Do AI Tokenomics Matter More Than Model Benchmarks? with Chris Potts

00:00

This episode is brought to you by Blitzie, the Autonomous Software Development Platform

本期节目由 Blitzie 赞助，它是 Autonomous Software Development Platform

00:05

built for enterprise scale.

专为 enterprise scale 打造。

00:07

Today's code bases have grown beyond human comprehension.

如今的 code bases 已经大到人类无法理解了。

00:10

Millions of lines, decades of tech debt, and complexity existing tools just can't fathom.

几百万行代码、几十年的 tech debt，还有现有工具根本摸不透的 complexity。

00:16

With Blitzie, thousands of specialized agents reverse engineer the code base, mapping

有了 Blitzie，数千个 specialized agents 会 reverse engineer 这个 code base，梳理

00:21

the architecture, dependencies, and business logic.

它的 architecture、dependencies 和 business logic。

00:24

With that context, the platform then autonomously executes entire epochs, writing, validating

有了这些 context，平台就会自主执行整个 epochs，编写、验证

00:29

and testing the code for every project.

并测试每个项目的代码。

00:32

The result, Fortune 500 enterprises, are able to modernize legacy systems and ship new

结果是，Fortune 500 企业能够把 legacy systems 现代化，并交付新

00:37

features five times faster.

功能，速度快五倍。

00:40

Want to try Blitzie on your code today?

今天想在你的代码上试试 Blitzie 吗？

00:42

Unlock 1 million lines of reverse engineering and 25k lines of code generation by visiting

想解锁 100 万行 reverse engineering 和 25k 行 code generation，就访问

00:47

blitzie.com slash sandbox.

blitzie.com slash sandbox。

00:53

We've started to enter a phase of AI that's not just about making models smarter.

我们已经开始进入 AI 的一个阶段，这个阶段不只是让 models 变得更聪明。

00:58

It's also about making them economically sustainable.

这同时也是在让它们在经济上可持续。

01:02

As reasoning models consume more tokens, context windows continue to grow, and agents become

随着 reasoning models 消耗更多 tokens，context windows 持续增长，而 agents 变得

01:07

embedded in more products and workflows, the economics of these systems are becoming

随着它们嵌入更多产品和 workflow，这些系统的经济账正变得

01:11

impossible to ignore.

无法忽视。

01:13

That's given rise to a new conversation around tokenomics.

这催生了一场围绕 tokenomics 的新讨论。

01:17

How we think about the costs, incentives, and trade-offs shaping next generation of AI.

我们如何看待塑造下一代 AI 的成本、激励和 trade-offs。

01:22

One person who's been thinking deeply about this is Stanford professor and Big Spin co-founder

一直在深入思考这个问题的人，是 Stanford 教授、Big Spin 联合创始人

01:27

Chris Pots.

Chris Pots。

01:29

His recent work argues that measuring AI progress requires looking beyond model benchmarks to

他最近的研究认为，衡量 AI 进展需要超越 model benchmarks，去

01:33

ask a different question.

问一个不同的问题。

01:35

What are our tokens actually buying us?

我们的 tokens 到底给我们换来了什么？

01:39

Here's Chris explaining how he thinks about tokenomics.

接下来是 Chris 解释他如何看待 tokenomics。

01:42

Another interesting moment to be in, as we're all being made aware of the true costs of all

另一个很有意思的时刻是，我们所有人都开始意识到所有

01:47

this AI usage.

这些 AI 使用的真实成本。

01:48

The analogy here is like, I used to cost me $20 to take a ride share to the airport Uber

这里的类比就像，我以前打网约车去机场要花 20 美元，用 Uber

01:54

or Lyft, and now it costs $90, but it's more like $20 to like $500 or something, right?

或 Lyft，现在要花 90 美元，但实际更像是从 20 美元到 500 美元左右，对吧？

01:59

I think what's happening is that the big providers are testing the waters on charging us

我觉得现在的情况是，那些大型 providers 正在试水，向我们收取

02:04

the true costs, plus whatever profit they need to make as they all try to gear up for

真实成本，再加上他们需要赚取的任何利润，因为他们都在努力为……做准备

02:08

IPOs and so forth.

IPOs 等等。

02:10

And in turn, that is very quickly leading people to ask questions like, what is the return

反过来，这很快会让人们开始问这样的问题：我们买的这些 tokens 的投资回报

02:15

on investment for all these tokens that we have purchased?

到底是多少？

02:19

And it's a very tricky area to be in because what does it mean to think about value in

而且这是一个非常棘手的领域，因为在这个

02:23

this context?

语境下，思考价值到底意味着什么？

02:24

I'm Sam Charrington, and this is the Twimmel AI podcast.

我是 Sam Charrington，这里是 Twimmel AI podcast。

02:28

For over a decade, I've been exploring the ideas and innovations shaping the future

十多年来，我一直在探索那些塑造 AI 未来的想法和创新

02:32

of AI through conversations like this one that help you understand what's real, what's

通过像这样的对话，帮助你理解什么是真实的，什么是

02:37

next, and what matters.

接下来，以及真正重要的东西。

02:39

Let's jump in.

我们开始吧。

02:49

I want to say thanks for coming on.

我想说，谢谢你来做客。

02:51

I've been looking forward to this conversation.

我一直很期待这次对话。

02:54

And I think where I'd love to start us off is to really dig into your background and

我觉得，我想先带我们从这儿开始：真正深入聊聊你的背景，以及

03:02

how it kind of got you to where you are now.

这一路是怎么让你走到今天这个位置的。

03:06

Yeah, my background is in linguistics.

对，我的背景是 linguistics。

03:10

Linguistics proper, not even natural language processing.

真正的 linguistics，甚至都不是 natural language processing。

03:13

I did my PhD on, among many other things, swears, what swears are like, why we swear,

我读博的时候，研究过很多东西，其中就包括脏话——脏话是什么样的，我们为什么说脏话，

03:19

what information they encode, what kind of taboos exist around them and so forth.

它们 encode 了哪些信息，围绕它们有哪些禁忌，等等。

03:23

And that was actually the trigger that got me into NLP because I wanted a lot of data

而这其实正是让我进入 NLP 的契机，因为我想要大量

03:29

of people swearing.

关于人们说脏话的 data。

03:30

I wanted to know what the context was like, what their intentions were.

我想知道 context 是什么样的，他们的意图是什么。

03:34

So I turned to corpora.

所以我转向了 corpora。

03:36

And from there, you start using NLP toolkits to add structure to those corpora.

从那里开始，你会用 NLP toolkits 给这些 corpora 加结构。

03:41

And then after a few years, maybe you're writing your own tools for doing that work.

然后几年之后，也许你就在写自己的工具来做这项工作了。

03:45

And then when you look back after 18 years or whatever it's been, you're just an AI person

然后等你回头看，过了 18 年还是多久，你就只是个搞 AI 的人

03:51

or an NLP person.

或者搞 NLP 的人。

03:52

But that is the true story.

但这就是真实的故事。

03:53

And I feel like if I had to, I could trace the lineage of every one of my current projects

而且我感觉，如果非要的话，我可以把我现在每一个项目的脉络

03:58

back to my fascination with why we care when someone drops an f-bomb.

都追溯回我对这个问题的着迷：为什么有人爆出一句 F 词时，我们会在意。

04:05

So are you an f-bomb dropper or did you come at it from the perspective of trying to

那你是自己会说 F 词的人，还是从试图

04:10

understand these others?

理解这些其他人的角度切入的？

04:13

I think very infrequently in my life, for my linguistics class, semantics and pragmatics,

我觉得在我人生里这种情况非常少，至于我的语言学课、semantics and pragmatics，

04:20

which is about linguistic meaning, on the final day, we always do a class on swearing.

这是关于语言意义的，在最后一天，我们总会上一节讲脏话的课。

04:25

And I review the history and we kind of tie all the course themes together.

我会回顾历史，也会把课程的所有主题串起来。

04:29

My handouts for that are full of swears.

我那节课的讲义里全是脏话。

04:33

But I only swear once in the lecture.

但在讲座里我只爆一次粗口。

04:35

I present the result that people remember things better if the utterance contains a

我会展示这个结果：如果一句话里包含一个

04:40

swear because it has a kind of emotional resonance, very primitive reaction.

脏话，人们会记得更牢，因为它有一种情感共鸣，一种非常原始的反应。

04:45

And so in that moment, I pick some fact from the course, some trivial thing.

于是在那一刻，我会从课程里挑一个事实，某个很琐碎的东西。

04:49

And I restate it with a swear and then I say, all of you will remember this for eternity.

然后我用一句带脏话的话重新说一遍，接着说，你们所有人都会永远记住这个。

04:54

But other than that, I'm very shy about it in the class.

但除此之外，我在课上对这件事特别害羞。

04:58

That's funny.

这挺搞笑的。

04:59

I'm sure there is loads of research on this, but I grew up in New York City.

我确定这方面有一大堆研究，不过我在 New York City 长大。

05:07

And as a New Yorker, I think that swearing is just kind of part of my natural language

而作为一个 New Yorker，我觉得说脏话差不多就是我的 natural language 的一部分

05:12

and way of communicating.

也是我交流方式的一部分。

05:14

And I married a Midwestern girl and she doesn't tolerate it at all.

而且我娶了一个来自中西部的女孩，她完全不能容忍这个。

05:19

She doesn't do it.

她自己不说脏话。

05:20

She doesn't tolerate it.

她不能容忍这个。

05:21

She won't tolerate it for me.

她可不会容忍我这样。

05:23

And it may for, we've been married for 30 years, so I adapt quickly, apparently.

而且可能，因为——我们已经结婚30年了，所以显然我适应得很快。

05:29

But for a long time, it took a lot of restraint to change that way of communicating, particularly

但很长一段时间里，要改变那种沟通方式需要很多克制，尤其是

05:38

when I'm communicating about something that I'm excited about or emotional about or want

当我在聊让我兴奋、让我情绪激动，或者我想

05:43

to convey the importance of, it's a really interesting topic.

传达其重要性的事情时，这真的是个很有意思的话题。

05:52

Well, we're not going to turn the podcast into a podcast about swearing, but I imagine

好吧，我们不会把这个播客变成一档聊脏话的播客，但我想

05:57

there was enough research there that we could if we wanted to.

那方面的研究足够多，如果我们想的话，完全可以做。

06:00

It's a fascinating area because it gets right to the heart of the culture that we've

这是个引人入胜的领域，因为它直击我们文化的核心，而我们已经

06:04

constructed and how it relates to our usage and everything else about us.

它是怎么被建构出来的，以及它如何关联到我们的使用和我们身上的一切。

06:08

Yeah, it's fascinating that we have swears.

