally the tasks aren't

算是非常基础的调查，几乎只需要极少的 training，从根本上说这些任务并不是

39:43

necessarily hard you just need to know what to look for and then be able to do it just accurately

一定很难，你只需要知道该看什么，然后能够准确地

39:47

and consistently and anybody who's ever run a lab most of these assays again are things that like

且一致地做出来，而且任何管过实验室的人，这些 assays 大多数，同样，都是那种

39:53

lab techs do interns interns will do not necessarily people with advanced degrees so the the skill is

lab techs 会做，实习生也会做，不一定是高学历的人，所以这个技能

40:00

really in just like consistently doing a fairly basic interpretation versus knowing something that's

其实就在于像这样持续做一个相当基础的解读，而不是知道某个

40:07

very complex and high level and we make these benchmarks with experts from the pharmaceutical

非常复杂、非常高级的东西；而且我们会和制药行业的专家一起做这些 benchmarks，

40:12

industry and yeah we can measure how long they take to do these tasks and it's typically like

对，我们可以测他们做这些任务要多久，通常大概

40:16

under a minute so it's pretty quick for somebody right can we let's look at another example sure

不到一分钟，所以对某个人来说挺快的，对吧？我们再看另一个例子，好

40:24

about this one yeah so what are we looking at here yeah so on the left we have a plasmid map a

这个吗？好，那我们这里看的是什么？好，左边有一个 plasmid map，一个

40:31

plasmid map is basically a big ring of DNA that's used to encode various proteins when biologists are

plasmid map 基本上就是一大圈 DNA，用来编码各种 proteins，当生物学家

40:38

trying to make something in a lab and the question itself is asking within this plasmid this kind of

想在实验室里做出某样东西时，而问题本身是在问在这个 plasmid 里这种

40:45

give me the percentage of a sequence that belongs to one set of genes versus another and then just

给我一个序列中属于某一组基因相对于另一组基因的百分比，然后就让它运行，把那个四舍五入到一位小数。

40:52

asking it to run to round that to one decimal place so this actually brings up a question from

所以这其实引出了你提出的一个问题。

40:58

you okay so first of all for the listeners this there's a very complicated circular diagram with

好，首先对听众来说，这里有一个非常复杂的圆形图，上面有很多数字、标签和颜色，你能想象这是为什么。

41:05

lots of numbers and labels and colors and you can imagine why also this one would be very

而且这个图对 VLM 来说会非常难解读，真的很难解读。

41:11

difficult for a VLM to interpret it to interpret but maybe backing up like why do I care if it can

但也许退一步讲，我为什么会在意它能不能——或者说需不需要——它是否能够以这种形式引入信息，对吧？

41:23

it need if it's able to bring in information in this form right like you could imagine just having

就像你可以想象，直接有一个像 CSV 这样的数值文件，或者别的东西，里面包含同样的信息。

41:29

a numerical like a CSV or something that has the same information why do we care that it can

那我们为什么会在意它能解读这个呢？

41:35

interpret this yeah so when scientists communicate with each other they use all of these methods

对，所以当科学家彼此交流时，他们会使用所有这些方法。

41:40

when I would you know if I were to deliver this data to a colleague I would likely use both but

换作是我，你知道，如果我要把这份 data 交给同事，我可能会两个都用，但

41:45

in a practical setting people interacting with these agents tend to want to give them minimal

在实际场景中，跟这些 agents 交互的人往往想给它们尽可能少的

41:51

information or give them just enough information that should be able to lead to the result that

信息，或者只给刚好够的信息，这些信息应该能导向他们

41:57

they're looking for in this case on the left you know this would be considered more than sufficient

想要的结果。在这个例子里，左边这个，你知道，会被认为有足够多

42:02

data somebody who's an expert in the field would be able to do this with no issue I think it's

data。这个领域里的专家做这件事完全没问题。我觉得

42:06

worth noting that the when the model actually gets this image it's quite a bit higher scale and

值得一提的是，当 model 实际拿到这张图时，它的 scale 要高不少，而且

42:11

visualization so it's fairly straightforward to go across but to Jonas's point this is something

visualization 也更强，所以扫过去相当直接；但按 Jonas 的观点，这件事

42:18

that practically speaking shouldn't take an expert more than a few minutes to actually parse out

实际来说，专家真要把它 parse 出来，不该花超过几分钟。

42:23

if they were to go through the question yeah and I would just add that in my work I often find that

如果他们要走这个问题的话，是的，我只想补充一点，在我的工作中我经常发现

42:28

visualizing data brings out information that is not apparent from the from the quantitative data

可视化数据能揭示出从定量数据中看不出来的信息

42:37

and it often is actually not apparent to the model as well so that you can imagine that

而且这些信息往往对model来说也不明显，所以你可以想象

42:44

learning how to interpret a diagram like this you might be able to spot certain structure in it

学习如何解读这样的图表，你可能会发现其中某些结构

42:51

that you wouldn't necessarily be able to spot yeah and our focus at hand shake is really how can we

而你不一定能发现。是的，我们在Handshake的重点真的是我们如何

42:55

build AI agents that are more economically valuable meaning more useful to real professionals

构建更具经济价值的AI agents，也就是对真实专业人士更有用

43:00

and when we ask scientists you know who practice and biotech and pharmaceutical companies what do they

当我们问科学家，你知道，在biotech和pharmaceutical公司从业的科学家，他们到底

43:05

actually want how do they want to use AI agents in their workflows like you know they don't want to

想要什么，他们希望如何在他们的workflows中使用AI agents，你知道，他们不想

43:10

reformat all the data and all the information they want to have the AI agent come into their workflow

把所有数据和所有信息都 reformat，他们想让 AI agent 进入他们的 workflow

43:15

as it exists in the native way and just have it start you know producing useful analyses and insights

就按它原生的方式存在，然后让它开始，你知道，产出有用的分析和 insights

43:20

right there yeah so drop in a like a image or drop in on PDF or something like that and say

就在那儿，对，所以你就丢一张图进去，或者丢一个 PDF 之类的进去，然后说

43:26

like to explain this fact or like something like that yeah yeah so actually speaking it's a

比如解释一下这个事实，或者类似这样的话，对对，所以其实说起来，这是一个

43:31

it's a labor handoff right if I were to format this into a table and get like all the data I you

这其实是一种 labor handoff，对吧，如果我要把这个整理成一张表，然后把所有数据都拿到，我，你

43:37

know I've just done the work as opposed to having it handed off to the agent so what's the closing

懂的，我就是自己把活干了，而不是把它交给 agent 去做，所以 closing

43:42

the loop so you work on the evil side and the data curation is this a benchmark that we want to

the loop 到底是什么，所以你在 evil side 做 data curation，这是一个我们想

43:49

help climb on like can you generate the environments the whole setup and do you want to train against

帮忙爬上去的 benchmark 吗，比如你能不能生成 environments、整套 setup，以及你想不想 train against

43:58

the benchmark do you do it you know held out what's what's the next step of handshaped have you

这个 benchmark 你们会做吗？你懂的，held out 的那种，handshaped 的下一步是什么？你们有没有

44:04

worked on solving this yeah so we do a lot of training data like selling training data to make

着手解决这个问题？对，所以我们做很多 training data，比如卖 training data，来让

44:10

models better at these economically valuable tasks and so yeah we think this benchmark really

模型在这些有经济价值的任务上表现更好，所以是的，我们认为这个 benchmark 真的

44:16

represents something that is critical for models to be useful in any kind of drug discovery you know

代表了某种关键的东西，让模型在任何 drug discovery 里都有用，你知道

44:20

AI is not going to cure cancer before it can interpret these very basic types of images and so to

AI 在能解读这些非常基础类型的图像之前，是治不了癌症的，所以为了

44:26

actually now help make that a reality we need to generate thousands of training data points

真正帮助把这变成现实，我们需要生成成千上万个 training data points

44:31

and to validate that these are actually useful training data we post train a lot of models in

并且为了验证这些确实是有用的 training data，我们会在内部 post train 很多模型

44:36

house and adjust the types of reward signals and metadata so the good thing about training data

并调整 reward signals 和 metadata 的类型，所以 training data 的好处是

44:40

as it can be easier than you know the real task the real deal where we give the models hints we

因为这会比，你知道，真正的任务、真正要做的事更容易，也就是我们给模型一些提示，我们

44:45

can frame some of these questions as multiple choice questions so yeah we try training like

可以把其中一些问题设计成选择题，所以对，我们试着训练像

44:50

Quinn models Kimi models Nemotron and we really struggle if we just do direct RL like JRPO

Quinn 模型、Kimi 模型、Nemotron，而且如果我们只是做像 JRPO 这样的 direct RL，我们真的会很吃力

44:56

training against you know these answers as is and that's because like as you mentioned earlier

直接拿这些答案按原样去 training，你知道，就是这样，那是因为，就像你之前提到的

45:01

there's just it's not a problem of visual understanding it's the process of how to break down

这根本不是 visual understanding 的问题，而是如何拆解

45:06

the answers just not in there so blindly JRPOing you're not going to get a roll out that has it

答案的过程，它根本不在里面，所以盲目地做 JRPO，你不会得到一个把它做对的 rollout

45:11

correct exactly you want diverse environments that guide it through better understanding of bio

正确。没错，你想要多样化的环境，引导它更好地理解 bio

45:18

and as a subsequent result it will start to generalize on this yes so yeah we want to build

然后作为结果，它就会开始在这一点上 generalize，是的，所以对，我们想要构建

45:24

we want to first understand the capabilities like where learning is stalling so there is some

我们想先理解一下这些能力，比如 learning 到底卡在哪里，所以会有一些

45:28

learning of just like learning to format the outputs and you know the short single answer format

learning，就比如学会怎么格式化 outputs，还有你知道的简短单一答案格式

45:33

and things like that but then learning quickly plateaus and so to understand where it's stalling

之类的，但然后 learning 很快就 plateau 了，所以为了理解它卡在哪里

45:38

like a big source of stalling is just in the vision encoder it lacks like a understanding of what

比如说一个很大的卡点其实就在 vision encoder 里，它缺乏一种理解，就是

45:45

this image is in the context of this biological question that has a lot of you know domain specific

这张图在这个生物学问题的语境里到底是什么，而这个问题有很多你知道的 domain-specific

45:50

terms that may be mapped to what is the red stuff in this image you know needs to be translated

术语，它们可能要 map 到“这张图里红色的东西是什么”，你知道，需要被翻译

45:55

from some four syllable word in the question and so one thing we do as a like a shortcut to help

从问题里某个四音节词翻译过来，所以我们做的一件事，就像一种捷径，来帮助

46:01

models is have them generate descriptions of the image that are contextualized on the question

模型，就是让它们生成对图像的描述，而这些描述是以问题为 context 来做 contextualization 的

46:06

and then we ask frontier models to solve the task using just the image description as an example

然后我们让 frontier models 仅使用图像描述来解决这个任务，作为一个 auxiliary type reward 的例子，这实际上只是在测试模型对图像的全局理解，以及那个 vision encoder 实际工作得有多好，所以那个，所以 vision encoder，如果我理解得对的话，不好意思，vision encoder，所以你生成一个文本描述，然后只基于这个文本描述做 inference，那么像这样的情况，那个，差距在哪里呢？是的，这几乎就像一个 baseline 或者 control，对吧？是的，所以我们发现，你知道，当我们确实使用这些模型时，它们生成的文本描述就是明显遗漏了一些关键细节，而这些细节就像是图像中最显著的东西，是你所需要的。

46:12

of an auxiliary type reward that really is just testing the model's global understanding of the

一种 auxiliary type reward，它其实只是在测试 model 对图像的 global understanding，以及那个 vision encoder 实际工作得怎么样。

46:17

image and how well that vision encoder is actually working so the so the vision encoder if I

所以那个 vision encoder——如果我理解得没错，抱歉——是 vision encoder，也就是说你先生成一个 textual description，然后只基于这个 textual description 做 inference。

46:24

understand this excuse me the vision encoder so you generate a textual description and then

所以像这种情况，问题在哪儿？gap 在哪儿？

46:31

only inference based on the textual description so what is the like for something like this what

对，这几乎就像一个 baseline 或者 control，对吧？

46:36

is the gap there yeah it's almost like a baseline or control right yeah so we find you know when

对，所以我们发现，你知道，当我们真的用这些 models 的时候，它们生成的 textual descriptions 就是会漏掉非常明显的关键细节，而那些细节是图像里最显著的东西，是你需要非常清楚地说出来的：这是一个产品，我们从蓝色换成了紫色，我们到底在……

46:41

we do use these models like the textual descriptions that they generate are just missing very

46:46

obviously key details that are like the most salient thing in the image that you would need to

46:51

answer the question and so you know if you describe the image like this image this plasmid image

回答这个问题，然后你知道，如果你把这张图像——这张 plasmid 图像——这样描述

46:58

as a series of rings and you don't you know quantify like the percentage of different aspects of

把它描述成一系列环，而你又没有，你知道，恰当地量化这些环不同方面的百分比

47:04

these rings uh properly then there's just no way to do this task from the image description at that