是啊，我们竟然有脏话这件事挺有意思的。

06:11

When the old swears lose their power, we invent new ones.

当旧的脏话失去力量时，我们就会发明新的脏话。

06:13

We pretend like nobody should use them, but as you say, people use them all the time.

我们假装好像谁都不该用它们，但就像你说的，人们一直都在用。

06:17

And it feels like an important part of being a language user that we've got them

而且感觉，作为语言使用者，我们能拥有它们

06:21

available to us.

并让它们随时为我们所用，是很重要的一部分。

06:23

Yes.

是的。

06:24

And the string of questions.

还有那一连串的问题。

06:25

I'd love to hear your take on kind of a linguist in the age of modern AI, you know,

我很想听听你的看法，你知道，关于在现代 AI 时代做一个语言学家，

06:34

transformers, statistical models, you know, this is an NLP used to be kind of coming from

transformers、statistical models，你知道，NLP 过去有点像是从

06:40

a linguistic perspective.

语言学视角出发的。

06:42

And now the entire field is shifted to a statistical perspective.

而现在整个领域已经转向了 statistical perspective。

06:46

And I'd love to hear your reflections on, you know, being on the other side of that

而且我很想听听你的思考，你知道，关于身处那个

06:50

transition, as well as maybe more importantly, ways that you think that kind of the traditional

过渡的另一边，以及也许更重要的是，你认为传统的

06:56

foundational linguistics is still important to the way we think about AI today.

foundational linguistics 在哪些方面仍然对今天思考 AI 的方式很重要。

07:02

These questions are on my mind all the time, yeah, because I operate at the intersection

这些问题一直在我脑海里，是的，因为我就在交叉点上

07:07

of all these different fields.

在所有这些不同领域里。

07:08

And I will say it's useful to distinguish in this context linguistics, you know, people

而且我想说，在这个语境下，把语言学区分出来是很有用的，你知道，有些人

07:13

in my department at Stanford study language and social identity, historical linguistics,

在 Stanford 我所在的系里研究语言和社会身份、历史语言学，

07:19

the structure of language.

语言的结构。

07:21

And they're just doing scientific investigation of language as a human phenomenon.

他们只是在把语言作为一种人类现象来做科学研究。

07:25

And they are not technologists and they're not trying to inform technology.

他们不是技术专家，也不是想为技术提供参考。

07:29

So their project is interestingly impacted by technological developments for NLP people

所以，对做 NLP 的人来说，他们的项目受到技术发展的影响很有意思，

07:35

who are, of course, participating directly in the engineering project.

当然，做 NLP 的人是直接参与这个工程项目的。

07:39

They're affected in a very different way by the rise of GNI and the kind of homogeneous

他们受到 GNI 崛起以及那种

07:43

nature of the solutions that people now adopt in that space.

同质化解决方案特性的影响，方式非常不同，而人们现在在这个领域正是采用这类方案。

07:47

So for the linguists, I feel like this is the most exciting moment that anyone could

所以对语言学家来说，我觉得这是任何人都能

07:52

have dreamed of.

梦想得到的最激动人心的时刻。

07:53

I feel incredibly privileged to be alive in this moment where humans encounter for the

我觉得无比幸运，能活在这个时刻：人类第一次

07:59

very first time non-human creatures that use our language very fluently.

遇到能非常流利地使用我们语言的非人类生物。

08:05

I think it's weirding us all out, but from the point of view of understanding the human

我觉得这让我们所有人都觉得怪怪的，但从理解人类

08:09

capacity for language, what a gift because you can ask about the mechanisms which are

语言能力的角度看，这是多好的礼物啊，因为你可以去问那些机制，它们是

08:15

different from humans, but obviously sufficient for achieving a certain kind of behavioral

与人类不同，但显然足以实现某种 behavioral

08:21

performance.

performance。

08:23

We can think about them as investigative tools.

我们可以把它们看作研究工具。

08:25

I mean, we train them on the internet.

我的意思是，我们在互联网上训练它们。

08:26

They're basically incredibly powerful distributional learners, and we can learn a lot from them

它们基本上就是极其强大的 distributional learners，而且我们可以从它们身上学到很多

08:31

about the true structure of language by just looking at the kinds of things that they learn.

关于语言真正结构的东西，只要看看它们学到的那些东西就行。

08:36

And it really gets at the heart of core questions in linguistics about how much of language

而这真的触及了语言学核心问题的核心：语言学习有多少

08:41

learning is innate and the nature of our capacity and whether it's statistical or symbolic.

是 innate 的，我们这种能力的本质是什么，以及它到底是 statistical 还是 symbolic。

08:46

All those things come flooding in in a completely fresh way.

所有这些都以一种完全崭新的方式涌进来。

08:50

And so whatever your reaction to language models is, it should be a significant one, right?

所以不管你对 language models 的反应是什么，它都应该是很重大的，对吧？

08:56

This should be causing you to rethink key questions.

这应该会促使你重新思考一些关键问题。

09:00

And that's all you could hope for as a scientist that you have new angles, new perspectives,

而作为一名科学家，你能期望的也就这些了：你有了新的角度、新的视角，

09:04

new questions reopened.

新的问题被重新打开。

09:06

That's been incredible.

这真是太不可思议了。

09:09

For NLP, I think it's a more uncertain prospect because pre the arrival of like pre-trained

对 NLP 来说，我觉得它的前景更不确定，因为在像 pre-trained

09:18

models, which for me would be like the LMO model back in 2017, 2018, before that there

models 出现之前——对我来说，那就像是 2017、2018 年的 LMO model——在那之前，有

09:25

was still a lot of statistical work, of course, and we were in the deep learning era.

当然，还是有很多 statistical work，而且我们当时处在 deep learning 时代。

09:29

But you could still, for example, do a PhD that was entirely about some specific phenomenon

但你仍然可以，比如，读一个 PhD，完全围绕某个具体现象

09:34

and maybe some very specific tweak to a model.

或者也许是对模型做某个非常具体的 tweak。

09:37

So you could say, I'm going to work on summarization and I've got a new idea about how to do that

所以你可以说，我要做 summarization，我有个新想法，关于怎么

09:41

well using deep learning models and that could be your PhD.

用 deep learning 模型把它做好，那就可以是你的 PhD。

09:45

And what we started to see 2018, 2019, 2020, especially with the arrival of GPT-3, that

而我们开始在 2018、2019、2020 年看到，尤其是随着 GPT-3 的到来，那

09:51

that was a very uncertain prospect because you might wake up one morning to find that

那是一个非常不确定的前景，因为你可能某天早上一醒来就发现

09:55

you had been completely scooped, that with essentially no effort, one of these large

你已经被完全 scoop 了，就是几乎不费吹灰之力，这些大型的

09:59

pre-training runs had done better than you at the thing that you'd worked so hard on.

pre-training 实验在你花了那么多心血的那件事上，反而比你做得更好。

10:04

And that caused an interesting, probably overall productive, but interesting and challenging

而这引发了一场很有意思、整体上大概也算有成效、但既有趣又充满挑战的

10:08

crisis for people, especially students who are trying to figure out what to do next with

危机，对很多人来说尤其如此，特别是那些正在琢磨接下来该拿

10:13

their PhD research.

自己的 PhD research 怎么办的学生。

10:15

But I think all of us felt a kind of real uncertainty in that moment.

但我觉得，在那一刻，我们所有人都感受到了一种真切的不确定感。

10:18

Yeah, I remember the anxiety of that time and I always felt it was kind of expressed

是啊，我记得那时的焦虑，而且我一直觉得它某种程度上表现为

10:26

as, you know, is research and NLP fundamentally like scale-limited already need a certain degree

就是，你知道，研究和 NLP 是不是从根本上就已经是 scale-limited 了，是不是已经需要一定程度的

10:33

of scale that only a handful of organizations have to do foundational research and is everyone

scale，只有少数几个组织才拥有，才能做基础研究，那是不是每个人

10:41

else going to be relegated to like poking the beast and seeing what it does.

否则就只会沦为去戳一戳那头野兽，看看它会有什么反应。

10:48

And I'm curious, do you feel like that was an anxiety that's passed or is it still very

我很好奇，你觉得那是一种已经过去的焦虑，还是它现在依然非常强烈？

10:54

present?

事情真的就那样发展了吗？

10:55

Has it panned out quite like that?

那你是怎么——你知道的——这件事对你来说最后是怎么解决的？

10:58

How do you, you know, how's it been resolved for you?

太有意思了。

11:00

So fascinating.

并没有解决。

11:01

Not resolved.

这是我经常和我的合作者以及学生们讨论的事情。

11:02

It's something I discuss a lot with my collaborators and with my students.

这是我经常跟我的合作者还有我的学生讨论的一件事。

11:06

We're all trying to figure this out in this moment.

这一刻，我们都还在努力把这件事弄明白。

11:09

I will say one concrete thing we did was orient a lot of our research toward interpretability.

我会说，我们做过的一件具体的事，就是把很多研究都转向了 interpretability。

11:16

Just the project of understanding how these models end up being so good at such hard tasks.

就是去理解这些 models 到底是怎么变得这么擅长这么难的任务的。

11:21

And the reason we did that is it's relatively inexpensive and it's also an area where clearly

我们这么做的原因是，它相对不那么贵，而且这也是一个领域，很明显

11:26

you would be explicitly hoping that models would get better because then there would be more

你会明确地希望 models 变得更好，因为那样就会有更多

11:30

to explain.

可以解释的东西。

11:33

Versus if you were doing that summarization project, you might quietly be hoping that there

而如果你在做那个 summarization 项目，你可能会暗自希望

11:37

wasn't going to be so much progress so that you could make the progress.

不会出现那么大的进展，这样你才能取得进展。

11:41

Let's hope the next model isn't good at summarization.

希望下一个 model 别太擅长 summarization。

11:43

I want to be the star of that show.

我想成为那个节目的明星。

11:45

That's, as I said, very uncertain.

就像我说的，这事儿非常不确定。

11:47

But if you're doing that kind of trip, you're like, let's get the new model released because

但如果你在做那种旅程，你会想，咱们赶紧把新的 model 发布出来吧，因为

11:49

now we're going to have even more structure to find, even more to explain.

现在我们要找的结构会更多，要解释的也会更多。

11:53

And that felt like a very productive choice.

而且感觉那是个非常有成效的选择。

11:56

I don't want to leave out the fact that it's also cheaper to do this research and that

我不想漏掉一个事实，那就是做这项研究也更便宜，而且

11:59

is significant.

这很重要。

12:00

And then I would say that right now, a lot of us are in a moment of thinking, we should

然后我会说，现在，我们很多人都在想，我们应该

12:05

do stuff that is weird and creative and out of the mainstream.

做一些古怪、有创意、脱离主流的事情。

12:09

We should be thinking about trying to achieve the next big thing because competing with these

我们应该琢磨着去实现下一个大事件，因为跟这些

12:15

massively resourced, incredibly creative and talented teams is just not a winning game.

资源极其雄厚、极具创造力和才华的团队竞争，根本赢不了。

12:21

So let's play a different game and hope that that's, as they say, where the puck is

所以，让我们玩点不一样的，并希望那就是人们说的，冰球要去

12:25

going, not where it is.

的地方，而不是它现在所在的地方。

12:27

What are some examples of that kind of thinking?

这种思维方式有哪些例子？

12:29

We've been thinking a lot about architectures because I have a lot of complaints about current

我们一直在大量思考 architectures，因为我对当前的

12:33

architectures.

architectures。

12:34

And I would say the other main theme right now for us in my group is thinking about data.

而且我想说，现在我们组里另一个主要主题就是思考 data。

12:40

You know, data have strange and wondrous properties.

你知道，data 有一些奇怪又奇妙的特性。

12:43

I think we don't understand how data affect models.

我觉得我们并不理解 data 是如何影响 models 的。

12:46

And that has all sorts of implications for security and safety and also the nature of the learning

而这对于 security 和 safety，以及 learning 的本质，都有各种各样的影响

12:51

that these models do.

也就是这些 models 所进行的 learning。

12:52

And really, data is fundamental.

而且说真的，data 是根本。

12:53

It's all data-driven learning.

这全都是 data-driven learning。

12:55

And so telling the full causal story from data to final model state feels like it will

所以，把从数据到 final model state 的完整因果故事讲清楚，感觉会对很多问题都意义重大。

13:00

just be significant for lots of questions.

但我不想漏掉 architecture 这一块，因为我觉得大家最终都走到的那套 architecture，也就是我们做得非常深、非常大的这些 staffed transformers，效率极其低下。

13:04

But I wouldn't want to leave out the architecture one because I feel like the architecture everyone

我们会希望它们利用所有这些 depth 和 representational power，去学习处理各种东西的 modular、recursive functions，以及各种令人兴奋的东西。

13:10

has arrived at, these staffed transformers that we make very deep and very large are tremendously

但我们发现的并不是这样，而这似乎是一个真正的机会，让我们直接 level up、做得更好。

13:16

inefficient.

低效。

13:17

We would hope they were using all that depth and all that representational power to learn

我们本来会希望它们是在利用所有这些 depth 和所有这些 representational power 来学习

13:23

modular, recursive functions for things and all sorts of exciting stuff.

modular, recursive functions，用来处理各种东西，以及各种令人兴奋的东西。

13:28

It is not what we find and that seems like a real opportunity to just level up and do better.

但这并不是我们发现的情况，而这看起来是一个真正的机会，可以好好升级一下、做得更好。

13:33

And maybe we could get massively more capable models with half the depth and a quarter of

也许我们能得到能力强大得多的模型，只要一半的 depth 和四分之一的

13:38

the representational width.

representational width。

13:40

And that would be transformative for the economics of AI in addition to leading to all sorts

而且这对 AI 的经济性会是变革性的，同时还会带来各种

13:44

of exciting things for capabilities.

令人兴奋的 capabilities 相关成果。

13:47

It's funny and maybe a bit validating for me to hear you say that because whenever I articulate

听你这么说，我觉得挺有意思，也许还让我有点被认可，因为每当我表达

13:55

a thought in that direction, particularly with folks that are coming from the frontier

那个方向的想法时，尤其是跟来自 frontier

14:01

labs or essentially the frontier labs, I get back this feeling that you're just not

labs，或者说本质上就是 frontier labs 的人聊时，我得到的反馈是，你只是不够

14:10

bit or less impilled enough like structures, that's old school thinking, you know, you're

bitter lesson-pilled，就像，结构这东西是老派思维，你知道，你……

14:17

just trying to like train some features, just collect a lot of data, throw it at the model

就是试着 train 一些 features，收集一大堆 data，然后一股脑丢给 model

14:22

and that's all you need.

而这就是你需要做的全部。

14:23

Okay, but here's my response to them.

好，但这是我给他们的回应。

14:26

Let's say rewind to 2017, we've got the transformer.

咱们假设倒回 2017 年，我们有了 transformer。

14:30

It's got absolute positional encodings and it's got a particular structure for its MLP

它有 absolute positional encodings，而且它的 MLP

14:36

layer, which is pretty narrow and pretty dense and a certain structure to its activations

layer 有一种特定的结构，它相当窄、相当 dense，而且对它的 activations

14:42

and its layer norms.

和它的 layer norms 有某种结构。

14:44

That's 2017.

那就是 2017 年。

14:46

The bitter less impilled thing to do would be to scale that up.

按照 bitter lesson 所暗示的，该做的就是把它 scale up。

14:50

But just consider, for example, how much it would cost to use the N-squared attention and

但你就想想，比如，用 N-squared attention 要花多少钱，而且

14:56

the absolute positional encodings would have a context window of 1 million.

absolute positional encodings 会有 1 million 的 context window。

15:00

This is the bitter less impilled thing, right?

这就是 bitter lesson 所暗示的做法，对吧？

15:02

Just keep scaling, but it would be absurd.

就一直 scale 下去，但这会很荒谬。

15:04

It would cost trillions of dollars to produce models that we all interact with right now.

要做出我们现在都在用的那些 models，得花数万亿美元。

15:08

What did people do instead?

那人们转而做了什么？

15:09

They thought hard about locality and they thought about how like positional encoding should

他们认真思考了 locality，也思考了 positional encoding 应该怎样

15:13

be favoring local relationships.

更偏向局部关系。

15:16

They completely rethought the MLP so that it's now wide and sparse.

他们彻底重新思考了 MLP，让它现在又宽又稀疏。

15:20

Everyone did careful work on the activation functions to make sure there weren't weird

每个人都在 activation functions 上做了细致的工作，以确保没有奇怪的

15:24

outliers so that they could quantize in a good way and so forth and so on.

outliers，这样他们就能以良好的方式进行 quantize，诸如此类，等等等等。

15:29

All of this analysis work built on intuitions about data and learning led to the model

所有这些建立在关于数据和学习的直觉之上的分析工作，最终造就了

15:33

that we have now, which is like a ship of theses compared to the 2017 transformer.

我们现在的模型；与 2017 年的 transformer 相比，它就像一艘由论文组成的船。

15:37

The only thing that survives is attention and the feed forward layer.

唯一幸存下来的是 attention 和 feed forward layer。

15:42

And I claim for you that none of that stuff is bitter less impilled.

而且我向你断言，那些东西没有一样是 Bitter Lesson 驱动的。

15:45

That was all analysis work that was meant to save based on priors and the data and priors

那全都是分析工作，本来是想基于 priors 和 data 以及 priors 来节省。

15:50

about how they knew learning would happen.

关于他们是怎么知道 learning 会发生的。

15:52

So I go back at them.

所以我就回头去怼他们。

15:53

You're not bitter less impilled enough apparently, although this is a reductio, I think.

显然，你并不 bitter，只是还不够被推动，不过我觉得这是一种 reductio。

15:58

Oh, I love this.

哦，我太喜欢这个了。

16:00

That's such a great response.

这个回应真是太棒了。

16:02

I think it also really calls out the relationship between data,

我觉得它也真的点出了 data 与

16:09

Macinturte and efficiency like core themes that you've been focused on and how they interrelate

Macinturte 和 efficiency 之间的关系，就像你一直关注的核心主题，以及它们如何相互关联

16:16

and support one another.

并且互相支持。

16:18

Yeah, absolutely.

对，绝对没错。

16:19

And this relates to one of my hot takes, you know, it's very fashionable, especially

而这和我一个暴论有关，你知道，现在特别流行，尤其是

16:22

among interpresearchers, but I think in general for people to say, we don't understand

在 interp researchers 当中，但我觉得总体来说，大家会说，我们不理解

16:26

how these models work.

这些 models 是怎么工作的。

16:27

It is also very mysterious to us.

对我们来说，它也非常神秘。

16:29

But the truth is that people in the field have very deep intuitions about how these models

但事实是，业内的人对这些 models 如何

16:34

work and that is the causal factor in us making so much progress because they could think

工作有非常深刻的直觉，而这就是我们能取得这么大进展的 causal factor，因为他们能够思考

16:39

analytically what would the structure of positional encodings and attention be so that I could

从分析上讲，positional encodings 和 attention 的结构得是什么样，我才能在 million context scale 上做到这一点？

16:44

do this at million context scale.

你只有基于深入的分析和洞察才能做到那种事，光靠猜是不行的。

16:47

You can only achieve that kind of thing based on deep analysis and insight, not by just

所以当人们说，哦，我们不懂，我就会说，我觉得你们懂的比你们表现出来的多得多。

16:51

guessing.

我觉得你们是懂的，至少不亚于我的修车师傅懂我的车是怎么运作的。

16:53

And so when people say, oh, we don't understand, I say, I think you understand much better than

确实有一些谜团，但你可以采取很多行动，并且在改进方面非常有效。

16:58

you're letting on.

你是在暗示。

16:59

I think you understand, at least as well as my car mechanic understands how my car works.

我觉得你懂，至少跟我那修车师傅懂我的车怎么运作差不多。

17:03

There are mysteries, but you can take a lot of action and be very effective in improving

有些东西仍然是谜，但你可以采取很多行动，并且在改进上非常有效。

17:07

things.

事情。

17:08

Why do you think they say that?

你觉得他们为什么会这么说？

17:11

Why do you think they say that they don't understand the models?

你觉得他们为什么说自己不理解这些 models？

17:14

There's just got to be some payback there.

这里面肯定有什么回报。

17:15

It's probably a paradox of expertise, right?

这大概就是专业知识的悖论，对吧？

17:18

So the more you do know, the more you feel like there are also mysteries and it's hard

所以你知道得越多，就越觉得其中也还有谜团，而且很难

17:22

to step back from that and be objective and say, well, we did make a phenomenal amount

从中抽身出来、保持客观地说，嗯，我们确实取得了惊人的

17:26

of progress and that can't be just because of happenstance.

进步，而这不可能只是偶然。

17:29

That was because we know a lot, but all you see as an expert is all the things that are

那是因为我们知道很多，但作为专家，你看到的全都是那些

17:34

still to be explained.