那到那个时候，就根本没法根据图像描述来完成这个任务

47:10

point and so you fundamentally just totally lost information in how you've represented the image as

所以你就从根本上完全丢失了信息，在你把图像表示成

47:15

a latent vector okay so Nick you prepared an example I think this is really helpful for uh to

一个 latent vector 的方式里，好吧。所以 Nick，你准备了一个例子，我觉得这真的很有帮助，呃，用来

47:24

understand what is the language the vision language model or whatever I'm actually struggling with

理解语言到底是什么——vision language model 还是什么——我其实一直搞不太清楚

47:31

and why is this how does a person actually go about interpreting all these blue lines yeah so

以及为什么会这样，一个人到底该怎么去解读所有这些蓝线，对，所以

47:40

on the right here we have what's known as a western blot if you want to I'm sorry a gel if you

在右边这里，我们有所谓的 western blot——如果你想的话——抱歉，一个 gel，如果你

47:45

want to think about this as like a physical object it's basically just a sheet of gelatin a scientist

想把这个当成一个实物来看的话，它基本上就是一张明胶，一位科学家

47:52

loads protein at the top so at the make very top of the y-axis and then the protein just kind of floats

会在顶部加载 protein，也就是在 y-axis 的最上方，然后 protein 就有点像往下漂

47:59

down small things float down quicker um larger things sort of hang at the top when you apply electrical

小的东西往下漂得更快，嗯，大一点的东西就有点悬在顶部，当你加上 electrical

48:04

current so in this case the scientists asked something very complicated um you don't need to worry

current 时，所以在这个情况下，科学家们问了一个非常复杂的问题，嗯，你不需要太担心

48:10

about this too much but basically what they're doing is they're just giving a setup and giving

这个，但基本上他们做的就是给出一个 setup，再给

48:13

some kind of experimental details basically on the left here um we can see this uh thing that looks

一些实验细节，基本上在左边这里，嗯，我们可以看到这个，呃，看起来

48:19

like a ladder this is a standard and then all of their samples are here to the right what they're

像 ladder 的东西，这是一个 standard，然后他们所有的 samples 都在右边，他们

48:23

basically saying is hey we have um a total of eight lanes here and then within those lanes we have

基本上是在说，嘿，我们这里总共有八个 lanes，然后在这些 lanes 里面我们有

48:29

treated each of these um lanes with like a different dose of some molecule and they're just trying

他们给这些，嗯，lane 分别处理了像是不同剂量的某种分子，然后他们只是想

48:35

to figure out um what changes are happening within these lanes and at what time point uh they're

搞清楚，嗯，这些 lane 里正在发生什么变化，以及在哪个时间点，呃，他们

48:41

starting to see um these changes occur and if you'll see they go from left to right from low

开始看到，嗯，这些变化出现；而且你会看到，它们是从左到右、从低

48:46

concentration to high concentration so another part of the setup here is they basically say hey

浓度到高浓度。所以这个 setup 的另一部分是，他们基本上是说，嘿

48:52

these are the sizes of each of the things in this lane so basically just focus on these three

这些是这条 lane 里每个东西的大小，所以基本上只要关注这三条

48:56

lanes and then once you focus on these three lanes the um question becomes quite simple it's basically

lane，然后一旦你关注这三条 lane，嗯，问题就变得很简单，基本上

49:02

just saying at what's lane do we first see a band in like all three of these pockets um and again

就是在说，我们在哪条 lane 里第一次看到，像在这三个区域里都有 band，嗯，然后再说

49:10

this is kind of something that you see with a lot of these tasks they look very noisy and very

这是你在很多这类任务里会看到的情况，它们看起来非常嘈杂，而且非常

49:14

complicated but when you fundamentally understand just what the question is asking they simplify

很复杂，但当你从根本上理解这道题到底在问什么，它们就会变简单

49:18

quite quickly um and basically they just boil down to uh pay attention to this element of this larger

而且很快，嗯，基本上它们就归结为，呃，注意这个更大的

49:25

diagram and then just give me some very basic analysis of what's occurring within that element and come

图表里的这个元素，然后只给我一些关于这个元素内部正在发生什么的最基础分析，并得出

49:30

to some very basic conclusion so in this case we can very quickly see that there is a band in all three

一个非常基础的结论。所以在这个例子里，我们很快就能看到，在第三条 lane 的那三个

49:37

of those like brackets of this third lane that says one microgram so the answer is one and there's

像括号一样的区域里，都有一条 band，上面写着 one microgram，所以答案是一，而且这里也有一点

49:43

a bit of human taste of this too so um if we go back um we can see that this bottom band at 20

人的判断成分，所以嗯，如果我们往回看，嗯，我们可以看到这条底部 band 在 20

49:51

to some might look like low or it may look like a smudge but if you were to ask a scientist this is

对某些人来说可能看起来像 low，或者看起来像一团模糊，但如果你去问科学家，这

49:55

very firmly a band so this is kind of the regions where we see um these models fail there's uh you

非常明确地就是一条 band，所以这些差不多就是我们会看到，嗯，这些模型失败的区域，那里有，呃，你

50:03

know a human taste element there's an attention element and then there's also just like a lot of

你知道，有一个人类审美元素，有一个 attention 元素，然后还有一大堆

50:07

noise in the visual image so um these three factors compound to make a question that is relatively

视觉图像里的噪声，所以呃，这三个因素叠加起来，会让一个对科学家来说相对

50:13

simple for a scientist quite difficult for a model and the distinction there would be if the model

简单的问题，对 model 来说却相当困难；而这里的区别在于，如果 model

50:19

doesn't interpret lane one as 20 is actually being something the answer it would go up is five

不把 lane one 解读成 20、而是把它当成别的某个东西，那答案就会是五

50:24

because that's where it clearly is there so yeah so if we look at opus five um yeah so

因为五明明就在那儿，所以对，如果我们看 opus five，呃，对，所以

50:30

to your point yeah you should be a scientist the one with the model yeah but uh so worth noting

就像你说的，对，你该当科学家，那个拿着 model 的科学家，对，但是呃，所以值得一提的是

50:37

you can't run this in opus 5.5 because of the security guard so it's run on opus five um but to

你在 opus 5.5 里跑不了这个，因为 security guard，所以它是在 opus five 上跑的，呃，但就像

50:42

your point yeah it answers five because it basically doesn't see that um band I've linked on the

你说的，对，它回答五，因为它基本上看不到那个，呃，我链接在……上的那个 band

50:49

20 bar to be an actual band um again there's a human taste element to this um you know a scientist

20 bar 才能算是一个真正的乐队，嗯，再说一次，这里面有个很人类审美的元素，嗯，你知道，科学家会非常明确地说，这是一个产品，我们把颜色从蓝色换成了紫色，我们到底在看什么？这个任务是什么？对，所以在这个任务里，这基本上是一张 medical astrology 图像，所以如果你，嗯，曾经在医生那里做过 biopsy，这实际上就是他们把它交给某个人，这个人在左边操作这个，然后只是看各种像是图像，来判断你是不是有某种 pathology，或者 tissue 是不是在某种程度上受损了，所以在右边这里，我们大概有，嗯，背景，很基本地说，这是一个，嗯，brain tumor，我们只是在找 pathology 的程度，而这里的问题实际上就是找出那两张 slides，嗯

50:56

would say that very clearly this is a product we've swapped from blue to purple what are we

51:02

what are we seeing here what is this task yeah so in this task this is basically a medical

我们这里看到的是什么？这个任务是什么？

51:06

astrology image so if you've um ever gotten a biopsy at a doctor this is effectively they hand it

嗯，所以在这个任务里，这基本上是一张 medical astrology image。

51:11

off to somebody who runs this on the left and is just looking at various like images to determine

所以如果你在医生那儿做过 biopsy，这其实就是他们把它交给一个人，那个人在左边跑这个，然后就看各种图像，来判断你是不是有某种 pathology，或者 tissue 是不是以某种方式受损了。

51:16

whether or not you have some kind of pathology or the tissue is damaged in some way so on the right

右边这里呢，大致就是，嗯，背景；非常基本地说，这是一个 brain tumor，我们只是在看 pathology 的程度。

51:22

here we have kind of the um background very basically this is a um brain tumor and we're just looking

这里的问题实际上就是找出那两张低于 60% 的 slides。

51:29

for the degree of pathology and the question here is effectively just to find the two slides um

所以是啊，这还远远没到可用的程度，如果你是一个做真实研究的科学家，基于这些 assays，你没法信任一个 agent 的结果。

51:35

of these nine that have like the most advanced uh disease state so um the scientist here gives

在这九个里，那些差不多是最晚期 disease state 的，所以呃，这里的科学家给出

51:43

a setup basically you can label these uh respectively and the question basically just boils down to

一个 setup，基本上你可以分别给这些 label，而问题基本上就归结为

51:49

find the slide that has none of these three purple dots of rich pink or to find white gaps so

找到没有这三个紫色点、rich pink 的 slide，或者去找 white gaps，所以

51:55

the purple dots are basically representative of a live cells um if we eliminate those we can find

这些紫色点基本上代表 live cells，嗯，如果我们排除掉这些，我们就能发现

52:02

that we are left with the six the rich pink we can see in the bottom left and also an f if we

我们就剩下那六个，左下角能看到的 rich pink，还有一个 f，如果我们

52:09

eliminate those um we're left with these four and then if we look for the defined white gaps

排除掉那些，嗯，我们就剩下这四个，然后如果我们去找那些明确的 white gaps

52:16

you can see kind of these like vesicles here these are these um spaces that you see where things

你能看到这里有点像 vesicles 的东西，这些就是这些，嗯，你看到的那些空间，那里东西

52:21

aren't very scattered and if we eliminate those we're left with these two slides so for a trained

不怎么分散，如果我们排除掉那些，我们就剩下这两张 slides，所以对于一个训练过的

52:27

histopathologist this is a very simple task and uh basically the answer here would just be H&I

histopathologist，这是个非常简单的任务，呃，基本上这里的答案就应该是 H&I

52:34

when the models do this uh they come to this answer so again this is opus five five um it

当模型做这个的时候，呃，它们会得出这个答案，所以再说一次，这是 opus five five，嗯，它

52:40

identifies G H&I um even at a cursory glance you can see that G is kind of the odd man out among

识别出 G、H&I，嗯，即使只是粗略看一眼，你也能看出 G 在这三个里面算是那个异类，在

52:46

these three um and even at a very basic level if you were to look at these um you wouldn't consider

这三个里面，嗯，而且就算在非常基础的层面上，如果你看这些，嗯，你也不会认为

52:51

this to be advanced necrosis so the implications here are like pretty obvious um you know if this

这是 advanced necrosis，所以这里的影响其实挺明显的，嗯，你知道，如果这

52:56

were to be a diagnostic test and you were to you know treat a patient based on the results um you

是一个 diagnostic test，而且你根据结果去，你知道，治疗一个病人，嗯，你

53:02

know you'd be actively damaging the person uh who falls in panel panel G where um H&I would be

知道，你就会实际上伤害那个，呃，被分到 panel panel G 的人，而那里，嗯，H&I 才应该是

53:09

correct so uh very real implications to getting these kind of things wrong uh quick question that

正确的，所以，呃，把这些事情弄错会有非常现实的影响，呃，快速问一句，那个

53:14

pops up here is do models often have incorrect answers and consistency so with uh Gemini or

这里冒出来的问题是：models 是不是经常会有错误答案和 consistency 问题，那用 uh Gemini 还是

53:21

Open AI or whatever also put G or do they all have their own different it varies so on a question

Open AI 或随便什么，也加上 G，还是说它们各自都不一样？这得看情况。所以面对一个

53:29

like this it would vary quite a bit on other ones uh it would be lestiverse so to your point on the

像这样的问题，它会有挺大变化；在其他问题上 uh 它会没那么 diverse。所以按你说的，关于

53:34

previous uh example um you know we we looked at kind of this band that appeared on five and was

之前的那个 uh 例子，嗯你知道，我们我们看了那种出现在 five 上的 band，而且

53:40

less clear on lane one um that sort of thing would be relatively consistent here where the image is

在 lane one 上就没那么清楚，嗯，那种情况在这里会相对 consistent，因为图像

53:46

a bit more noisy and there's a bit more nuance the actual analysis I would expect you to be quite

有点更 noisy，也有更多细微差别；实际分析我觉得你会

53:50

a bit more diverse just because the artifacts that we're actually viewing um themselves are quite

diverse 得多，只是因为我们现在实际看的这些 artifacts，嗯，本身就很

53:55

diverse and the models often screw up translating the biological terminology into this purple pink white

diverse，而且 models 经常会把生物学术语转成这种紫、粉、白的时候搞砸

54:02

inconsistent ways and then there's also you know like how pink like what really is pink um that's

而且方式还不一致，然后还有，你知道，比如粉色到底怎么算，什么才算真正的粉色，嗯，这

54:08

a question of again you have to translate that biology knowledge into this common speak and the