还没被解释清楚的东西。

17:37

Partly also, it's just a narrative in the field and it does stretch back to days when I think

部分原因也在于，这只是这个领域里的一种说法，而且它确实可以追溯到那些我觉得

17:40

we had very little understanding of how these models worked and possibly because a lot

我们对这些 models 如何运作了解得还非常少的日子，也可能是因为很多

17:44

of them weren't that good that was very little to explain.

models 当时都没那么好，所以没什么可解释的。

17:47

And so that's just been slow to catch up with how much progress we have made in understanding

所以，这就一直迟迟没能跟上我们在理解方面已经取得了多大进展

17:52

the kind of intuitive human level mechanisms that these models are operating with.

也就是这些 models 运作时采用的那种直觉性的、人类水平的机制。

17:57

Also wanted to ask you about DSPY, I forgot about this as we were talking earlier, but you

另外也想问问你关于 DSPY 的事，我们之前聊的时候我忘了这个，但你

18:04

were involved in DSPY, which, well, I'll let you talk about it, but I'm curious how

参与了 DSPY，这个嘛，我会让你来讲，但我很好奇它怎么和你的研究联系起来，还有，你知道，我们聊过的一些支柱。

18:12

it connects into your research and like, you know, some of these pillars that we've talked

哦，有很多很棒的脉络。

18:18

about.

而且我能给你提供一个多么 meta 的线索啊，因为我们刚才在聊做研究要有策略。

18:19

Oh, there's lots of wonderful strands.

这确实源自我的学生 Omar Qatab。

18:21

And what a meta strand I could offer you because we were talking about being strategic

他是 DSPY 背后的远见者，现在仍然是它的负责人。

18:25

with research.

伴随着研究。

18:26

This does stem from Omar Qatab, my student.

这确实源自 Omar Qatab，我的学生。

18:29

He's the visionary behind DSPY and still its lead.

他是 DSPY 背后的远见者，现在也仍然是它的负责人。

18:33

And he just had the intuition early on that we should rethink what it means to make a scientific

他一开始就有一种直觉：我们得重新思考，做出科学贡献到底意味着什么。

18:37

contribution.

以前，我们总是从论文的角度来看，把论文当成你所能做的这类贡献的起点和终点。

18:38

Previously, we saw it in terms of papers as the beginning and the end of all of this kind

他说，我们应该转而思考项目，思考如何赋能人们。

18:43

of thing that you would contribute.

所以对他来说，论文只是一个更广泛贡献的一部分，而这个贡献实际上可能以 open source 或 open weights 的 release 为核心，让人们能够做成大事。

18:45

We should instead, he said, think about projects and about empowering people.

他说，我们应该转而思考项目，思考如何赋能人们。

18:50

And so for him, the paper is one part of a broader contribution that might actually be

所以对他来说，这篇论文只是一个更广泛贡献的一部分，而这个贡献实际上可能

18:55

centered on an open source or open weights release that would allow people to do big

以 open source 或 open weights 发布为中心，让人们能够做

19:01

things.

大事。

19:02

And that's where you find impact and that's the nature of a contribution going forward.

而那就是你能看到影响力的地方，也是未来贡献的本质。

19:06

And DSPY is a kind of embodiment of that, although he made a similar investment with

而 DSPY 就是这一点的一种体现，尽管他在

19:10

the Colbert retrieval model.

Colbert retrieval model 上也做了类似的投入。

19:13

And then people built on what he did.

然后人们在他所做的基础上继续发展。

19:15

And then you really saw it take off where open source contributions made it easier and

然后你真的看到它起飞了，open source 的贡献让那项技术变得更容易用，而且

19:19

easier to use that technology leading to more and more impact.

越来越容易使用，从而带来了越来越多的影响力。

19:22

And of course, DSPY is another wonderful example because in investing in this community,

当然，DSPY 是另一个绝佳的例子，因为通过投资这个社区，

19:29

and in the open source resource itself, he built a huge following.

以及这个 open source 资源本身，他建立了一个庞大的追随者群体。

19:33

There are lots of startups, mine included, where the core tech stack for the LMS is built

有很多初创公司，包括我的那家，它们的 LMS 核心 tech stack 都建立在

19:37

on DSPY, and that has made life so much easier.

DSPY 上，而这让日子好过太多了。

19:41

And then of course, it was a platform for him and for us to really think in an innovative

然后当然，对他和我们来说，它也是一个平台，让我们真正以创新的

19:46

way about prompt optimization and agentic workflows and all of those things.

方式去思考 prompt optimization、agentic workflows 以及所有这些东西。

19:51

Yeah, I was thinking not too long ago the degree to which model strength as a correlate

是啊，我不久前还在想，model strength 作为与

20:00

to model size, I suppose, and capability has kind of overcome the need for an explicit

model size 相关的东西，我猜，以及 capability，在多大程度上已经有点让我们不再需要一个显式的

20:09

framework like DSPY, DSPY.

像 DSPY、DSPY 这样的 framework。

20:12

Yeah, there are kind of two levels to that.

是啊，这里面其实有两个层面。

20:14

One would be just the engineering side where DSPY is great four years ago because it's

一个就是工程侧，DSPY 在四年前很厉害，因为

20:20

kind of hard to construct the code around one of these systems in a way that's modular

围绕这些系统之一来构建 code 有点难，要做到模块化

20:25

and reproducible and so forth because pecking out something where you've got a prompt

又可复现等等，因为你要敲出某个东西，里面有 prompt

20:29

string in the middle of your code with some slots in it, it's very error prone and it leads

string 夹在你的 code 中间，里面还有一些 slots，这非常容易出错，而且会导致

20:33

to bad system designs.

糟糕的 system designs。

20:35

And DSPY solved that.

而 DSPY 解决了那个问题。

20:36

And you could think that the need for that is diminishing somewhat because now we all

你可能会觉得，对这种东西的需求正在有所减弱，因为现在我们所有人都

20:41

specify these systems in English and have the coding agents do them.

用英语来指定这些系统，然后让 coding agents 去做。

20:45

And before that, there were another hundred frameworks that solve that particular part

而在那之前，还有另外一百个框架，专门解决拼图里的那个特定

20:51

of the puzzle.

部分。

20:52

Oh, yeah, there's always competition and I think at that level of just thinking about

哦，对，竞争总是有的，而且我觉得在那个层面，单纯考虑

20:55

programming interfaces and APIs, they can all learn from each other and so DSPY learned

programming interfaces 和 APIs 的时候，它们都能互相学习，所以 DSPY 从 PyTorch 学到了

21:00

a lot from PyTorch in terms of layer wise design and the kind of modularity that introduced

很多，在 layer-wise design 以及它引入的那种 modularity 方面

21:05

and then of course, you would hope that everyone kind of slurps up all these interesting

然后当然，你会希望每个人都能把所有这些有趣的

21:09

innovations and it leads to everyone being better.

创新都吸收进去，这会让每个人都变得更好。

21:12

There's lots of evidence of that at the level of interfaces.

在 interfaces 这个层面上，有很多证据能证明这一点。

21:14

I would maintain for you that even if we have agents actually writing the code for these

我还是会坚持跟你说，即使我们有 agents 真的在为这些

21:18

systems, it's great for us and for them if they write it in something that actually expresses

系统写 code，这对我们和他们都是好事，如果他们用某种能真正表达

21:22

these systems as modular components so that we can audit them so that they can change them.

这些系统为 modular components 的东西来写，这样我们能 audit 它们，这样它们也能修改这些系统。

21:27

It just feels like good engineering practices for any agent to think in a modular way and

这感觉就像，任何 agent 用 modular 的方式思考，都是很好的 engineering practices，而

21:32

that's what DSPY encodes.

这正是 DSPY 所编码的东西。

21:35

The other side is like the prompt optimization side and I believe people have that the need

另一边就像是 prompt optimization 那一侧，而我相信人们都有那种认识：

21:40

to be careful with your prompts is diminishing over time.

要小心对待你的 prompts 的必要性正在随着时间减弱。

21:43

I understand that narrative, but people should also, for example, just run like a simple

我理解这种叙事，但人们也应该，比如说，就直接跑一个简单的

21:49

annotation study where they use a few different models or the same model a few times on slightly

annotation study 里，他们会用几个不同的 models，或者同一个 model 在略有

21:54

different data.

不同的 data 上跑几次。

21:55

They will be blown away by the amount of variation that still exists.

他们会被仍然存在的 variation 之大给震到。

22:00

To be charitable, let's say that these LMs disagree about fundamental facts about how

说得客气点，这些 LMs 在一些基本事实上都达不成一致，比如到底该怎么

22:04

to label certain texts or what kind of response to give.

label 某些文本，或者该给什么样的 response。

22:08

We all kind of slip past this because we feel like, hey, they're smart and they're good

我们都有点把这一点滑过去了，因为我们觉得，嘿，它们很聪明，表现也很好

22:11

and they're getting better, but if you quantify it, it's pretty disturbing and the next step

而且还在变得更好，但如果你把它量化出来，其实挺让人不安的，而下一步

22:15

from that is to think about having all those agents optimize a prompt so that their behavior

就是去想，让所有这些 agents 去 optimize 一个 prompt，好让它们的行为

22:20

is at least consistent and then you're right back at that DSPY vision.

至少是一致的，然后你就又回到了那个 DSPY 愿景。

22:25

It's interesting that you say that because I don't feel like that necessarily aligns

有意思的是你这么说，因为我不觉得那一定符合

22:31

with my recent experience and in particular, one thing that I've noticed that's been surprising

我最近的经历，尤其是我注意到的一件很让人意外的事，

22:38

is how well aligned, I guess, maybe that's not the right word, but how similar the responses

就是它们有多 aligned，我猜，也许这个词不太对，但就是这些回复有多相似

22:47

I get to a query across different models.

我在不同模型上针对一个 query 得到的。

22:51

For example, these are often kind of what I would call a like a casual prompt, a casual

比如说，这些通常算是我会称之为一种随意的 prompt，一种随意的

22:58

query, something that I might type in the Google and it will now generate an LLM response

query，就是我可能会在 Google 里输入的东西，然后它现在会生成一个 LLM 回复

23:06

for me and it's kind of AI mode and I'll take the same thing and put it into ChatGPT and

给我，而且这有点像 AI mode，我会把同样的东西输入到 ChatGPT 里，然后

23:15

maybe Claude.

也可能是 Claude。

23:16

It surprises me that the responses are often very, very similar, like very similar structure,

让我惊讶的是，这些回答经常非常非常相似，像是结构非常相似，

23:25

very similar facts, very similar citations and stepping back, there are lots of ways

事实非常相似，引用非常相似，而且退一步看，其实有很多种方式

23:34

that they could answer or approach these different questions but it seems like the models

它们可以回答或处理这些不同的问题，但看起来 models

23:39

or the training or the system prompts or something is all kind of converged on something that

或者 training，或者 system prompts，或者别的什么，都好像收敛到了某个东西上，

23:44

makes the models express themselves very similarly which seems to be at odds with the last

让 models 表达自己的方式非常相似，这似乎和最后

23:52

thing you said about the need to optimize prompts or the impact of the individual prompt.

你说的关于需要优化 prompts，或者单个 prompt 的影响那一点相矛盾。

23:57

I'm open-minded, but for example, we did a thing recently we were writing a grant and

我持开放态度，但举个例子，我们最近做了件事，我们当时在写一份 grant，而且

24:02

we needed a title and you want to be strategic with these titles, so we come up with a

我们需要一个标题，而你想在这些标题上讲究策略，所以我们会自己先想出一

24:05

whole bunch of them ourselves and then we all disagree on what would be the best.

大堆标题，然后我们所有人对哪个最好都意见不一。

24:09

So let's find out what the agents think, so ask a few anthropic models and a few GPT

那我们就看看 agents 怎么想，去问几个 Anthropic models 和几个 GPT

24:15

models, which of these five titles, which is the best.

models：这五个标题里，哪个最好。

24:18

So you get a different answer from all of them, along with a detailed rationale about why

结果你会从它们每个那里得到不同的答案，还附上一份详细的 rationale，解释为什么

24:23

obviously, of course, the choice that the model has made in that moment is the best one.

显然、当然，model 在那一刻做出的选择就是最好的那个。

24:28

This is great because then we can think about which one of these arguments is most persuasive,

这很棒，因为这样我们就能想想这些论证里哪一个最有说服力，

24:32

but if you are hoping for consistency at a subjective labeling task, which this is one,

但如果你在一个 subjective labeling task 上指望一致性，而这正是这样一个任务，

24:38

you can see right there that you're going to have a real problem unless you give very specific

你能看到，就在那里，你将会遇到一个真正的问题，除非你给出非常具体的

24:41

criteria and then you're kind of also constructing a prompt for them and you might want to

标准，然后你其实也是在给它们构造一个 prompt，而你可能想

24:47

manage them differently.

用不同的方式管理它们。

24:48

There is a real, I don't have evidence for this yet, but we have an intuition at Bigspin

确实有一个真实的——我现在还没有证据，但我们在 Bigspin 有一种直觉，

24:53

and the research we've done that you get a kind of paradox that the more requirements you

以及我们做过的研究显示，你会遇到一种悖论：你加的要求

24:57

add, actually the more variation you'll see because the different models will key into

越多，实际上会看到越多的变化，因为不同的 models 会抓住

25:01

different sub parts of the requirements.

要求的不同子部分。

25:04

And since they do it very concertedly, you can actually get systematically biased behavior

而且因为它们非常协同地这么做，你实际上会得到带有系统性 bias 的行为。

25:10

from something that you thought was a very good specification.

从一个你原本觉得非常好的 specification 出发。

25:14

And that again calls for this idea that what you need to do is figure out what the label

而这又回到了这个思路：你需要做的是弄清楚 label

25:17

was ought to look like and then have some automatic optimization process get the model there.

本来应该长什么样，然后让某种自动优化过程把模型带到那个状态。

25:24

And that's what things like Jepa and Meet Pro were for.

而 Jepa 和 Meet Pro 这类东西就是干这个的。

25:27

A topic that I really wanted to, a topic that I would really like to dig into with you

有一个话题我一直很想，有一个话题我特别想跟你深入聊聊

25:33

based on our previous conversation was the idea of tokenomics.

基于我们之前的对话，就是 tokenomics 这个概念。

25:38

It's something that people are talking about a lot recently.

这是最近大家聊得很多的一个东西。

25:44

I think folks that use Cloud Code, for example, have a very visceral experience with anthropic

我觉得比如用 Cloud Code 的人，对 anthropic 会有非常切身的体验

25:50

changing the terms around uses, but it's happening under the covers with all of these large providers.

改变术语的使用方式，但这正在所有这些大型 provider 的底层发生。

25:57

And so I think way more now than six months ago, we're all a little antsy with the relationship

所以我觉得现在比六个月前，我们对与这些大型 model provider 的关系以及我们获得的价值都更加焦虑不安。

26:08

we have with these big model providers and the value that we get.

我们最近写了一篇文章讨论这个问题。

26:12

We recently wrote an article about this.

稍微谈谈这与你们更广泛的研究有什么关联，同时也聊聊当你们开始深入这个领域时发现的一些事情。

26:16

Talk a little bit about how it ties into your broader research, but also some of the things

是的，又是一个身处其中的有趣时刻，因为我们所有人都开始意识到所有这些 AI 使用的真实成本。

26:25

that you found when you started to dig into this area.

当你开始深入这个领域时，你所发现的东西。

26:29

Yeah, another interesting moment to be in, as we're all being made aware of the true

是的，又是一个很有意思的时刻，因为我们都开始意识到所有这些AI使用量的真实成本。

26:36

cost of all this AI usage.

我看到Ed Zitron发了一条推文，就是一张截图，有人注意到Copilot告诉他们上个月花了500美元，如果他们继续以现在这种方式使用Copilot的新构建，下个月就会是11,000美元，这真是实打实的吓人。

26:37

I saw a tweet from Ed Zitron, just a screenshot from someone who was noticing that Copilot

这里的类比就像，我以前坐网约车去机场，Uber或Lyft，花20美元，现在要90美元，但这更像是从20美元到500美元之类的，

26:44

was telling them that their last month was $500, and if they keep up the way they are

我当时在跟他们说，他们上个月花了 $500，要是他们继续照现在这样

26:49

with Copilot's new building, it will be $11,000 in the next month, which is real sticker

用 Copilot 的新 building，下个月就会变成 $11,000，这真是实打实的

26:56

shot.

价格冲击。

26:57

And the analogy here is like, I used to cost me $20 to take a ride share to the airport

这里的类比就像，我以前打网约车去机场要花 $20

27:03

Uber or Lyft, and now it costs $90, but it's more like $20 to like $500 or something,

Uber 或 Lyft，现在要花 $90，但这更像是从 $20 到 $500 之类的，

27:11

right?

对吧？

27:12

Right.

对。

27:13

If only the slope will be as shallow as it were, right?

如果斜率能像以前那么平缓就好了，对吧？

27:16

That's right.

没错。

27:17

We start to wish for those easier stories.

我们就开始怀念那些更轻松的故事了。

27:19

Yes.

是的。

27:20

And so what will happen?

那会发生什么呢？

27:21

I mean, I think what's happening is that the big providers are testing the waters on

我的意思是，我觉得现在的情况是，那些大provider正在试探性地

27:25

charging us the true costs, plus whatever profit they need to make as they all try to

向我们收取真实成本，再加上他们为了准备IPO之类的事情所需要赚取的利润。

27:30

gear up for IPOs and so forth.

为 IPO 之类的事情做准备。

27:33

And in turn, that is very quickly leading people to ask questions like, what is the return

反过来，这也很快让人们开始问这样的问题：

27:37

on investment for all these tokens that we have purchased?

我们买的这些 tokens，投资回报率到底是多少？

27:42

And it's a very tricky area to be in because what does it mean to think about value in

而这是一个非常棘手的领域，因为在这个语境下，

27:47

this context, even if we focus in on people who are doing just coding with coding agents?

思考价值意味着什么，即使我们只聚焦在那些用 coding agents 做编程的人身上？

27:53

Can we agree on what it means to add value?

我们能对“增加价值”的含义达成一致吗？

27:55

I mean, we have a few measures in mind like making a pull request or committed lines of

我的意思是，我们心里有几个衡量标准，比如发起一个 pull request，或者 committed lines of code

28:01

code that last in the repo for a while, or documentation touched or skill files created,

能在 repo 里留存一段时间，或者 documentation 被改动，或者 skill files 被创建，

28:08

but we might also worry that that's not capturing the value for many kinds of sessions we have,

但我们也可能担心，这并没有捕捉到我们许多种 sessions 的价值，

28:14

which are more open-ended and about discovery.

这些更开放，也更关乎探索发现。

28:16

So that's the first question is just solving this value issue, right?

所以第一个问题就是解决这个价值问题，对吧？

28:20

Let's just agree on what it would mean to add value for a coding agent.

我们先就什么叫做为 coding agent 增加价值达成一致吧。

28:24

I like this line of inquiry because, to me, it's the response to this thing that drives

我喜欢这条探究思路，因为对我来说，它是对这件事的回应，这件事让我

28:33

me crazy, which is, oh, big tech company CEO.

发疯，那就是，哦，大型科技公司 CEO。

28:41

This year, 95% of our code will be generated by AI.

今年，我们 95% 的代码将由 AI 生成。

28:47

It's like, yeah, what does that really mean at that level?

就好像，是啊，在那个层面上这到底意味着什么？

28:53

Out of the details beneath there, but is that a good thing for a man?

抛开底下那些细节不谈，但那对一个人来说是好事吗？

28:59

Oh, another dimension, right?

哦，还有一个维度，对吧？

29:00

Which is that code of liability or an asset?

也就是说，那个 code 到底是负债还是资产？

29:03

Right.

对。

29:04

Right.

对。

29:05

And this idea of like you articulate as kind of code longevity in the code base, that's

而且你把它阐述成 code base 里的某种 code longevity，这个想法挺有意思的。

29:10

an interesting way to think about it.

可能有很多有意思的思考角度，但现在几乎没人在想。

29:11

There's probably a lot of interesting ways to think about it that very few are thinking

大概有很多有趣的思考角度，但现在很少有人会想到

29:15

about right now.

这些，就在当下。

29:16

And all of these fall victim to the standard thing that once you make it a metric, it's

而所有这些都会掉进那个老问题里：一旦你把它变成一个 metric，它

29:19

no longer useful to you.

就不再对你有用了。

29:21

If we said, well, it's completion of projects, right?

如果我们说，好吧，那就是看项目的完成情况，对吧？

29:24

Well, then everyone would just have many projects that they completed, but they could all

那大家就只会有一大堆自己完成的项目，但它们可能全都

29:28

be liabilities and add very little value.

是 liabilities，带来的价值少得可怜。

29:31

But one framework we could offer that we did in the research you alluded to is, let's

但我们可以提供的一个 framework，是我们在你提到的那项研究里做过的，那就是，让我们

29:34

think about this like economist might, so we might have like a consumer price index.

像经济学家那样来思考这件事，比如我们可能会有类似 consumer price index 的东西。

29:39

And the first step will be what's the basket of goods that we're going to consider, you

而第一步会是，我们要考虑的 basket of goods 是什么，你

29:43

know?

知道吗？

29:44

In that standard land, it would be like the price of eggs and the cost of rent and other

在那种标准语境里，它就会像鸡蛋的价格、房租和其他

29:48

kinds of tangible goods.