又是个问题，你还是得把那些生物学知识翻译成这种通俗说法，而

54:14

models uh don't do very well in consistent ways there on thresholds and I'm assuming part of

models 嗯，在 thresholds 上没法用一致的方式做得很好，而且我猜一部分

54:19

the expertise that handshake has is expert graders that can differentiate what rich pink means to me

handshake 的专业能力就在于有 expert graders，他们能区分浓郁的粉色对我来说到底意味着什么

54:27

this is all beautiful rich pink uh I think you know uh if you were to actually like the one

这全是漂亮的浓郁粉色，呃，我觉得你知道，呃，如果你真的像那个

54:33

I struggle with not having seen this before is the third one the defined white gaps right

我之前没见过的、让我纠结的，是第三个，那个定义好的白色空隙，对吧

54:38

when we look at B H and I see white gaps everywhere I think if I really think about it I would probably

当我们看 B H 的时候，我到处都能看到白色空隙，我觉得如果我认真想想，我大概

54:45

get this right but um the value is domain experts that can label this with consistency accuracy and

能把这个做对，但嗯，价值在于 domain experts 能一致、准确地标注这个，而且

54:54

harder questions than this yeah and the experts are very important so to your point about like knowing

比这更难的问题，是啊，而且专家非常重要，所以说到你那个点，比如怎么知道

54:59

which uh you know what how pink is too pink um these are things that are passed down almost by

哪个，呃，你知道，粉到什么程度算太粉，嗯，这些东西几乎是靠

55:05

oral tradition by scientists so when you look at a paper um you know you have all these nice

科学家口口相传传下来的，所以你看一篇 paper 的时候，嗯，你知道，你会看到所有这些漂亮的

55:10

graphs that are actually the amalgamation of hundreds or thousands of assays and you

graphs，而这些其实都是成百上千个 assays 的汇总，然后你

55:14

may have like some representative samples shown in the um papers themselves but the training

可能会看到，嗯，paper 本身里展示了一些 representative samples，但 training

55:21

corpus that the models are actually handed is very um constricted compared to the amount of data

corpus 是模型实际拿到的，非常，嗯，受限，跟 data 的量比起来

55:26

that scientists are actually processing so fundamentally um these are things that the models lack

科学家实际在处理的，所以从根本上说，嗯，这些东西正是模型所缺乏的

55:33

a critical mass of training data on just by virtue of how scientists report the data

关于这些东西的 critical mass 的 training data，而这仅仅是因为科学家报告 data 的方式

55:38

you're seeing Nick's tacit knowledge on display you know he's spent years in the wet lab

你看到的是 Nick 的 tacit knowledge 在现场展现，你知道，他在 wet lab 里干了好多年

55:42

and even he's struggling to articulate exactly that experience that he's lived

而且连他自己也很难准确说清楚他亲历的那种经验

55:48

yeah awesome okay so then uh the language the these the frontier models and other models are

对，太棒了，好，那……呃……这些语言……这些……这些 frontier models 和其他模型都……

55:56

struggling with this task um but if I just were to put this into opus or whatever and said hey

在这个 task 上很吃力。呃，但如果我直接把这个丢进 opus 或者别的什么里，然后说，嘿，

56:04

saw this problem it would likely reach for a tool right so what why not what why is this an important

看到这个问题，它很可能就会去找个 tool，对吧？所以，为什么不呢？为什么这是一个重要的……

56:11

problem if we can just solve them with tools I mean maybe you can't this one specifically but

问题呢？如果我们直接用 tools 就能解决，我是说，也许这个具体的问题不行，但……

56:14

yeah so we can't um even when we do run these as agent tasks and we provide a bunch of you know

对，所以我们解决不了。呃，即使我们真的把这些作为 agent tasks 来跑，而且我们提供了一堆，你知道，

56:20

Python libraries for scientific purposes like sci-fi and a bunch of image manipulation libraries

用于科学目的的 Python libraries，比如 sci-fi，还有一堆图像处理 libraries

56:26

like open cv the accuracy is significant that significantly better than the vlm's but it's still

像 open cv，accuracy 显著比 VLM 的更好，但还是不到 60%，所以对，还是远没到可用的程度——如果你是做真实研究的科学家，基于这些 assays，你没法信任 agent 的结果，而且你还得根据这些结论做下一步决定。对，agent 离理解还差得远，但已经比只调用一次 VLM 好很多了，我觉得这正好指出了现在 VLM architecture 的真正短板，因为就像 Nick 指出的，人类科学家做这个任务时，这个任务里不需要用到任何 tools，他们就是看看图像，想一会儿，眼睛扫一圈，然后他们就知道了。

56:31

under 60 percent and so yeah like the it's still nowhere near usable like you would not be able to

56:37

trust the results of an agent if you were a scientist doing real research uh based on these assays

56:42

and you need to make your next decision based on the conclusions yeah the agents are still far from

然后你得根据这些结论来做下一步决定。对，agents 还远远没有

56:46

understanding but they are a lot better than just a single call to the vlm which I think points

理解，但它们比只对 vlm 做一次 single call 要好得多，我觉得这指向

56:52

to real shortcomings right now in that vlm architecture because human scientists like Nick

当前 vlm architecture 里真实存在的短板，因为像 Nick 这样的人类科学家

56:57

pointed out doing this task there's no tools involved in this task for a human to do it they just

指出，做这个任务时，人类做这个任务根本不涉及 tools，他们只是

57:02

look at the image they think about it for a bit they their eyes scan around and then they know the

看看图像，想一会儿，眼睛四处扫一扫，然后他们就知道

57:06

answer and I think that just points to a lot of shortcomings in the ways that we're training vision

答案是，而且我觉得这恰恰指出了我们在训练 vision

57:11

language models and provides like a clear incentive of how we should be making these models better

language models 的方式上有很多不足，也给了一个很明确的动力，去想该怎么把这些模型做得更好

57:18

in this example the model that's incorrect is mostly just suffering from visual intelligence right not

在这个例子里，出错的那个模型主要就是卡在 visual intelligence 上，对吧？不是，

57:26

yep biological understanding or understanding of the task that seems like how do you close that

对，是 biological understanding，或者对 task 的理解，那看起来就像是你怎么把那个

57:34

loop it's is it biological visual intelligence that you need to train on is it just the vision

loop 闭合起来？是说你需要在 biological visual intelligence 上训练吗？还是说只是 vision

57:39

encoders not the best examples not the best and I'm assuming a lot of the value actually comes from

encoders 不够好？examples 不够好？而且我猜很多价值其实来自

57:46

the other side where like you have a very clear example the vision is fine the model is just

另一边，比如你有一个非常明确的 example，vision 没问题，模型只是

57:51

not the best at solving the task is this it's all of the above yeah so a lot of the tasks we find

不太擅长解决这个 task。是这样吗？是以上所有。对，所以我们发现的很多 task

57:57

especially like when we run open source models like Kimmy the straight visual capabilities like

尤其是我们跑像 Kimmy 这样的 open source models 时，那种直接的 visual capabilities 就……

58:02

localization are just lacking you know even if we reframe this task as Nick kindly explained it

localization 就是很欠缺，你知道吧，哪怕我们按 Nick 好心解释的那样把这个 task 重新表述一下

58:08

in a explain like I'm five language as like find the first column where there's three blue lines

用那种“给五岁小孩解释”的说法，就像去找第一列里有三条蓝线的地方

58:15

in that world a lot of the models are still not able to do the tasks but the more advanced

在那个语境下，很多 models 还是做不了这些 tasks，但更 advanced 的

58:21

models that do have these basic visual capabilities like astra we find is a lot better at just basic

那些确实具备这些基础 visual capabilities 的 models，比如 astra，我们发现它们在基础的

58:25

vision tasks they've definitely done a lot of work there on the vision computer vision side

vision tasks 上要好得多，他们在 vision、computer vision 这边肯定做了很多工作

58:30

they're still not able to map the biology into a simple tap vision task and then do the vision

他们还是没法把 biology 映射成一个简单的 tap vision task，然后再去做 vision

58:37

task and it's really not super easy to disentangle where the errors come from necessarily like sometimes

task，而且说实话，要把这些错误到底从哪来的 disentangle 清楚，也未必那么容易，你知道，有时候……

58:42

it's that just the fact that there's a two-step process that has to both happen with no errors

就是因为有一个两步流程，必须两步都无差错地发生

58:48

that allow errors to compound somewhere in there do you have some intuition as to

从而让错误在其中某个环节累积；你有没有一些直觉，关于

58:53

where those gains come from is it from astra is a very good agent harness it is trained to test things

这些提升从哪来，是因为 astra 吗？astra 是一个非常棒的 agent harness，它被训练来测试东西

58:59

try different possibilities to solve it like rigorously check its answer and you know if you throw

尝试不同的可能性来解决它，比如严格检查它的答案，而且你知道如果你把

59:06

this at just a really good coding agent in my intuition it may solve it on this type of task

这个交给一个真的很强的 coding agent，按我的直觉，它可能会在这种 task 上解决它

59:12

or is it actually like they've trained in a lot more bio and it's it's picked up something

还是说其实是因为他们在 bio 上训练得多得多，然后它学到了一些东西

59:17

like this and the nuances that it should find yeah I don't know any of the details of what's

比如这个，以及它应该找到的那些细微差别。是啊，我不知道任何细节，关于到底

59:22

going on in the model but I feel like they've done a lot of computer vision research and there's

model 里在发生什么，但我觉得他们做了很多 computer vision 研究，而且还有

59:27

been a lot of breakthroughs in the actual architecture for computer vision specifically

具体到 computer vision 的实际架构，已经有很多突破了

59:31

cool do you have a sense for how good the models need to be before

挺酷的。你觉得这些模型得好到什么程度，才——

59:37

scientists will will sort of or or let's put it firm it differently it would be useful in a lab

科学家会……呃，或者说，我们换个说法，它在实验室里会有用

59:44

in a like human out of the loop lab in the loop kind of situation yeah it would have to be

在那种类似 human out of the loop、lab in the loop 的场景里，对，那它必须得

59:49

highly highly accurate um it's hard to quantify the exact um it's hard to quantify that but

非常非常准确，呃，很难量化确切的……呃，很难量化这个，但

59:55

somewhere under 90 percent would be more or less insurable I mean there's a real economic

差不多低于 90% 的话，或多或少还算可承保吧。我是说，这里确实有实实在在的经济……

60:02

and um also in some cases public health impact of getting this kind of thing wrong

而且呃，在某些情况下，这类东西一旦出错，还会有公共卫生方面的影响

60:07

so in this example if you were to answer five instead of one and somebody were to run with that

所以在这个例子里，如果你回答的是 5 而不是 1，然后有人真的拿这个去执行

60:12

they would be performing their next experiment experiment at like a 5x concentration

他们下一次实验会以大约 5x concentration 来进行

60:17

of what is actually required um and basically going down a experimental path that is just

相对于实际所需，嗯，基本上就是走上一条实验路径，而这条路径简直

60:23

fraught with payroll by virtue of the initial investigation just being completely wrong

风险重重，因为最初的调查本身完全错了

60:29

um although the experiment was performed correctly all right so but in you can imagine you have

嗯，虽然实验执行得是对的，好吧，但你可以想象，你有

60:34

these linchpin experiments right like that like some things matter than more than others and so you

这些关键实验，对吧，就像有些事情就是比其他事情更重要，所以你

60:40

really want to be basically 99.99 percent on certain tasks and other just you could there's like

真的会希望在某些任务上基本上做到 99.99%，而其他任务就，你可以，差不多有

60:46

more room for erity of a sense of how how well models are calibrated to their confidence on these

更多空间留给 erity，一种感觉，就是 models 在这些事情上对自己的 confidence 到底 calibrated 得有多好

60:54

things that's a great question um I guess the the easy answer would be they speak with a degree of

这些事情，这是个好问题，嗯，我想最简单的答案会是，他们说话时带着某种程度的

61:04

confidence um that is not warranted based on the data that we're seeing right yeah so um if we

confidence，嗯，基于我们看到的数据，这并不合理，对吧？是啊，所以嗯如果我们

61:11

look at opus's answer in this case it is saying like very definitively the answer is five it's

看 opus 在这个例子里的回答，它非常明确地说答案是 5，它

61:16

saying there's no discernible 20 kda product it doesn't give any wiggle room there it's a very very

说没有可辨识的 20 kda product，它在这一点上没留任何回旋余地，这是一个非常非常

61:21

confident narrator um which if you're a human trying to um you know take the feedback coming from

自信的叙述者，嗯，如果你是个人类，想嗯，你知道，接收来自

61:28

the model um is particularly problematic because uh you don't even know that hey I may need to hedge

model 的反馈，嗯，这尤其成问题，因为呃你甚至不知道，嘿，我可能需要留点余地

61:35

against the potential error in this calculation are in this analysis um but which in a way compounds the

来应对这个计算里的潜在错误，或者这个分析里的潜在错误，嗯，但这某种程度上加剧了

61:42

issue yeah and we actually run an experiment in our paper where we have the model not only answer the

问题，是啊，而且我们在论文里其实做了一个实验，让 model 不仅回答

61:47

question but also do what's called verbalize confidence you know just self report its confidence

问题，还做所谓的 verbalize confidence，你知道，就是 self report 它的 confidence

61:51

you know nothing mathematical here totally vibes based based on the models intuition but yeah we

你知道，这里完全没什么数学成分，全靠 vibes based，基于 model 的直觉，但 yeah，我们

61:56

find they're notoriously uncalibrated and they often say they're max confidence uh basically

会发现它们出了名地 uncalibrated，而且经常说自己 max confidence，呃，基本上

62:02

near 100 and the answer is wrong yeah it's clearly uh two things I want to know if you actually go

接近 100，然后答案是错的，yeah，很明显，呃，我想知道两件事，如果你真的去

62:07

back to the website um you're the yeah this if you scroll up it's the first time we're seeing

回到网站，um，你那个，yeah，这个，如果你往上翻，这是我们第一次看到

62:14

Gemini be quite good um if you scroll up actually even just to the rankings there yeah so you know

Gemini 相当不错，um，如果你往上翻，其实哪怕只是看排名那里，yeah，所以你知道

62:21

it's not too often that you see Gemini 3837 better than five six opus mues I don't think there's

不常能看到 Gemini 3837 比 five six opus mues 更好，我不觉得有

62:30

much you can comment on why a lab is doing good but is there something you've noticed with Gemini

太多可以评论的，关于为什么一个 lab 做得好，但有没有什么你注意到的，关于 Gemini

62:36

and frontier science in general I would say I've noticed that it's pretty good on image

以及总体上的 frontier science，我会说，我注意到它在 image 上挺不错的

62:43

like vqa style tasks in general and I think maybe that speaks to Gemini launching from the very

像 vqa 风格的任务总体上，我觉得这可能也说明了 Gemini 从一开始就是作为 multimodal model 推出的

62:49

beginning as a multimodal model yeah so maybe uh falling up on the question about lab and the loop

嗯，所以也许回到关于 lab 和 loop 的那个问题

62:57

we have there's a lot of attention on recursive self improvement uh right now and I'm wondering

我们现在有很多关注都在 recursive self improvement 上，嗯，我在想

63:04

I have kind of two I can see two worlds one in which you can kind of in it just do general

我大概有两种……我能看到两个世界：一种是你可以在里面直接做通用的

63:10

RSI and you solve that plus solve the vision problems and you kind of solve the problem or you can

RSI，然后你解决了它，再加上解决了 vision 问题，你就差不多把问题解决了；或者你可以

63:16

imagine that the models really need need to be able to think uh natively in biological data in order

想象模型真的需要能够原生于 biological data 来思考，才能

63:25

to solve some of problems you have a sense for which one is going to be the answer or any thoughts

解决其中一些问题。你觉得哪一种会是答案，或者有什么想法

63:32

about that yeah so I think when we're talking about producing a good machine learning model there's

关于这个？嗯，所以我觉得当我们在谈论产出一个好的 machine learning model 时，有

63:37

two dimensions one is given a data set you know what are the training algorithms neural architectures

有两个维度，一个是给定一个 data set，你知道有哪些 training algorithms、neural architectures

63:43

and all that to maximize the model's performance on that data set you know that's called sample

以及所有这些，来最大化 model 在那个 data set 上的 performance，你知道，这在 machine learning 里叫 sample

63:48

complexity and machine learning that's the holy grail of machine learning research to get low sample

complexity，而这是 machine learning research 的圣杯，就是要做到低的 sample

63:52

complexity when people talk about recursive self improvement they're talking about uh AI inventing

complexity；当人们谈 recursive self improvement 时，他们说的是，呃，AI 在发明

63:58

new ideas typically to bring down that sample complexity better architectures um but then the other

新想法，通常是为了把那个 sample complexity 降下来，更好的 architectures，呃，但另一个

64:03

access is to make a better model is just to have better data more data what type of data what kind

方向是把 model 做得更好，就是要有更好的 data、更多的 data，什么类型的 data、什么

64:09

of reward signals and that data and obviously you need both of those uh doing well in order to actually

样的 reward signals，以及那些 data，而且显然这两边你都得做好，呃，才能真正

64:14

have a good model at the end of the day and to define what is better data and good data at some level

最终得到一个好的 model，并且在某种程度上定义什么是更好的 data、好的 data

64:21

it really needs a connection right with the external worlds you can't just uh be in a loop um of

它确实需要跟外部世界建立连接，对吧，你不能只是呃待在一个循环里，嗯，思考自己的 data。咱们假设，你知道，model 觉得自己能完成这些 VQA tasks，所以它只是在生成 data，呃，然后给自己正在生成的 data 提供答案，而这些答案是它自己造的。你知道，就像我们之前讨论过的，这些 models 在这种 data 上的 confidence level 其实并没有很好地 calibrated，嗯，你很容易就能看到这变得非常失控。嗯，特别是像我们在一些 training data 实验里确实试过创建 synthetic data，嗯，我们看到 synthetic data 往往呃要么是逼真的，要么是简单的，呃简单的，但几乎从来不是既逼真又难的，因为那些逼真又难的 data 有 edge cases。

64:27

thinking your own data let's let's assume you know the model thinks it's able to do these VQA tasks

思路了。你自己的 data——咱们、咱们假设你知道，model 觉得自己能做这些 VQA 任务

64:32

and so it's just generating data uh and giving itself answers for the data that is generating that

所以它就只是在生成 data，呃，然后给自己正在生成的 data 给出答案，那些

64:39

are self-created and you know as we previously discussed the models are really not calibrated in

答案是自我创造的；而且你知道，就像我们之前讨论过的，这些 models 真的在……方面没有 calibrated

64:44

their confidence level on this data um you can easily just see that going very off the rails um

他们在这份数据上的信心水平，嗯，你很容易就能看出来那会彻底跑偏，嗯。

64:49

especially like we did try in some of our training data experiments to create synthetic data um

尤其是，比如我们在一些 training data 实验里确实试过创建 synthetic data，嗯。

64:54

we see the the synthetic data is often uh either realistic or easy uh easy but it's almost never

我们看到，synthetic data 往往呃，要么是 realistic，要么是 easy，呃 easy，但它几乎从来不是

65:02

realistic in hard and that's because the hard data that's realistic has edge cases

既 realistic 又 hard，而这是因为 realistic 的 hard data 会有 edge cases

65:06

and when we prompt these models to generate synthetic data with edge cases it ends up like

而且当我们用 prompt 让这些 models 生成带 edge cases 的 synthetic data 时，结果会变成

65:10

violating physical impossibilities creating these molecules that are super unrealistic and so we

违反物理上的不可能性，造出这些超级不现实的 molecules，于是我们

65:16

think that you know you really do have to be grounded in some uh human workflow human intuition

就觉得，你知道，你真的必须扎根在某些嗯 human workflow、human intuition 里

65:22

like when we bring a new scientist uh into the lab or even society advancing uh this type of

就像我们把一个新 scientist 嗯带进 lab，或者甚至整个社会在推进嗯这类

65:28

technology you know we're not doing it from first principles we're doing it from an understanding

technology 时，你知道，我们不是从 first principles 出发，而是基于一种理解

65:33

of the physical intuition behind how these molecules work which comes from these assays and other

也就是这些 molecules 如何运作背后的 physical intuition，而这来自这些 assays 和其他

65:40

educational material that is straight up just grounded in the workflows and biological research

教育材料，而这些就是直接扎根在 workflows 和 biological research

65:45

that's been done and so I think it's really important that uh this type of RSI work that is done

已经做过的东西里，所以我觉得嗯，这类正在开展的 RSI work 真的很重要

65:52

here will have to be really integrated into the biological research workflows that exist today

这必须真正整合进今天已有的生物研究 workflows 里

65:58

and just trying to do it from scratch you know create the machine god in a temple in silico

而如果只想着从零开始做这件事，你知道，在 in silico 的神殿里造一个机器神

66:03

and then have it be able to do this kind of wet lab work in the real world it will just be lacking

然后还想让它能在现实世界里做这种 wet lab 工作，那它只会缺

66:08

too much uh intuition on what is actually interesting and what matters and the mental map

太多，呃，对什么才真正有意思、什么才重要，以及 mental map 的直觉

66:15

essentially of how this how these molecules actually work in a real organism right so let's walk

本质上就是这些 molecules 在真实 organism 里到底怎么运作的，对吧。那我们来走

66:21

through I don't have a clear picture in my mind if so I have a you know I have my new Neil lab

一遍——我脑子里没有清晰的画面。假如说，我有，你知道，我有一个新的 Neil lab

66:28

and I'm doing biology research and I want to create some amazing bio AI and so I come to you

我在做生物研究，想做出一些很厉害的 bio AI，所以我来找你

66:35

and I what do so you have these benchmarks but presumably there's what more data backing

然后我问，那我该做什么？所以你有这些 benchmarks，但想必还得有，什么呢，更多 data 在背后支撑

66:41

that you guys are generating or procuring what what if what's that process look like what how do

你们是在生成还是在采购？那、那如果……那个流程到底是什么样的？

66:48

we interact so that I end up with the data I need yeah there are basically two routes that can be

我们要怎么对接，才能让我最后拿到我需要的数据？

66:53

taken um so there is the wet lab approach so we're able to procure creates um work with partners to

嗯，基本上有两条路可以走。一条是 wet lab 的方式，所以我们可以 procure creates，嗯，跟合作伙伴一起做 wet lab 的工作，嗯，你知道，就是沿着这些实验室在找的方向。

67:00

get wet lab work um you know in the vein of what these labs are looking for uh it's usually very

呃，这通常非常定制化，所以得按个案来做，但一般来说，如果你想要，我们可以 build。

67:06

custom so it would have to be done on like a per case basis but typically if you want it we can build

这就是 wet lab 方面的情况。第二种类型是那种……嗯，pre-bait，要么是 image artifacts、professional artifacts，要么就是基本的 text prompts。

67:12

that's uh the story answer for the wet lab aspect the second flavor that these come in are these

这些是由我们网络内的专家以非常大的规模制作的，用来在某种……上 train 这个 model。

67:18

kind of um pre bait either image artifacts professional artifacts are just like basic text prompts

有点像，嗯，pre bait，要么是 image artifacts、professional artifacts，就像基本的 text prompts

67:26

that are made by experts within our network at a very large scale to train the model on some kind of

这些是由我们 network 内的专家以非常大的规模制作的，用来在某种东西上 train 这个 model

67:33

specific workflow or capability that these frontier companies are looking to actually develop that

这些 frontier companies 真正想开发的具体 workflow 或 capability，是能嗯更准确地接入 drug discovery 流程中的某一个环节，或者你知道，如果你只是想在 biotech 或 medical imaging 里比较通用地做事，嗯，像 vials 这种，会相对更 generalized 一点，对应到你想构建的 capability

67:39

could be um to plug in more accurately to one portion of the drug discovery process or you know if

那为什么公司不……如果你是在用 CRO 之类的东西来生成数据，为什么公司不自己生成数据呢？为什么我不干脆直接这么做呢？

67:47

you're just trying to work generally within biotech or within medical imaging um you know something

是啊，嗯，barrier to entry 很高，因为说实话，这是最大的……这是最大的问题之一

67:53

like vials where it's a bit more generalized to the capability that you're trying to build

所以嗯，在这个行业里，这些公司实际需要的数据量是高度 scaled 的

67:58

and why aren't companies if you're why aren't companies generating if you're using a CRO or something

那为什么公司不——如果你是在用 CRO 之类的东西，为什么公司不生成呢

68:05

to generate data why not just why wouldn't i just do that yeah well it's uh there's a lot of

来生成 data，为什么不直接——我为什么不直接那么做呢？是啊，嗯，这个嘛，呃，有很多

68:10

barrier century because to be honest that's one of the biggest uh that's one of the biggest things

barrier century，因为说实话，那是最大的呃，那是最大的事情之一

68:15

so um in this industry the volume of data that these companies actually need are highly scaled

所以嗯，在这个行业里，这些公司实际需要的数据量是高度 scale 的

68:21

um they just need a lot of it you know we don't work on the scale of hundreds of units who work in

嗯，他们就是需要很多，你知道，我们不是按几百个单位的规模来做，通常做的是

68:27

thousands if not tens of thousands typically so um there's a lot of trust you have to give to a CRO

几千甚至上万个，所以嗯，你得对 CRO 给予很多信任

68:33

with that type of data and i think also um more um a more complex issue to solve is the expertise

对于那类数据，而且我觉得嗯，还有一个更嗯、更复杂的问题要解决，就是专业能力

68:41

so uh these workflows are very dynamic and the needs of these labs update on a weekly if not

所以呃，这些 workflows 非常动态，这些实验室的需求现在每周甚至

68:48

daily basis on those nowadays and um the way that handshake is built is from the ground up it's an

每天都会更新，而且嗯，handshake 是从底层构建起来的，它是一个

68:54

expert network we can reach out to people within these various fields of professions and source them

专家网络，我们可以联系到这些不同专业领域里的人，并把他们获取过来

68:59

very quickly um and also in a very vetted and robust way to make sure things create just like

非常快，嗯，而且还要用经过严格筛选、很稳健的方式，确保能产出那种

69:05

very strong data on these specific workflows um that is fairly bulletproof compared to most other