各种有形商品。

29:50

What are engineering goods that we might track?

那我们可能追踪的 engineering goods 有哪些？

29:52

But eggs might be a summary or a pull request, a bug fix or something like that.

但鸡蛋可能是一个 summary、一个 pull request、一个 bug fix 之类的东西。

29:58

Or we could think broadly because we both use these coding agents, requirement discovery,

或者我们可以想得更宽泛，因为我们都用这些 coding agents、做 requirement discovery，

30:04

right?

对吧？

30:05

Knowledge accumulation.

Knowledge accumulation。

30:07

These are things that we don't currently track, of course, even as engineers, but might

这些是我们目前不追踪的东西，当然，即使作为工程师，但它们可能

30:11

be behind our intuition that these coding agents are making us productive, even if it's

是我们直觉背后的原因，即这些 coding agents 让我们更高效，即使它

30:15

not reflected in the PR counts or whatever, right?

没有反映在 PR 数量或其他什么上，对吧？

30:18

I mean, in a sophisticated approach, you might say, I don't want more PRs because this

我是说，在一个成熟的方法中，你可能会说，我不想要更多的 PRs，因为这

30:22

is just a certain kind of busy work that doesn't relate to the actual goals I have.

只是某种瞎忙，与我实际的目标无关。

30:27

What are the actual goals?

实际的目标是什么？

30:28

It's completing valuable projects and so forth.

是完成有价值的项目等等。

30:30

I could do it with fewer PRs, but I had, you know, really robust code.

我可以用更少的 PRs 来完成，但是，你知道，我有真正 robust code。

30:35

I'd be possibly happy with that.

我可能对此会挺满意。

30:37

So we got to figure out what the basket of goods is, but then we could start to track it relative

所以我们得搞清楚 basket of goods 是什么，但之后我们就可以开始跟踪它相对于

30:41

to token usage.

token usage 的情况。

30:42

And that would be the consumer price index.

那就会是 consumer price index。

30:44

So for any time period, we could just say, I've got my tokens spent and I've got my

所以对任何时间段，我们都可以直接说，我有我花掉的 tokens，我也有我

30:48

goods produced, tokens divided by goods produced is a pretty rough measure of the purchasing

产出的 goods，tokens 除以 goods produced 是衡量那些时间段里 token 的 purchasing

30:56

power of the tokens in those time periods.

power 的一个相当粗略的指标。

30:59

Then you would do the standard consumer price index thing of making what they call a hedonic

然后你就会做标准的 consumer price index 那套，做他们所谓的 hedonic

31:03

adjustment.

调整。

31:04

So you could just say, maybe quality is improving over time, so you'd pick some measure

所以你可以就说，也许质量随着时间在提升，那你会选某个指标

31:07

for that and make an adjustment to the line.

据此对那条线做个调整。

31:10

And when we did that study, we did code survival.

我们做那项研究时，用的是 code survival。

31:13

So the number of lines of code that survives more than four days in the repository, we made

就是在 repository 里存活超过四天的 lines of code 数量，我们做了

31:18

an adjustment upward because that rate is going up.

向上调整，因为这个比率在上升。

31:21

That's surprisingly short.

这短得有点意外。

31:23

Four days.

四天。

31:24

Four days survival.

四天的生存期。

31:25

You could miss.

你可能会漏掉。

31:26

So again, all this is around measurement and I'm happy to just be starting this discourse

所以再说一次，这一切都是围绕 measurement 的，而我很高兴能刚开始这场讨论

31:30

because we can see it's important to the economics of AI and it seems like the work

因为我们可以看到，这对 AI 经济学很重要，而且看起来这项工作

31:33

isn't being done at a high enough rate for us to get a clear picture.

没有以足够高的速度推进，让我们无法看清全貌。

31:37

So we could make it longer and maybe the adjustment would be different.

所以我们可以把它做长一点，也许调整就会不一样。

31:40

I think currently for the data we have, which is this sweet chat benchmark, which was released

我觉得目前就我们有的数据来说，也就是这个 sweet chat benchmark，它是由

31:45

by researchers at Stanford, it's about 6,000 real coding sessions, all the metadata, everything

Stanford 的研究人员发布的，大约有 6,000 个真实的 coding sessions，所有 metadata，所有东西

31:51

you'd want.

你会想要的。

31:52

What we see with Opus 4.6 usage in the time period we have, which is February to mid-April

在我们目前有的这个时间段里，也就是今年2月到4月中旬，我们从 Opus 4.6 的使用中看到的是，tokens 的购买力在下降。

31:58

of this year, a decline in the purchasing power of tokens.

这个 CPI 正在下降。

32:02

That CPI is going down.

再说一次，我只是想提出这个问题：是因为我们的商品篮子选错了，还是因为我们实际上从这些 tokens 中得到的价值更少了？

32:05

And again, I just want to open the question, is it because we have the wrong basket of

有一点我可以肯定，这背后绝对有一个因果因素：今年2月，大多数 tokens 都用于生成 code，而这跟我们刚才谈到的结果有关。

32:10

goods or is it because we're actually getting less value from these tokens?

商品，还是因为我们实际上从这些 tokens 中获得的价值更少了？

32:15

The one thing I can say that's definitely a causal factor here is that in February of

我能确定地说，这里肯定有一个因果因素，那就是

32:21

this year, most of the tokens went to producing code, which relates to the outcomes we just talked

在今年二月，大部分 tokens 都用于生成 code，这和我们刚谈到的结果有关

32:27

about.

大概。

32:28

By mid-April, it was quite split between code generation thinking and also explanation

到四月中旬，它在 code generation 的思考和

32:34

to the user.

给用户的解释之间已经相当分裂了。

32:36

And so that split now is going to have an effect on the things we're measuring and that

所以这种分裂现在会影响我们正在衡量的东西，而那

32:40

might be cause for reflection.

可能值得反思。

32:42

There's value in those explanations that's not reflected in PRs, but might be reflected

那些解释里的价值并没有反映在 PRs 里，但可能反映

32:46

in something like knowledge discovery.

在像 knowledge discovery 这样的东西里。

32:48

And I see that coming up within the same time frame, it's become very common to now

而且我看到，在同一时间段内，这一点也出现了，现在变得非常常见的是

32:55

talk about the token efficiency of a new model that's been released with the implication

聊聊一个刚发布的新模型的 token efficiency，它暗示着

33:01

of being tokens of internal use tokens, thinking tokens versus, you know, per token

这些是 internal use tokens、thinking tokens，相对于，你知道，每个

33:09

of output, I guess is maybe a way to think about it.

output token，我猜这也许是一种理解方式。

33:11

Another fascinating dimension.

另一个很有意思的维度。

33:13

And this actually relates all the way back to the theme of efficiency for these architectures.

而这其实一路都能联系回这些架构的效率主题。

33:17

So here's a claim I'll make for you.

所以我来给你提个主张。

33:19

Based on my read of the literature on inference time scaling, what's sometimes called test

根据我对 inference time scaling 文献的阅读，它有时被称为 test

33:25

time scaling, which is just having the models generate lots of tokens at the moment that

time scaling，也就是让模型在那个时刻生成大量 token

33:30

you ask them a question.

你问他们一个问题。

33:31

So those scaling trends, everything we're seeing now is completely in line with those predictions,

所以那些 scaling trends，我们现在看到的一切都完全符合那些预测，

33:36

which is you get pretty good gains for a while with the more tokens you spend on a log

也就是说，随着你把更多 tokens 花在 log

33:42

scale.

scale 上，你会在一段时间里获得相当不错的提升。

33:43

So this is jumping up quite a lot, but you do see it reflected in performance improvements,

所以这个涨得挺多的，但你确实能看到它反映在 performance improvements 上，

33:47

but it flattens out over time.

但会随着时间逐渐平缓下来。

33:50

And it's not like this curve sky rockets.

而且并不是说这条曲线会一飞冲天。

33:53

It's sobering.

这挺让人清醒的。

33:54

You got to spend a lot of tokens for small gains in performance.

你得花很多 tokens，才能换来 performance 上的一点点提升。

33:59

We all do this.

我们都会这么做。

34:01

We all do this.

我们都会这么做。

34:02

And we're just seeing it now play out.

而我们现在只是看到它正在上演。

34:03

And when people talk about token efficiency and worry about this, I think what they're seeing

当人们谈论 token efficiency 并为此担心时，我觉得他们看到的

34:06

is just the real lesson of what we already projected from inference time scaling.

其实只是我们早就从 inference time scaling 中预测到的真正教训。

34:11

And this is independent of the approach to inference time scaling you're taking, whether

而这和你采用的 inference time scaling 方法无关，无论

34:15

it's, you know, multiple parallel, you know, multiple parallel inferences or some kind

是，你知道，多个并行，你知道，多个并行 inferences，或者某种

34:23

of Oracle or, you know, any number of other schemes, it's just fundamental to inference

Oracle 的，或者，你知道，其他任意数量的方案，这对 inference

34:29

time scaling.

time scaling 来说就是根本性的。

34:30

That's a great question, right?

这是个好问题，对吧？

34:31

I think we know that it's independent of some of those things like the parallel work

我想我们知道，它不依赖于其中一些东西，比如 parallel work

34:36

versus having it do lots of long chains, but some of the other factors you mentioned,

而不是让它做大量 long chains，但你提到的其他一些因素，

34:40

I think we just don't know.

我想我们就是不知道。

34:42

And that's why I said it relates back to the question of efficiency for these architectures.

这就是为什么我说，它又回到了这些 architectures 的 efficiency 问题。

34:46

If we made a fundamental change to how the models work, maybe these trade-offs would

如果我们对 models 的工作方式做出根本性的改变，也许这些 trade-offs 会

34:51

be very different.

会非常不一样。

34:52

I mean, after all, so all of this stuff is a kind of patch job on the fact that there's

我的意思是，毕竟，所以所有这些东西都算是在给一个事实打补丁，那就是

34:57

no recursion in the depth.

depth 里没有 recursion。

34:59

It's a fixed depth.

它就是一个固定 depth。

35:01

And so the only recursion we can get, the on open ended notion of computation is by generation.

所以，我们唯一能得到的 recursion，也就是那种开放式的 computation 概念，是通过 generation 来实现的。

35:06

But if we had models that could be recursive, maybe fewer tokens for larger gains, I think

但如果我们有可以 recursive 的模型，也许更少的 tokens 就能换来更大的收益，我觉得

35:11

we don't know.

我们不知道。

35:12

Yeah, I mean, in the end, we're going to spend the cost on compute or tokens.

是啊，我的意思是，到最后，我们反正都要把成本花在 compute 或者 tokens 上。

35:17

So this might not affect our bills in the end, but it is a fascinating question.

所以这最终可能不会影响我们的账单，但这是个很吸引人的问题。

35:21

What are the true scaling laws and what's possible in this space?

真正的 scaling laws 是什么，在这个领域里什么又是可能的？

35:24

And you're right to push back.

你说得对，是该反驳一下。

35:26

We talk about these things like they were like platonic ideals of laws, scaling law and

我们谈论这些东西时，就好像它们是某种柏拉图式的理想法则、scaling law，以及

35:31

folks that, right?

诸如此类的东西，对吧？

35:33

But even for the scaling laws for pre-training, you know, there's lots to discover there.

但即便是 pre-training 的 scaling laws，你知道，那里也还有很多东西有待发现。

35:37

And many of the stories of progress are actually like transcending the scaling law.

而且很多关于进步的故事，其实都是在超越 scaling law。

35:42

We see like better improvements than those laws predicted because everyone worked so hard

我们看到比那些 scaling laws 预测的还要好的提升，因为每个人都非常努力。

35:46

behind the scenes to do very innovative things, which maybe relates to our bitter lessons

在幕后做一些非常创新的事情，这也许和我们那些惨痛教训的

35:50

discussion.

讨论有关。

35:51

Any particular example come to mind of that?

这方面有没有什么特别的例子浮现在你脑海里？

35:54

Data usage in the nature of the data really matters and overtraining the models really matters,

数据的使用方式、数据本身的特性真的很重要，而 overtraining 模型也很重要，

35:58

which is kind of pushing up against the standard scaling law presentation.

这有点是在跟标准的 scaling law 说法相抵触。

36:02

And now I'm just going to speculate, I should check on this.

现在我只是在推测，我应该去核实一下。

36:04

But things like mixture of experts might have really flipped the script on what it means

但像 mixture of experts 这样的东西，可能真的彻底改写了“计算 parameters 意味着什么”这件事

36:09

to count parameters and then turn how these laws relate.

以及这些 laws 之间如何关联。

36:11

And then I think maybe even also stuff like the context window and so forth.

然后我觉得，可能甚至还包括像 context window 之类的东西，等等。

36:15

This is another thing to check, but I just speculate that we've seen larger gains from

这是另一个需要核实的事情，但我只是猜测，我们看到的更大收益来自

36:20

pre-training than you would have predicted by those early scaling laws papers suggesting

pre-training，比你从那些早期 scaling laws 论文里会预测到的还要大，那些论文暗示

36:24

that there is some innovative thing that was happening on top of pure scaling.

在纯粹的 scaling 之上，还有一些创新性的东西在发生。

36:29

Thinking about the concept of a market basket, one kind of pushback that comes up for me is

想到 market basket 这个概念，我脑海中会冒出的一种质疑是

36:38

in the, you know, the real economy, you know, A's is different than milk is different than

在，你知道，实体经济里，你知道，A's 和牛奶不一样，和

36:46

beef, et cetera, et cetera.

牛肉不一样，等等，等等。

36:49

And they're all influenced by different factors, you know, production, for example.

而它们都受到不同因素的影响，你知道，比如生产。

36:57

Whereas what you've done with this kind of CPI basket with tokens is kind of like more

而你用这种带 tokens 的 CPI basket 做的事情，更像是

37:06

like analogies, like here's the typical bundle of work and, you know, what it requires

一种类比，比如这是典型的一组工作，而且，你知道，它需要什么

37:11

from a consumption perspective.

从消费的角度来看。

37:14

The tokens aren't fundamentally different, like they're the same tokens.

tokens 本质上没什么不同，它们就是同样的 tokens。

37:17

It's just like how much it takes to do this versus how much it takes to do that versus

就只是像做这个需要多少，对比做那个需要多少，再对比

37:21

how much it takes to do that.

做那个需要多少。

37:22

You know, tell me what I'm, what I'm missing there and, you know, what does kind of characterizing

你知道，告诉我，我，我漏了什么，而且，你知道，算是刻画

37:30

these products, you know, give you in your analysis.

这些产品，你知道，在你的分析里到底能给你带来什么。

37:34

Yeah, fascinating to think about.

是啊，想想挺有意思的。

37:36

One thing I could insert there is the tokens are different at the level of being used

我能补充的一点是，tokens 在被用于

37:40

for code generation or skill file writing or explanation or thinking, right?

code generation，或者写 skill file、解释或思考时，是不一样的，对吧？

37:46

Those are different kinds of tokens that probably do feel tangibly different to us.

那些是不同类型的 tokens，对我们来说，很可能确实能感觉到明显不同。

37:51

So is that an element in your think?

所以，这是你思考中的一个因素吗？

37:52

Because I think I was thinking from our conversation that you had 10 different, almost like tasks,

因为我觉得，我是从我们的对话里想到，你有 10 个不同的、几乎像是 tasks 的东西，

37:59

like 10 different types of tasks from the domain of code generation, which, you know,

像是来自 code generation 这个 domain 的 10 种不同类型的 tasks，你知道，

38:05

if they were all kind of largely code generation, you know, that is the part that had some

如果它们大体上全都是 code generation，你知道，那就是有某种……的那部分

38:09

dissonance for me.

对我来说有种违和感。

38:10

But if you're talking about, like if your, your basket is like creative writing versus,

但如果你说的是，比如你的，你的 basket 是像 creative writing，对比于，

38:16

you know, a few code generation things that are kind of in different versus, you know, summarization

你知道，几个 code generation 的东西，它们有点属于不同的类别，对比于，你知道，summarization

38:24

versus editorial, commenting feedback, those, you know, may be more fundamental.

对比于 editorial、commenting feedback，那些，你知道，可能更根本。

38:30

Yeah.

嗯。

38:31

So I think this is very significant when we have done some research on this as well at

所以我觉得这非常重要，因为我们自己在这方面也做过一些研究，在

38:35

the level of what kinds of session types exist and in turn, what kinds of users are there.

这个层面上：存在哪些 session types，以及反过来，有哪些类型的用户。

38:40

So you might notice of your own behavior.

所以你可能会注意到你自己的行为。

38:42

I guess this is reflected in your comment that sometimes you want to quick check in on

我猜这也体现在你说的那句话里，就是有时候你想快速确认一下

38:45

a question.

一个问题。

38:47

Sometimes you want a quick PR to get fired off.

有时候你想快速发一个 PR 出去。

38:51

Sometimes you want to be in a mode of deep collaboration.

有时候你想进入一种深度协作模式。

38:54

Sometimes you're partnering with the AI.

有时候你是在和 AI 搭档。

38:55

Sometimes you're delegating the work and so forth and so on.

有时候你是在把工作委派出去，诸如此类，等等等等。

38:59

And the outcome measures that we choose should be sensitive to this.

而我们选择的 outcome measures 应该对这一点敏感。

39:02

We shouldn't penalize the agent if your chat interaction with it about some scientific

如果你的 chat interaction 是和它围绕某个 scientific 的，我们不应该惩罚 agent

39:07

question didn't lead to a PR was never on the table in the first place.

问题并没有导向一个 PR，这件事从一开始就根本不在讨论范围内。

39:12

Whereas if you're trying to get some work delegated, that's actually a coding task and

但如果你是想把一些工作委派出去，那其实是个 coding task，而且

39:16

all it does is chat with you, that would feel quite unproductive.

它做的只是跟你聊天，那就会让人觉得挺没产出的。

39:20

So we need to bring that in and that would be a higher level discovery process of what

所以我们需要把这一点引入进来，而这会是一个更高层次的 discovery process，去搞清楚人们

39:24

people are trying to do and so forth.

想做什么之类的事情。

39:27

And thinking about the notion of value, is this something that you're anticipating?

再想想 value 这个概念，这是你预期中的东西吗？

39:35

It strikes me that that's an entire research thread that one could go into.

我觉得这整个就是一条可以深入研究的 research thread。

39:41

I don't know if that's a linguist or an economist or computer scientist, probably interdisciplinary.

我不知道这算是语言学家、经济学家还是计算机科学家的事，大概是跨学科的。

39:50

Most interesting questions.

最有趣的问题。

39:52

Is that something that you're working on or was it something that you put out there

那是你正在做的东西，还是你把它抛出来，让别人接手去推进的？

39:59

for someone to take up and run with?

我不确定。

40:01

I am not sure.

我可以告诉你，这个想法的脉络是这样的：我们创办了 Big Spin 这家 startup，因为我们希望看到更多人从 AI 中受益。

40:02

I can tell you the lineage of this idea is that we founded this startup Big Spin because

不管你爱它还是恨它，它都已经来了。

40:08

we would like to see more people benefit from AI.

而且我希望这些好处能分配得更均匀。

40:12

Whether you love it or hate it, it's here.

不管你爱它还是恨它，它都已经来了。

40:15

And I would like the benefits to be more evenly distributed.

而我希望这些好处能分配得更均匀一些。

40:19

And I can tell that that will mean bringing on board many more people than currently

40:25

benefit from AI.

40:26

Right now, I would say that it's mostly experts deriving real value.

我能看出，这意味着要让比现在多得多的人受益于 AI。

40:31

And a lot of the world is currently even trying to figure out what this is all about as a

Right now, I would say that it's mostly experts deriving real value.

40:34

tool or an entity in their lives.

现在的话，我会说主要还是专家在获取真正的价值。

40:37

So we would like to have more access and more productivity.

And a lot of the world is currently even trying to figure out what this is all about as a

40:41

And that implies making the user experience is much better.

tool or an entity in their lives.

40:46

Trying out what interaction patterns lead to success for people, meeting them where

而世界上很多人现在甚至还在试图搞清楚，这个东西作为一个工具，或者说作为他们生活中的一个存在，到底意味着什么。

40:50

they are in a kind of adaptive way.

它们是以一种自适应式的方式。

40:52

The whole list of things that you might worry about if you were a product manager who

如果你是一个产品经理，手上有一个已经部署的 AI 产品，

40:55

had some deployed AI product.

那你可能会担心的所有事情。

40:58

And I think by that root and from that perspective, we just ended up worrying about our own

而我觉得，从那个根源、从那个视角出发，我们最后就只是在担心我们自己的

41:04

token usage increasing and wondering whether there's real value there.

token 用量在增加，并且怀疑那里是否真有价值。

41:07

And it just happened to collide actually just like three weeks ago with this emerging

而且它实际上就在大概三周前，刚好撞上了这个正在兴起的

41:12

narrative on the back.

背后的叙事。

41:13

I think of all these rumors about IPOs, about what the return on investment was.