在这些具体 workflows 上非常强的数据，嗯，和大多数其他相比这相当可靠

69:12

like CROs are smaller vendors awesome well thanks for showing us it's always refreshing for i think

就像 CROs 是更小的供应商，太棒了，嗯，谢谢你给我们展示，我觉得这总是让人耳目一新

69:18

the majority of people will never actually see these science uh data sets walk through as examples

大多数人其实永远都不会真正看到这些科学呃 data sets 被当作例子一步步讲解

69:24

i encourage people to try it themselves like if you scroll down here to your list of tasks uh

我鼓励大家自己试一试，比如你在这里往下滚到你的任务列表，呃

69:29

just click on any of them i actually tried it with the second version so we did that one in the left

就点其中任意一个，我其实用第二个版本试过，所以左边那个我们做过了

69:34

if you scroll down a little bit more there's another one uh a little bit more i i tried it i just

如果你再往下滚一点，还有另一个，呃，再往下一点，我、我试过，我刚刚

69:38

third in the chat GPT it it's purely hallucinates that one on there the one that you're on right now

在 ChatGPT 里试了第三个，它、它纯粹是在那儿胡编，就是你现在正开着的那个

69:44

this one oh oh yeah yeah the same the same task on the right oh sorry um this this so i it looked like

这个啊，哦哦对对，右边那个同样的、同样的任务，哦抱歉，呃，这个这个，所以我，它看起来像

69:52

the example you were talking about yeah i took the screenshot i took the question it did as you

你刚才说的那个例子，对，我截了图，我拿了那个问题，它做得跟你一样

69:56

said it used some uh code interpreters like started bounding boxing and it spewed out eight with

它说它用了一些，呃，code interpreters，比如开始做 bounding boxing，然后它就蹦出了 eight，带着

70:03

pure confidence and uh you know there there's a lot of ways to go but it's it's refreshing to

十足的自信，而且，呃，你知道，有，有很多路可走，但真正让人耳目一新的是去

70:09

actually see what these tasks are dive deep um appreciate your expertise uh excitingly i think we

看到这些任务到底是什么、深入进去，嗯，感谢你的专业分享，呃，令人兴奋的是，我觉得我们

70:15

have more benchmarks that you've done so we're gonna bring in more of the handshake team are

还有更多你做过的 benchmarks，所以我们会请来更多 handshake team 的人，还有

70:20

any closing closing question thoughts oh it's uh i always ask this what why should why should

有什么收尾的、收尾的问题或想法吗？哦，这个，呃，我总会问这个：为什么应该，为什么应该

70:26

scientists uh or why should AI people be interested in in these kinds of problems

科学家们，呃，或者为什么 AI 领域的人应该对这类问题感兴趣

70:32

carrying cancer is the absolute best application of AI and we should all be focused on carrying

carrying cancer 是 AI 绝对最好的应用，我们都应该专注于 carrying

70:36

cancer yeah again to the social contract like there's a lot of societal risk in AI i think everybody's

cancer，对，又回到社会契约，就像 AI 里有很多社会风险，我觉得每个人都

70:43

aligned in that regard especially with recent developments um but the ultimate reward and the

在这方面是 aligned 的，尤其是考虑到最近的一些进展，嗯，但最终的回报以及

70:48

social contract that we've kind of signed with the companies developing these models is that there

我们某种程度上和开发这些 models 的公司签下的社会契约是，

70:52

will be true innovation in either the um you know drug or human health space um you know for like

会有真正的创新，在嗯你知道药物或人类健康领域，嗯你知道为了像

70:59

better infrastructure and if that isn't very specifically met um i i think it would be a hugely

更好的基础设施，而如果这一点没有被非常具体地满足，嗯，我我我觉得那会是一个极其

71:07

problematic for everyone involved yeah i know and i would also argue that AI for science in general

对每个参与者来说都成问题的局面，是的我知道，而且我还会认为 AI for science 总体上

71:15

is really at right now at the frontier of a lot of the machine learning and AI research right

现在真的处在很多 machine learning 和 AI research 的前沿，对吧

71:23

solving these kinds of problems we saw actually although you can solve lots of image problems using

解决这类问题，我们其实看到，尽管你可以用

71:31

the current generation of models when you get to these more complicated and domain specific

当前这一代 models 解决很多 image problems，但当你碰到这些更复杂、更 domain specific 的

71:36

questions you actually might demand different architectures different data regimes and

你实际可能会提出的问题，可能会要求不同的 architectures、不同的 data regimes，而且

71:41

nothing's like that yeah and that's the life sciences especially are one of the domains that

没有什么像那样，对，而 life sciences 尤其属于那种领域，它

71:46

might be so complex that it's just beyond human understanding and you know we really do need AI

可能复杂到完全超出人类理解，你知道，我们真的需要 AI

71:52

actually to be able to make these problems go away awesome i would add um there's a lot that you can

真正能让这些问题消失，太棒了。我想补充一下，嗯，有很多你可以

71:59

add to so periodic put out their model neon recently uh it's state-of-the-art better than all

补充的，所以 Periodic 最近发布了他们的 model Neon，呃，它是 state-of-the-art，比所有

72:05

models and cheaper on material sciences tasks uh friend shout out to Rohan who works there

models 都更好，而且在 material sciences tasks 上更便宜，呃，给在那工作的朋友 Rohan 点个赞

72:11

is a researcher that doesn't have a background in any of this right he's done model training was

他是一位研究者，完全没有这方面的背景，对吧，他做过 model training，之前是

72:16

that open AI um material sciences by all this is new but within you know the company's only been around

在 OpenAI 吗，嗯，material sciences，所有这些都很新，但你知道，这家公司才成立没多久

72:23

for so long they're able to put out state-of-the-art models so if you work in AI language models

他们长期以来都能推出 state-of-the-art 模型，所以如果你做 AI language models 相关工作

72:29

there's a lot of impact you can have even if you don't have domain expertise okay we're back with

即使你没有领域专业知识，也能产生很大影响。好，我们回来了。今天

72:35

Colin and Sahil and Slicks has joined us again about to talk about finance finance maxing

Colin 和 Sahil 也在，Slicks 也再次加入我们，接下来要聊 finance、finance maxing

72:40

Colin you're new to the show give yourself a quick introduction yeah so join handshake AI

Colin，你是这个节目的新面孔，简单介绍一下自己吧。对，我加入了 Handshake AI

72:44

in February and before that was in finance for 13 years at a major bulge racquet bank doing

是在二月，之前在一家大型 bulge bracket bank 做了 13 年金融，做

72:51

energy and then a lot of real estate private equity awesome in the intro we talked a bit about

能源，然后做了很多 real estate private equity。很棒，在开场里我们聊了一点

72:56

hill climbing finance um but we haven't seen the actual tasks examples reasoning traces or models

hill climbing finance，嗯，但我们还没看到实际的 tasks、examples、reasoning traces 或 models

73:02

fail uh you have atlas walk us through it give us the high level yeah so atlas is uh

fail 的情况，呃，你有 Atlas，带我们过一遍吧，先讲个大概。对，所以 Atlas 是，呃

73:08

hundred tasks on a new finance benchmark in 13 different environments so we really want to create

一个新的 finance benchmark 上有 100 个任务，分布在 13 个不同的 environment 里，所以我们真的很想打造一个尽可能贴近真实世界的 environment，这样当 agent 进入其中一个 environment 时，它就是在扮演银行里某个人的角色。

73:14

a an environment that's seen away the real world as much as possible so when an agent enters one of

它有七到十二个团队成员，有半打客户要打交道；拿到指令后，它还得通过 email、calendar、drive，以及分析师会用到的每一个工具，去做一轮类似事实调查的事情，从而把上下文理清。

73:19

these environments it is taking on the persona of someone on a bank it has you know seven to twelve

这其实就像银行家真正的工作。

73:26

team members it has half a dozen customers that has to interact with and it has to contextualize

很多人好奇银行家为什么赚那么多钱。

73:32

itself after it gets instructions by going on a kind of a fact finding mission across email calendar

他们赚这么多钱并不是因为做超复杂的数学——那其实相对直接——他们赚钱是因为……

73:39

drive every every tool that an analyst would use and this is you know this is like a banker's real job

推动每一个，每一个分析师会用的工具，而这就是，你知道，这就像是银行家真正的工作

73:47

a lot of people wonder why bankers make so much money they don't make a lot of money because they're

很多人想知道为什么银行家赚那么多钱，他们赚那么多钱不是因为他们

73:51

doing super complex math it's actually like relatively straightforward but they make money because

在做超级复杂的数学，其实这相对来说挺直白的，但他们赚钱是因为

73:56

they have to do it under pressure 24 seven on call and basically quarterback uh the world economy

他们得在压力下干活，7×24 小时随叫随到，基本上就是给世界经济当四分卫

74:03

so you know when when Pepsi calls or Coca-Cola calls or Google calls or your company wants to IPO

所以你知道，当 Pepsi 打来电话，或者 Coca-Cola 打来电话，或者 Google 打来电话，或者你的公司想 IPO

74:09

you call a banker that banker gets a junior banker he's a you know he or she's 24 or 25

你就找银行家，那个银行家再找个初级银行家，他——你知道——他或她才 24 或 25 岁

74:15

and they are running all the analysis that makes the market confident in in buying whatever you're

而他们在做所有分析，让市场有信心去买你正在

74:21

doing to invest in the economy for tomorrow so that was that was the inspiration for atlas finance

做的、为明天经济投资的事，所以那就是 atlas finance 的灵感来源

74:26

and we can kind of talk through the environment uh you know basically we staffed them like a banker

然后我们可以大概聊聊这个环境，呃，你知道，基本上我们是像给银行家配团队那样给他们配人

74:31

would be staffed um they have access to all these different tools uh they again they're taking on

那样配人，嗯，他们能用所有这些不同的工具，呃，他们又是在扮演

74:36

a persona and then they have to deliver their their products like a like a human wood and then we

一个角色，然后他们得交付他们的产品，就像、就像真人会做的那样，然后我们

74:41

grade it with super rigorous rubrics that are expert built and we could talk about this more later

用专家构建的超严格 rubrics 来给它打分，这个我们之后可以再聊

74:47

the difficulty of rubrics in a semi verifiable domain but that's kind of the inspiration for atlas

rubrics 在 semi-verifiable domain 里的难度，但这差不多就是 atlas 的灵感来源

74:53

i guess the last thing i would ask about atlas is how do models currently perform so in bio you know

我猜关于 atlas 最后想问的是，模型目前表现如何，比如在 bio 里，你知道

74:57

stuff was sub 30 percent yeah it's similar it's it's sub 20 percent okay it's very poor um you know

那东西低于30%，对，差不多，是低于20%，好吧，非常差，嗯你知道

75:05

frankly when we talk to part of the challenge of this is when we talk to experts in the field

说实话，我们聊的时候，部分挑战在于，当我们和这个领域的专家交流时

75:11

they're not comfortable using this stuff you know it's virtually unusable there's a there's a phrase

他们不太愿意用这些东西，你知道，这几乎没法用，有句话是

75:16

as a banker you're taught uh very early that if it's 95 percent right it's 100 percent wrong

作为银行家，你很早就被教导，如果它 95% 是对的，那它就是 100% 错的

75:22

because it takes just as much time if not more time to fix a flawed output as it is to just

因为修复一个有缺陷的输出，花的时间一样多，甚至更多，就跟直接...

75:27

build from scratch or use a template like templates exist people use them all the time sometimes

从零搭建，或者用一个 template，就像 template 本来就存在，大家一直在用，有时候会用。

75:32

we pretend like they don't but it's sometimes it's not that hard to wire up a template uh and the

我们装作好像大家不用，但其实有时候把 template 接起来也没那么难，呃，而且

75:38

act of wiring up the template contextualizes the human so they feel comfortable with what they're

把 template 接起来的这个过程，会让人有上下文感，让他们对自己正在看的东西感到舒服，所以

75:42

looking at so you know AI applications or tools for finance have to be so good that they can

你知道，金融领域的 AI applications 或工具必须做到足够好，才能

75:50

surpass that that inherent advantage of what currently exists uh and frankly you know they just

超越现有方案那种固有优势，呃，而且说实话，你知道，他们只是

75:56

say that you know it's not here yet it's great for note-taking uh or you know remembering you know

说，你知道，这个还没真正到位，它拿来做 note-taking 很棒，呃，或者你知道，用来记住，你知道

76:01

six six meetings of interactions with your client but it's it's not there in terms we don't have the

六次、六次跟客户互动的会议，但它还没到那个程度，从某种意义上说，我们没有

76:06

the reliability and the consistency of the output it's still a huge improvement to where it was

output 的可靠性和一致性，不过跟以前相比，这仍然是巨大的进步。

76:11

a year ago like you know a year ago you could look at an output and you would know immediately

一年前，你知道，一年前你只要看一个 output，马上就能知道。

76:16

oh this is wrong this is broken like i would never trust this model the challenge now is they

哦，这是错的，这是坏的，我绝对不会信任这个 model。

76:21

look and they sound and they read very well very professionally so a lot of times the formatting's

现在的挑战是，它们看起来、听起来、读起来都非常好，非常专业。

76:27

right uh the structure can be right but you know in in some of our examples in atlas you know these

所以很多时候 formatting 是对的，呃，structure 可能也是对的。

76:33

are these are tasks that again take like 15 to 30 hours and so they have over four thousand cells

但你知道，在我们 Atlas 的一些例子里，你知道，这些任务又是那种要花 15 到 30 个小时的。

76:40

with a value or a label or a formula in them and like i think the one that i'm going to go through

所以里面有超过四千个 cells，每个里面都有一个 value、一个 label 或一个 formula。

76:44

has over 2600 live formulas which is just like a logic chain of you know a decision that that

而且我觉得我要过一遍的那个有超过 2600 个 live formulas。

76:51

produces a value and you know the key errors are in less than 50 of those you know 2,600 cells

这就像一条逻辑链，你知道，一个决策会产生一个 value。

76:59

so would you rather go and try to find the 50 cells or would you rather just you know build it

所以你是更愿意去找那50个cells，还是更愿意就，你知道，直接把它做出来？

77:04

build it from scratch right it's it's very challenging and frankly we've seen before you know

从零开始做，对吧？这真的，这真的非常有挑战性，而且说实话，我们之前也见过，你知道

77:10

even pre-LOM there were there were companies that were approaching the industry and they're saying

甚至在pre-LOM之前，也有，也有公司在进入这个行业，他们会说

77:14

hey you can use machine learning to do you know standardized underwriting or build templates and I

嘿，你可以用machine learning来做，你知道，标准化的underwriting，或者搭templates，然后我

77:19

actually you know mess around with one of these in like 2018 2019 and he said hey we just bought

其实我，你知道，在2018、2019年左右试过其中一个，然后他说，嘿，我们刚买了

77:24

an apartment uh two months ago you know seventy thousand dollars we didn't tell them the price because

一套公寓，呃，两个月前，你知道，七万美元，我们没告诉他们价格，因为

77:29

like you know seventy thousand dollars why don't you tell us what you think it's worth and they

就像，你知道，七万美元，你为什么不告诉我们你觉得它值多少钱，然后他们

77:34

came back and they gave us a range that was actually pretty good it was like plus or minus five

回来给了我们一个区间，实际上相当不错，大概是正负5

77:38

percent of what we bought but we realized that you know in these live processes these markets are so

我们买的东西里占百分之多少，但我们意识到，你知道，在这些实时流程里，这些市场太

77:44

efficient that if you're off by plus or minus five percent you're either never going to do a deal

高效了，你要是差个正负百分之五，你要么永远做不成一笔交易

77:49

and you're never going to generate fees or allocate capital or or make loans right for your

要么永远赚不到费用，或者没法 allocate capital，或者没法给你的

77:53

anti-customer or you're going to overpay on everything and so with leverage you know I would have

anti-customer 做合适的贷款，要么你什么都会多付钱，所以有了 leverage，你知道，我本来会

77:58

been out of the money twenty percent day one if I've been using this automated skill set so it's

第一天就 out of the money 20%，如果我一直在用这套 automated skill set，所以这

78:03

you know even though in some metrics we're like we're super close in terms of usability uh we

你知道，虽然在某些 metrics 上我们就像，在 usability 方面我们超级接近，呃，我们

78:09

know we have a long way to go in terms of deployment okay I think the one before you jump in actually

知道在 deployment 上我们还有很长的路要走，好吧，我觉得在你插话之前，其实

78:14

the one thing I would say we're obviously proud of the benchmark but like it's it's it's it's in

我想说的一点是，我们当然为这个 benchmark 感到骄傲，但就像，它，它，它，它是在

78:19

there's things a good step in the direction of simulating concepts or things that a banker has to do

有些东西确实是朝着模拟概念、或者模拟一个银行家必须做的事情迈出的不错一步

78:24

but there's a long way to go before we're like fully simulating like a full bank and all the variables

但距离我们真正完整地模拟一整家银行、以及所有会起作用的变量，还有很长的路要走

78:32

that play right and so I mean this is still just kind of the tip the iceberg yeah let's look at it

而且这些变量都得对，所以我意思是，这还只是冰山一角。对，我们来看看吧

78:37

uh you know as as Colin pulls it up I was going to say uh you know the tendency for me so we had a

呃，你知道，等 Colin 把它调出来的时候，我本来想说，呃，你知道，我的倾向是……所以我们之前有一个

78:43

similar ingredient issue at uh doing frontier code you just do an easy medium hard and that easy

类似的 ingredient 问题，呃，在，呃，做 frontier code 的时候，你就是搞 easy、medium、hard，然后那个 easy

78:48

you know let the models have like 70 80 percent uh but and then medium in hard uh would be the

你知道，让 models 大概有 70、80% 吧，呃，但然后 medium 和 hard 呃，就会是那个

78:54

sort of gradation there um yeah let's let have look yeah so this slide is just kind of high level

差不多就是那种难度分级吧。嗯，对，我们来看看。对，所以这张幻灯片只是大概、概括地

79:00

like what a task really is and so you know with the the agent receives the prompt uh it gets its

说明一个 task 到底是什么，所以你知道，agent 接收到 prompt，呃，它就会拿到它的

79:07

instructions and it's not we're not hand holding the model we're not giving it every specific

给 instructions 的时候，并不是说——我们不是在手把手带着 model，我们不会把每一个具体的

79:13

assumption uh and resource that it needs but we may give it an attachment or point to an email

假设，呃，以及它需要的资源都给它，但我们可能会给它一个 attachment，或者指向一封 email

79:18

that has the attachment and then you know number two it kind of kicks off the discovery chain

那封 email 里有 attachment，然后你知道，第二点，它会启动这个 discovery chain

79:23

which is exactly how a real human being has to get you know up to speed and one of these jobs

这恰恰就是一个真人要在一份这样的工作里，你知道，快速上手时得经历的过程

79:28

it's like you you're getting thrown into different scenarios hey like go figure it out right like

就像你——你会被扔进不同的场景里，嘿，就像，自己去搞清楚，对吧，就像

79:32

that's like you know one of the most common phrases on a bank fork figure it out right so the agent

这就像，你知道，在 bank fork 上最常听到的话之一：figure it out，对吧，所以 agent

79:37

has to do basically what a human has to do which is find the assignment and then it's it it it disperses

基本上就得做一个人必须做的事，也就是找到任务，然后它——它它它就会分散开

79:43

through the environment to find all the different sources so again it's it's calendar it's one note

到整个环境里去找到所有不同的来源，所以再说一遍，它是——它是 calendar，它是 OneNote

79:51

it's slack it's drive it's attachments in the email um and then it actually goes through

是 Slack，是 Drive，是邮件里的附件，嗯，然后它实际上会走一遍

79:58

its its computation so it'll it'll you know it would build the the model first in python and then

它的，它的 computation，所以它会，它会，你知道，它会先用 Python 把 model 建出来，然后

80:04

it would it would use that as kind of its source of truth and it would rebuild and excel so as

它会，它会把这个当作它的 source of truth，然后它会重建一个 Excel，以便

80:09

usable for human and then you know deliver it we either in a data room or an outlook uh just like

让人能用，然后，你知道，把它交付出去，我们要么放在 data room 里，要么通过 Outlook，呃，就像

80:16

a human would be like hey you know mr. md mrs. md please find attached let me know if you have any

一个人会说：嘿，你知道，MD 先生、MD 女士，请查收附件，如果有任何

80:21

questions you know sincerely right so it's very realistic and then and then we would grade it

问题，你知道，此致，对吧，所以它非常真实，然后，然后我们会给它评分

80:26

on a on a rubric that we built with humans to again try to mimic best mimic the um

按照一个，一个我们和真人一起搭建的 rubric，再次尝试去模仿，尽量最好地模仿，嗯

80:34

the grading or the judgment of like a VP because in the real world you know there's there's

像 VP 那样的评分或判断，因为在现实世界里，你知道，有，有

80:39

several layers of review and we we also had to do the same layers of review where you know it gets

好几层审核，而且我们我们也得走同样的几层审核，你知道，它会

80:44

passed from the analyst and associate to the VP who needs to sign off on it before it goes to

从 analyst 和 associate 传到 VP，VP 需要签字确认，然后才会到

80:49

the md before it goes to the investment committee so this is kind of the environment where the agent

MD，再到 investment committee，所以这就是那种环境，agent

80:57

agent works in one example uh this is row and stern partners and again we created uh you know

agent 就是在这样的环境里工作。举个例子，呃，这是 row and stern partners，然后我们又一次创建了，呃，你知道，

81:02

13 different personas and you know they they could either be a a group in a larger bank or

13 个不同的 persona，你知道，他们他们可能是大银行里的一个团队，或者

81:09

you know a boutique advisory and in this case the agent has assumed the persona of Julian uh and

你知道，一个 boutique advisory，而在这个案例里，agent 采用了 Julian 的 persona，呃，并且

81:16

is being directed to its uh information from sanjay who's the the staffing managing director

正从 sanjay 那里接收它的，呃，信息，sanjay 是那个，那个 staffing managing director

81:23

and then there's also going to be additional information uh particularly around the the assumptions

然后还会有额外的信息，呃，尤其是围绕那些假设

81:27

for the model from from Nathan who's the partner at the client and then just it's a very it's a very

对这个模型来说，来自、来自 Nathan，他是客户那边的合伙人，然后就是，这是一个非常、非常

81:33

you know realistic environment in that there's other projects going on you know there's thousands

你知道，很真实的环境，因为还有别的项目在进行，你知道，有成千上万

81:39

of emails but there's other projects with other clients and other staffings and the agent has to

封邮件，但还有别的项目、别的客户、别的人员配置，而且 agent 必须

81:44

know you know where to focus its attention right and so i can know i'm only i'm only focused on project

知道，你知道，该把注意力放在哪儿，对吧，所以我能知道，我只、我只专注于 project

81:49

falcon if it if it's from dan or greys ignore it it's irrelevant right here it my here it might be

falcon；如果它来自 Dan 或 Greys，就忽略它，它不相关，对吧，这里它可能

81:55

sending me something but that's not relevant for this task so it's you know just just like a human

给我发东西，但那和这个任务不相关，所以就是，你知道，就像一个人一样

82:00

one and so this this kind of talks through shows the the instruction and the email so the the key

所以这个、这个有点像走一遍流程，展示了指令和邮件，所以关键

82:07

instruction says to you know to look through your inbox uh you know for project falcon and this

指令说，你知道，去翻你的收件箱，呃，你知道，找 project falcon，然后这个

82:13

this starts the discovery phase and so you'll see in the email and this will come up later

这会启动 discovery phase，所以你在 email 里会看到，这个后面也会再提到

82:19

that we're very clearly giving it context i mean frankly it might be it might be too much

就是我们非常明确地在给它 context，我是说，坦白讲，这可能，可能有点太多了

82:23

context like it's word in some ways we're handholding it but we're trying to make it fair um

context，就像在某些方面它是 word 一样，我们是在手把手带着它，但我们想让它公平，嗯

82:28

because it's realistic to have too much context yeah yeah that's true some some of some of it could be

因为 context 太多其实是挺真实的，对对，确实，有些，有些，其中有些可能会是一种干扰，嗯

82:32

a distraction um but you know project falcon is uh you know 2019 sponsor lbo with a three three

但你知道 project falcon 是，呃，你知道，2019 年的 sponsor LBO，带一个 338(h)(10) election

82:40

eight h ten election so you know and lbo is is is is pretty standard for a bank i would say the

所以你知道，LBO 对银行来说算是挺标准的，我会说，tax treatment 有点独特

82:47

the tax treatment is is a little unique but we're not hiding the ball here right it's like it is in

但我们不是在藏着掖着，对吧，就像它就在 instructions 里，就在 kickoff email 里，就在 one note 里

82:53

the it is in the instructions it is in the kickoff email it is in the one note it is in the the

就在那个，那个……

82:59

deal terms from Nathan who's the client and we also see from the trajectory that the agent

来自客户 Nathan 的 deal terms，而且我们也能从 trajectory 看到，这个 agent

83:05

in this case one of the two frontier models opened all these sources and looked at it and understood

在这个案例里，两个 frontier models 中的一个打开了所有这些 sources，看了它们，并且理解了

83:09

what it was supposed to do and it you know also gets it's it's uh it's submission instructions

它应该做什么；而且你知道，它也拿到了，它的，呃，它的 submission instructions

83:16

i like your are these internal tools you guys used uh yeah suit face to everything yeah we built

我喜欢你的——这些是你们用的 internal tools 吗？呃，是的，suit face to everything，是的，我们建了

83:21

we built uh our engineering we have great engineers on this that built you know all these

我们建了，呃，我们的工程团队——我们这方面有很棒的工程师，他们建了，你知道，所有这些

83:25

internal tools right so try to you know recreate the environment as much as possible um and then

internal tools，对吧？所以尽量，你知道，尽可能还原那个环境，呃，然后

83:31

again this is the um this is the inbox and this is the the email from Nathan who's sending joy in

再说，这是那个，呃，这是 inbox，而这是 Nathan 发来的那封 email，他在把 joy 发进来

83:38

like the key the key insights for the election so again like an lbo for those who are unfamiliar at

像是这次 election 的关键，关键 insights，所以再说一次，就像 LBO 一样，给那些不熟悉的人

83:43

home it's like it's basically just a leverage buyout winery where you're using debt to buy a company

嗯，基本上就是一个 leverage buyout，就是你在用债务买下一家公司

83:48

and you're hoping that it grows over time and you saw it for more money and you know you can you

然后你希望它随着时间增长，再以更高的价格把它卖掉，你知道你可以，你

83:52

can use debt to juice the returns of your equity right like very basic just like a house yeah just

可以用债务来放大你 equity 的回报，对吧，就像很基础的，就像房子一样，对，就

83:57

exactly just like a house just like a house uh the only difference the only wrinkle for this scenario

完全正确，就像房子一样，就像房子一样，呃，唯一的区别，这个场景唯一的小问题

84:03

is that a three three eight eight ten election it's it basically treats the transaction as an

就是 three three eight eight ten election，它基本上把这笔交易当作一个

84:11

asset sale for tax purposes and there's certain scenarios where it makes more sense basically like

asset sale 来做税务处理，而且在某些场景下它更有意义，基本上就像

84:16

you save money on taxes and you might pay the the sour uh more money as compensation because

你在税上省了钱，而且你可能付给 the sour 呃更多钱作为 compensation，因为

84:22

ultimately that becomes their tax liability but there's there's very specific accounting that

最终那会成为他们的 tax liability，但是有，有非常具体的 accounting，这个

84:27

goes into this treatment and a piece of that accounting is ignoring the existing goodwill on the

会进入这个处理流程，而这部分会计处理里有一点是忽略了卖方

84:33

seller's balance sheet right and so we'll come back to that but you know this is basically just a

balance sheet 上现有的 goodwill，对吧，所以这个我们回头再说，但你知道这基本上只是

84:41

high level view of what the agent did it took you know 73 actions 101 tool calls took him 36

一个 agent 做了什么的概览，它执行了你知道 73 个 actions、101 次 tool calls，花了 36

84:47

minutes which is a huge savings over like a human particularly like the context retrieval and stuff

分钟，这相比人类来说节省太大了，尤其是像 context retrieval 这些东西

84:52

I think that's like a no-brainer um you know you could save hours just going through the inbox

我觉得这根本不用想，嗯你知道光是翻收件箱

84:58

and like the notes and and and pulling that together uh because you know this this actual

和笔记、把这些整理到一起，就能省好几个小时，呃因为你知道这个实际的

85:03

full model rebuild took our experts you know it over 15 hours and we how long will you estimate the

full model rebuild 花了我们专家超过 15 个小时，然后我们……你估计这些

85:09

tasks takes I would I would think this would probably take 15 hours I mean between you know going

任务要花多久？我会……我会觉得这大概要花 15 个小时，我的意思是，在你知道要去……

85:14

through and having to again check what Nathan said and what Sanjay said and you know the old version

通读一遍，还得再去核对 Nathan 说了什么、Sanjay 说了什么，以及你知道的那个旧版本

85:20

I mean even that just contextualizing yourself can take a couple hours what's harder to quantify is

我是说，光是让自己把上下文捋清楚就可能要花好几个小时，而更难量化的是

85:25

like what's the value of time in not waiting for information right if some of this information is

比如说，不用等待信息的时间价值到底是什么，对吧，如果其中有些信息是

85:32

coming in piecemeal live and you're just like sitting on your hands if you can if you can feed

零散地实时进来的，而你只能干等着，如果你能，如果你能把这些喂给

85:36

that to an agent and like let it let it know when it's ready for your review like that's that's a

一个 agent，然后比如让它，让它准备好给你 review 的时候通知你，那那真的是一个

85:43

great way to allocate time more efficiently right also just to jump it is this the answer correct

更高效地分配时间的好办法，对吧；另外插一句，这个答案是对的吗

85:49

is it a successful it is not correct uh uh the big issue is that when it was building its model

这算成功吗，它并不正确，呃，呃，大问题是，当它在构建自己的 model 时

86:01

in Python it didn't realize that it was making a decision on the existing goodwill that we talked

用 Python 的时候，它没意识到自己是在基于我们谈过的现有 goodwill 做决定

86:06

about and you know if if it had if it had flagged it for review or it had um you know said I'm not

大概就是，你知道，如果，如果它当时把它标记出来做 review，或者它，嗯，你知道，说我不

86:13

sure about this let me think about it like if you if you go and ask this model today how is

太确定这个，让我想想，就像如果你，如果你今天去问这个 model，那

86:19

existing goodwill on the sellers balance sheet treated from the perspective a very a very

卖家 balance sheet 上现有的 goodwill，从一非常，一个非常

86:23

purchaser for a three three eight eight ten tax treatment it knows the answer it's like well in

买方，按 three three eight eight ten 这种 tax treatment 来看，它知道答案，就像，嗯，在

86:30

that scenario you would not actually use it in the basis and it's a relic this calculation it doesn't

那种情况下，你实际上不会把它用在 basis 里，而且这是个老古董，这个 calculation，它不

86:34

apply to knowledge excuse me it doesn't apply to knowledge exactly exactly it doesn't apply the

适用于 knowledge，抱歉，它不适用于 knowledge，对，对，它不适用于

86:39

knowledge during the key step so you know in in action 20 it's supposed to you know basically

knowledge 在这个关键步骤，所以你知道，在，在 action 20 里，它应该，你知道，基本上

86:46

build this this calculation you know this this this whole model uh and it just it this one line here

构建这个，这个 calculation，你知道，这个，这个，这个整个 model，呃，而它只是，它这一行这里

86:53

where it subtracts the excess goodwill from the excess basis for the step up calculation

就是它从 excess basis 里减去 excess goodwill，用来做 step up calculation

86:58

basically poisons the entire valuation the entire model and it doesn't it doesn't have there's

基本上就把整个 valuation、整个 model 都毒化了，而且它没有，它没有——

87:04

nothing in the trajectory that suggests it was it wasn't sure about the decision it was making

trajectory 里没有任何东西表明它对自己正在做的决定不确定

87:09

or you know going through its reasoning and frankly those are the worst kind of errors because if

或者说，你知道，走一遍它的 reasoning，坦白说，那些是最糟糕的一类错误，因为如果

87:14

if it knows that there's something to surface to you like at least that's helpful but here it's

如果它知道有东西需要呈现给你，至少那还有帮助，但在这里它是

87:18

just hidden in that it didn't realize it was making a key decision but again this is

只是藏起来了，因为它没意识到自己正在做一个关键决定，但再说一次，这

87:23

what's what's ironic or interesting about it is when a human being does this they would absolutely

它反讽或者说有意思的地方在于，当人类这么做的时候，他们绝对会

87:28

this would be if this was my you know job this would be my full focus I probably would have talked to

这会是——如果这是我的，你知道，工作，这会是我的全部焦点，我大概会去跟……聊

87:33

an accountant and a lawyer you know two times each before I sent this to my VP it's like hey did I

在我把这个发给我 VP 之前，我找了个会计和一个律师，你知道，各问了两次，就像：嘿，我这个弄对了吗

87:38

get this right like I've never done one of these before like can you help me and they're like yeah

我以前从没做过这种，你能帮我一下吗；他们说：嗯，不是，我们之前就是这么做的，对吧

87:42

no this is how we did it before right and this information was in the environment it was it was

而且这些信息就在 environment 里，就在，就在邮件里，在 OneNote 里，你知道，那是个非常真实的呃 application

87:46

in the email it was in the one note you know it was a very realistic uh application

但它就是没意识到自己在干什么，然后它基本上，呃，你知道，就把它发出去了

87:54

but it just it didn't realize what it was doing and it basically um you know it shipped it

它，抱歉，它检查了自己的工作，它检查了工作，它把 Python 和 Excel 对了一遍

88:00

it sorry it checked its work it checked its work on the it checked the python against the excel

而且它对上了，它就像在说：太好了，我巴不得出任何错，完美，像这样，这个可以

88:06

and it tied it out it was like great I'm dying to make any errors perfect like this this is good

发了，对吧，然后它，它把它发给了 Sanjay，你知道，它把错误发出去了，而且它没标出来

88:11

to go right and then it it mailed it to Sanjay and it you know it shipped the error and it didn't flag

要走对，然后它，它把它邮件发给了 Sanjay，然后它，你知道，它把错误发出去了，而且它没有 flag

88:17

anything you know uh as a as a major concern and this is the this is the this is the actual

你知道，呃，作为一个主要关注点，这就是那个，这就是那个，这就是实际的

88:25

deliverable that created and this is why it's so challenging today to audit this stuff um and and

交付物，而这就是为什么今天 audit 这些东西、呃，并且

88:33

use it because you know on its surface it actually looks you know like a fairly you know robust model

使用它这么有挑战性，因为你知道，表面上看，它确实看起来，你知道，像一个相当，你知道，稳健的 model

88:39

right and again there's there's over 2600 formulas in here and over 4,000 cells but it's it's just

对吧，而且再说一遍，这里面有超过 2600 个 formulas 和超过 4,000 个 cells，但就只是

88:48

this line it's just this row 16 that like blows up everything and what it does is it the effective

这一行，就只是这个 row 16，它把一切都搞炸了，而它实际造成的

88:55

impact is that it it decreases the the error it it varies like the true answer about like 10%

影响是，它会降低误差，它会让真实答案大概变动 10% 左右

89:02

in terms of cash flow and it messes up the the cash you know every every period thereafter so

在 cash flow 方面，而且它会把之后的每一个 period 的 cash，你知道，都搞乱，所以

89:08

this could do two things right this could completely embarrass you in front of your client

这可能会造成两件事，对吧，这可能会让你在客户面前彻底难堪

89:14

and it could lose you the mandate because the worst thing that you know can happen to one of these

而且这可能会让你丢掉这个 mandate，因为你知道，对这类 advisory 公司来说，最糟的情况可能就是——

89:19

advisory companies you know whether it's a major consultant or a major bank on the M&A side

不管是大咨询公司，还是 M&A 那边的大银行——

89:25

what are they selling they're selling trust and competency and confidence right like the phrase

他们卖的是什么？卖的是信任、专业能力和信心，对吧，就像那句话说的——

89:30

like nobody got fired for hiring McKinsey or Goldman Sachs right it's like because they just have

没人会因为雇了 McKinsey 或 Goldman Sachs 被开除，对吧，就是因为他们有那种 brand equity——

89:34

the brand equity where when they walk into the room and they tell you something you you trust them

当他们走进会议室，跟你说点什么的时候，你会相信他们是对的。

89:39

that they're right so the worst thing that can happen is you go in there and the CFO and the CEO

所以最糟的情况就是，你进去汇报，CFO 和 CEO 看着你的 output，然后觉得你的 net basis step 有点不对劲——

89:44

is looking at your output and like something's weird about your net basis step hold on don't go to

“等等，先别翻下一页，我们能不能聊聊这个？”

89:50

the next page like can we can we talk about this and then you know you get a lump in your throat and

然后你知道，你嗓子眼一紧，喉咙像哽住了，而且……

89:53

you're like oh man like you know what I do and they're like well you know your competitor yesterday

你就会说，哎，你知道我是做什么的，然后他们说，嗯，你知道你竞争对手昨天

89:59

came in like that their numbers their numbers different like why why is it different right and then

就这么来了，他们的数字，他们的数字不一样，就会想，为什么，为什么会不一样，对吧，然后

90:03

one of you's wrong and you know people have lost you know multi-million dollar deals for like way

你们俩有一个是错的，而且你知道，有人已经因为远低于那个的差异

90:09

less than that yeah show or you're an investor and you you know you overpay or underpay because

丢掉了你知道的几百万美元的交易，是啊，或者你是投资人，而且你知道，你会多付钱或少付钱，因为

90:17

you know again like 5% off you know the true number doesn't sound like a lot but it compounds over time

你知道，再比如偏离真实数字 5%，你知道，听起来不算多，但它会随着时间复利累积

90:24

and particularly in finance where everything is data driven and number backed becomes a huge issue

尤其是在金融领域，一切都由数据驱动、由数字支撑，这就成了一个大问题

90:30

so again it's it's functionally unusable and frankly like I wouldn't you know it

所以再说一次，它实际上根本没法用，坦白说，我不会，你知道，它

90:35

it'd be very difficult to go through this thing without some kind of attention helping you

如果没有某种 attention 帮你，要把这东西过一遍会非常困难

90:40

understand where you should be looking you know every like the formatting is clean like the

你得明白该往哪儿看，你知道的，就比如格式都很干净，像是那个

90:44

structure makes sense but there's these little gremlins in the model that like when they pop up

结构也说得通，但模型里总有这些小捣蛋鬼，它们一冒出来

90:49

you know it becomes a usability and a trust issue I think it's always cool to see these trajectories

你知道的，这就成了可用性和信任问题。我觉得看这些 trajectories 总是很酷

90:55

an actual example and environment walk through it how do you translate this to the RL side so

一个实际例子和环境，把它走一遍，那怎么把这个转化到 RL 这边，所以

91:02

one is the rubric of what you pull two is there's a lot of work here that's done for 30 minutes

一个是你拉取什么的 rubric，二是这里有很多工作是为了 30 分钟的

91:07

of this correct in this case there's one error sometimes there's multiple how do you transition

这个正确，在这个例子里有一个错误，有时候有多个，那你怎么把那个转

91:12

that out into useful reward signal because you did a lot of inference and a lot of it's correct

化成有用的 reward signal，因为你做了很多 inference，而且很多都是对的

91:18

maybe start off with the rubric of how this is graded and answer as you yeah so creating the rubric

也许先从这该怎么评分的 rubric 开始，然后你来回答，对，所以就是创建这个 rubric

91:24

is by far the most challenging part of you know the semi verifiable domains generally it's very

这绝对是 semi verifiable domains 里最有挑战性的部分，你知道，通常很容易也很直接地说，比如这个输出是 0%，你知道，他们什么都没返回，像一张空白页或者假的，好，明白了；或者这个输出是 100%，就是，哇，他们做得完美，你知道，我一点都不会改，这就是完美。

91:31

easy or straightforward to say like this output is a 0% you know they return absolutely nothing

真正难的是中间每一步，尤其是当有多个变量在起作用的时候，对吧？也许 model A 更擅长，你知道，那个 cash flow built down，但 model B 更擅长 balance sheet，那你怎么、你怎么、你怎么平衡这两件事，对吧？

91:38

like a blank page or fake okay understood or this output is 100% like oh they nailed it perfectly

所以我们花了很多时间和专家一起构建这些 rubrics，目标就是让它们尽可能 deterministic，对吧，所以我们、我们……

91:44

and I you know I wouldn't change anything this is perfect what's really hard is every step in

而我，你知道，我不会改任何东西，这很完美，真正难的是每一步，在

91:49

between particularly when there's there's multiple variables that play right maybe maybe model A is

中间，尤其是当有，有多个变量在起作用的时候，对吧，也许，也许 model A 是

91:54

better at you know the the cash flow built down but model B is better at the balance sheet well how

更擅长，你知道，那个 cash flow built down，但 model B 更擅长 balance sheet，那怎么

92:01

do you how do you how do you balance those two things right and so we spent a lot of time building

你怎、你怎么、你怎么平衡这两件事，对吧，所以我们花了很多时间构建

92:06

these rubrics with experts and the goal is to make them as deterministic as possible right so we we

这些 rubrics，和专家一起，目标是让它们尽可能 deterministic，对吧，所以我们、我们

92:14

intentionally create the environment the inputs to be you know there's only one right answer right

有意地把环境设计成，输入进去的时候，你知道，只有一个正确答案，对吧

92:22

and you know there may be conventions in the economy or the you know the industry where there

而且你知道，经济里或者你知道，行业里可能会有一些惯例，在这些地方

92:27

could be two defensible or reasonable you know approaches but we just we you have to make absolutely

可能有两种都站得住脚、或者合理的，你知道，做法，但我们只是，我们，你必须绝对

92:35

sure in the environment that okay there are two reasonable approaches but actually this note from

确保在环境里，好吧，有两种合理的做法，但实际上，来自

92:39

Sahil says you know use this approach right so then we can we can grade it on the rubric so that's

Sahil 的这条注释说，你知道，用这种做法，对吧，这样我们就能，我们就能按 rubric 来评分，所以这就是

92:45

how we that's how we approach to the rubrics generally the rubrics are about 50% just hard checks

我们，这就是我们处理 rubric 的方式，一般来说，rubric 里大约 50% 就是 hard checks

92:53

so this is like a simplification but you know half of the rubric is the numbers like is you know

所以这算是一种简化，但你知道，一半的 rubric 就是那些数字，比如，你知道

92:59

is is the year one EBITDA right is the year two cash balance right and we think that's reflective

第一年的 EBITDA 对不对，第二年的现金余额对不对，而且我们认为这能反映出来

93:04

and that is reflective of the industry standards like you can argue it's even higher so I add a high

这也反映了行业标准，你甚至可以说它更高，所以我加一个 high-level 的层面：它就像解决一个任务的关键步骤，会有某种 binary answer，你应该有 step one 到 five，你应该有这个，就是这种 binary 的是或否，对对对。所以对，我们会给实际的 numerical outputs 打分，你知道，一半的分数是 criteria，而且总体而言，在我们的 benchmark 里，我想平均有 146 个不同的 criteria，然后剩下的分数会分配到 construction 上，所以不一定是那些数字，而是真的深入到 process 里。所以这就是我们一直在说的 gate，而且这也是，你知道，我会把它也算作一种 construction output，因为你知道，这个 step up。

93:09

level it's like the key steps of solving a task there's some binary answer you should have

分层来看，就像解决一个任务的关键步骤，有些 binary 的答案，你应该有

93:15

step one to five you should have this that's this binary us or no yeah so yeah so we we'd grade

第一步到第五步，你应该有这个，这就是 binary 的，是或否，嗯，所以对，所以我们会去评分

93:20

the actual numerical outputs for you know half of the the points are criteria and and in general

实际的 numerical outputs，你知道，一半的分是 criteria，而且、而且一般来说

93:27

and in our benchmark there's I think on on average 146 different criteria and then the remainder of

在我们的 benchmark 里，我觉得平均有 146 个不同的 criteria，然后剩下的

93:35

the the points are allocated across construction so not necessarily you know the numbers but

那些、那些分会分配到 construction 上，所以不一定是，你知道，那些数字，而是

93:42

but really drilling down into the process and so this this is the gate that we've been talking about

但真正深入到这个流程里，所以这、这就是我们一直在说的那个 gate

93:49

and this is you know I would also consider a construction output because you know the step up

然后这个，你知道，我也会把它看作一种 construction output，因为你知道这个 step up

93:54

is built as purchase price plus estimated working capital iabilities west tax basis in current assets

它是按 purchase price 加上预估的 working capital liabilities，再加上 current assets 里的 tax basis 来构建的

94:00

west tax basis appreciable assets and recovered straight line over 15 years this is this was

还有 appreciable assets 的 tax basis，并按 15 年 straight line 回收——这就是，这当时是

94:05

discussed and shown in the environment multiple ways you know we knew that the agent looked at the

在 environment 里用多种方式讨论过、展示过；你知道，我们知道 agent 看过 source material，而这个呢，你知道，拿它来评分是公平的，结果它又搞错了，因为你知道

94:11

source material and this is you know fair to grade it on and it got it wrong again because you know

它里面的某个东西，你知道，它的 weights 不知怎么的——它看着自己正在算的东西，却忘了，你知道

94:17

something in you know it's it's weights did to looked at what it was calculating and forgot you

知道这是一个特殊例外，然后，呃，可能又退回到它在其他 scenario 里对 goodwill 的处理方式了；但这就是我们考虑 rubric 的方式，我们可以用这个 rubric，所以

94:23

know that this is a special exception and and kind of maybe reverted to its treatment of good will

抱歉，这个 gate 会把这个 section 里其他所有东西都关掉，就像 VP 会做的那样，对吧，如果我

94:29

for other scenarios but that's how we think about the rubric and we can use the rubric and so

针对其他场景，但这就是我们考虑 rubric 的方式，而且我们可以用这个 rubric，所以

94:34

sorry this gate shuts off everything else in this section it's just like a VP would right if I

抱歉，这个 gate 会把这一部分里其他所有东西都关掉，就像 VP 会做的那样，对吧，如果我

94:41

you know if I'm showing if I'm going through a model you know with my managing banker or whatever

你知道，如果我在展示，如果我在和 managing banker 一起过一遍 model 或者什么的

94:48

and you know row nine is wrong like they you know they check everything else but once they get the

然后你知道，第九行是错的，他们会检查其他所有东西，但一旦他们拿到

94:53

row nine that's wrong you know there's a bus in your model or there's a grand win or something

那个有问题的第九行，你知道，你的 model 里有个 bus，或者有个 grand win 之类的

94:57

you know they circle it and send it back right they don't just keep they don't keep checking

你知道，他们会把它圈出来然后退回来，对吧，他们不会就一直，他们不会继续检查

95:01

because number one they've all trusted you because they you know they see this error and it's

因为第一，他们都已经信任你了，因为他们，你知道，他们看到这个错误，而且很明显

95:07

clearly you didn't get this right so why are you going to get everything else right but also

很明显这个你都没弄对，那你凭什么会把其他所有东西都弄对呢？但另外

95:11

because these errors compound again this is you know this is over 2000 interconnected formulas

因为这些错误会叠加，再说一遍，你知道，这是超过 2000 个相互关联的 formulas

95:17

and so if there's one error that error pollutes every other number in it so and because it's realistic

所以只要有一个错误，那个错误就会污染里面的其他每一个数字，所以，而且因为它很真实

95:23

we shut off the gate for every you know all 180 points in the section doing with the tax mechanics

我们把那一节里跟 tax mechanics 有关的所有 180 个点对应的 gate 都关掉了，你知道吧

95:30

and so we can use this rubric which again has a lot of specificity on on the numbers and you can

所以我们可以用这个 rubric，它对数字这块还是有很多具体说明的，而且你可以

95:36

you can actually just logically go through well it actually got you know years one two and three

你其实可以直接顺着逻辑过一遍，嗯，它其实把你知道的第一年、第二年和第三年

95:41

correct for these numbers it got yours for wrong so now you can you can direct your attention to

在这些数字上做对了，但你的那部分做错了，所以现在你可以把注意力转到

95:46

what happened between year three and year four that there there's an error there right and so we

第三年和第四年之间发生了什么——那里有个错误，对吧——然后我们

95:51

can use we can use the rubric and we use the trajectory to come up with kind of the failure modes

可以用，我们可以用这个 rubric，然后用 trajectory 来总结出一些 failure modes

95:56

and understand which models are strong and weak in in various types of tasks the three biggest

并理解哪些 models 在各种类型的任务里强、哪些弱。三个最大的

96:04

issues that we saw across models for this benchmark were financial logic where it was unable to

我们在这个 benchmark 里跨 models 看到的问题，是 financial logic，它没法

96:11

to you know exercise this kind of expertise in context with the environment as well as scoping

就是，你知道，在跟 environment 交互的 context 里运用这类专业能力，同时还要做 scoping

96:21

completeness so it might it might have gotten through you know steps one through nine but it left

完整性，所以它可能，可能已经走完了，你知道，第一步到第九步，但它把

96:29

step ten unfinished right so it's so it's an unfinished model so yeah financial logic still

第十步留成了未完成，对吧，所以它就是个，就是个没做完的 model，所以是啊，金融逻辑还是

96:34

you know we you have to make the tasks more complicated but it's still failing these you know

你知道，我们，你得把 tasks 弄得更复杂，但它还是在这些上失败，你知道

96:40

pretty consistently across every every model family I think for on rubric specifically right like you

而且相当一致，横跨每一个，每一个 model family，我觉得尤其是针对 rubric，对吧，就像你

96:47

the shape of this task can be used for evaluation it could also be used for RL and so how you think

这个 task 的 shape 可以用于 evaluation，也可以用于 RL，所以你怎么看待

96:53

about the score probably differs right to your point if the model is getting a lot of the things

关于 score 可能就会不一样，对吧，按你说的，如果 model 把很多东西都

96:59

right I don't want to give it no you know reward at all during an RL run but for the sake of an

做对了，我不想在 RL run 期间完全不给它 reward，但为了一个

97:05

eval yeah where you're like can I actually use this thing in the real world like add a bank where

eval 嗯，就是你会想，我到底能不能在真实世界里用这个东西，比如放到一个银行场景里，让

97:11

a banker would trust it that is a different sort of threshold right on what you need and so

一个银行家会信任它，那又是另一种门槛了，对吧，就看你需要什么，所以

97:16

one thing the team did which which I was really cool is just separating you know what's truly

团队做的一件事——我当时觉得特别酷——就是把真正

97:21

critical for real world usage where if you're going to evaluate it like what would just make

对实际应用至关重要的东西拆出来：如果你要评估它，什么会让

97:27

this totally untenable like that tax treatment example like that's a critical component that would

这个变得完全没法用，比如那个税务处理的例子，那就是一个关键部分，会

97:32

just lose you trust immediately in a meeting with your client or with your VP but you're

立刻让你失去信任，在跟客户或者跟你的 VP 开会的时候，但你是

97:39

breaking it out in a way where when you're doing an RL run you can sort of still have you know

把它拆出来，这样当你做 RL run 的时候，你还是能有点，你知道

97:44

enough granularity in the reward to say hey this still scores like you know if you look at the

reward 里足够的 granularity，来说嘿这个还是能得分的，你知道，如果你看那个

97:49

thing it's 30 points right maybe this whole section's 180 points

这个……是30分对吧？可能这整一节是180分。

What Frontier Labs Buy: A Data Deep Dive with Handshake AI — Garrett Lord & Team

Play Queue