我想到所有这些关于 IPO 的传闻，关于投资回报到底是多少。

41:18

And then all these CEOs came out and said, all our spend was enormous.

然后所有这些 CEO 都出来说，我们的支出全都大得惊人。

41:21

And we want to scale back and we're walking back our claims from a few months ago.

而且我们想 scale back，也在收回几个月前说过的话。

41:25

And that is just a fascinating thing to witness in the narrative here.

而在这整个叙事里，这真是让人叹为观止的一幕。

41:29

I'm wondering are you also, does this research also attempt to project forward?

我很好奇，你们是不是也——这项研究是不是也尝试向前推演？

41:35

In theory, you could create a model for anthropics cost and spend based on publicly available

理论上，你可以根据公开可用的

41:46

data and some presumptions and give us a sense for how close we are to paying full freight

数据和一些假设，为 Anthropic 的成本和支出建一个 model，并让我们大致了解

41:55

for our tokens versus if we're only paying 10% for our tokens.

我们离为我们的 tokens 支付全价有多近，还是说我们只为我们的 tokens 支付 10%。

42:00

You could then project what that cost might look like over time as we're paying more

然后你就能推演，随着我们付得越来越多，这个成本随时间可能会变成什么样。

42:06

and more of the full cost.

以及更多完整成本。

42:08

Yeah, I don't have again fascinating questions.

是啊，我还是没有那些引人入胜的问题。

42:10

I don't have resolving answers.

我没有能解决问题的答案。

42:11

I am glad I am not tasked in some organization with projecting spend on all of this stuff

我很庆幸自己没有在某个组织里被安排去预测所有这些事情的支出

42:16

because I think it would be basically possible.

因为我觉得这基本上是有可能做到的。

42:18

For the time period that I was describing for our little CPI experiment, anthropic change

在我为我们那个小小的 CPI 实验所描述的那段时间里，anthropic 改了

42:25

the default reasoning on the model at least two times.

model 上的 default reasoning，至少两次。

42:28

So we see like they launched it with default reasoning high.

所以我们看到，他们发布它的时候 default reasoning 设得很高。

42:32

We have a mysterious sudden rise in the token usage, which we cannot explain.

我们的 token usage 出现了一次神秘的突然上涨，我们无法解释。

42:36

And there's a new baseline.

而且现在有了一个新的 baseline。

42:37

They lowered it to medium as the default.

他们把它默认调低到了 medium。

42:39

They patched a bunch of bugs that were related to context management and then turned it back

他们修了一堆跟 context management 相关的 bug，然后又把它调回了 high。

42:43

up to high.

而且所有这些都会影响总的 token output。

42:45

And all of these things have an effect on the total token output.

你可以想象，他们还改了默认的 context window，这意味着人们能在任何时刻吞下多得多的东西。

42:47

As you can imagine, they also changed the default context window, which meant people

你可以想象，他们还改了默认的 context window，这意味着人们

42:50

could swallow up much more stuff at any given moment.

能在任何特定时刻吞下多得多的东西。

42:55

So imagine trying to predict what token spend is going to be like when you have all these

所以想象一下，要预测 token spend 会是什么样，当你有所有这些

42:59

exogenous events in addition to changes that we don't even know about and questions about

exogenous events，再加上我们甚至都不知道的变化，还有关于

43:06

where the value actually lies.

价值到底在哪里的问题。

43:08

Very difficult.

非常难。

43:09

And then the true cost of a token, the estimates very wildly for every dollar we spend

然后是一个 token 的真实成本，每花出去一美元，估算会非常剧烈地波动

43:14

that could be as low as two and as high as 20.

可能低到 2，高到 20。

43:18

And I think this is just because it's hard to factor in things like R&D and future build

我觉得这只是因为很难把像 R&D 和未来的 build

43:22

out and depreciation and all of that stuff.

out、depreciation 以及所有这些东西都考虑进去。

43:25

I think at the current moment, we just don't know, but there couldn't be a more significant

我觉得眼下我们还不知道，但对全球经济来说，基本上没有比价值在哪里、谁来付钱、付多少更重要的问题了。

43:29

question for the global economy, basically, than where the value is and who's going to

是啊。

43:35

pay and how much.

在你的文章里，你创造了 tokenflation 这个词，至少用来描述最近 token economics 的表现。

43:37

Yeah.

我猜你觉得这种情况会继续下去。

43:38

In your article, you coined the term tokenflation to describe at least the recent behavior

看起来确实在继续。

43:44

of token economics.

token economics 的。

43:47

I imagine you see that continuing.

我想你看到这种情况在持续。

43:49

Seems to be continuing.

看起来确实还在持续。

43:50

Yeah.

43:51

Yeah.

43:52

So the picture that we get from the CPI, a picture of tokenflation, yes, your token

所以从 CPI 里我们看到的图景，就是 tokenflation，是的，按我们这里能想到去衡量的一切来看，你的 token 已经买不到过去能买到的东西了。

43:56

is not buying you what it once did, according to everything we can think to measure here.

而且就算把模型变好这个因素也调整进去，对吧？

44:02

And even adjusting for models getting better, right?

这一点很关键，因为如果这只是一个模型思考更多、变得更稳健，而我们所有人从中获得的结果都在指数级变好的故事，那这些支出看起来就会完全合理，但我们看到的并不是那幅图景。

44:04

That's critical there because if it was just a story of models thinking more and being

这很关键，因为如果这只是一个关于 models 想得更多、并且变得...

44:08

more robust and we were all getting exponentially better outcomes from this, then the spend

更加健壮，而且我们都能从中获得指数级提升的结果，那么这笔支出

44:13

would look completely rational, but that's not the picture that we see.

看起来就完全合理，但我们看到的并不是这样一幅图景。

44:16

And so we have to do some hard thinking about what's going to happen and how to improve

所以我们必须认真思考一下会发生什么，以及如何改善

44:20

the situation.

这种情况。

44:22

Let's dig into that a little bit more.

让我们再稍微深入聊一聊这个。

44:25

How would you articulate what you're seeing?

你会怎么描述你所看到的情况？

44:26

The models are getting, quote, unquote, better.

这些模型正在变得，所谓，更好。

44:31

There's a set of open questions about, are the reported ways that models are better,

有一系列开放问题：那些被报告出来的模型变好的方式，

44:39

actually reflective of some intrinsic betterness, like in that question brings, it is often

是否真的反映了某种内在的更好，就像那个问题引出的那样，它往往

44:47

about like benchmarking and learning the benchmarks overfitting that kind of thing.

是关于像 benchmarking 和 learning the benchmarks、overfitting 这类事情。

44:53

And then there's the kind of question of chattiness and the volume of thought that it requires

然后还有一个问题，就是聊天量和

45:04

a given model generation to produce an answer.

某一代模型生成一个答案所需的思考量。

45:07

What are other factors that you see?

你还看到哪些其他因素？

45:09

Yeah, we can pick that apart as well.

是的，我们也可以把这一点拆开来讲。

45:11

So and this relates to a line I've had consistently, which is that we should think

所以，这和我一直坚持的一条线有关，那就是我们应该

45:17

in terms of systems, not in terms of models.

从系统的角度来思考，而不是从模型的角度来思考。

45:19

So in the data that we've got, Sonnet, as Opus and Sonnet, 4.5 versus 4.6, those two

所以，在我们拿到的数据里，Sonnet——拿 Opus 和 Sonnet 来说——4.5 对 4.6，那两次 generation 变化，我猜，都是真正的 model 变化。

45:26

generation changes, those are real model changes, I assume.

但另一方面，你知道，这里面有个事儿，它也意味着，你不该在那些你没有足够专业能力去评估答案的领域使用这些东西，可那又恰恰是你最需要帮助和支持的地方。

45:30

I think they did something very substantive at the level of the weights.

我觉得他们在 weights 层面做了非常实质性的工作。

45:34

And everybody immediately saw that that led to like a 5x increase in token usage.

而且大家立刻就看到，这导致 token 使用量大概增加了 5 倍。

45:39

And this was related to the introduction of adaptive thinking.

而这跟 adaptive thinking 的引入有关。

45:44

Now fix that.

现在修复这个问题。

45:45

That's the level shift that we already took, and maybe we're seeing improvements there

那是我们已经经历过的 level shift，也许我们在那里看到了一些改进

45:49

that are worthwhile.

这些改进是值得的。

45:50

It gets hard to say, but let's assume there was a level up in improvement.

这很难说，但我们假设改进上有一个 level up。

45:54

Then for the period that we did our CPI experiment for, that's a fixed model, Opus 4.6.

然后，在我们做 CPI 实验的那段时间里，那是一个固定模型，Opus 4.6。

46:01

So all the code improvements that we saw in the data relate to the product.

所以我们在数据里看到的所有 code 改进，都和产品有关。

46:06

This has to relate to things like them turning the knobs on the adaptive thinking, changing

这肯定和这些事情有关，比如他们去调 adaptive thinking 的那些旋钮、改变

46:10

things about the system prompt, changing things at the level of the product, and that's

system prompt 相关的东西、在产品层面做改动，而

46:14

where the improvements were.

改进就出在这些地方。

46:16

And so that shows you that even for a fixed model, we can get varied for now comes through

所以这说明，即使 model 是固定的，我们也能得到变化，现在这些变化来自

46:20

these things because they really are sophisticated engineered systems at this point.

这些东西，因为它们现在真的是很精密的工程化系统。

46:24

Yeah.

对。

46:25

And so was the product in this case specifically, cloud code or, oh yeah, and so we know

所以在这个情况下，产品具体是 cloud code 吗？还是，哦对，所以我们知道

46:31

there's tons of stuff there.

那里有超多东西。

46:33

Yes, I believe we know that these are all cloud code sessions that we kept in our data.

是的，我相信我们知道，这些都是我们保存在数据里的 cloud code sessions。

46:37

Sweet chat is broader than that and involves a couple of other coding agents.

Sweet chat 比这更广，还涉及另外几个 coding agents。

46:40

But I think I can say that all our data, our cloud code sessions using Opus 4.6.

但我想我可以说，我们所有的数据，我们使用 Opus 4.6 的 cloud code sessions。

46:45

Have you seen any evidence that changes via API usage, experience, similarly dramatic

你有没有看到任何证据表明，通过 API 使用发生的变化，会经历类似戏剧性的

46:56

variation in performance?

性能波动？

46:59

Oh, fascinating.

哦，有意思。

47:00

To kind of control for a lot of that product level stuff, all the prompts that are hidden

为了稍微控制很多那种产品层面的东西，所有那些被隐藏的 prompts。

47:04

from us, all of those affordances.

来自我们，所有那些 affordances。

47:06

Yeah, I don't know.

是啊，我也不知道。

47:07

But that's a nice thing to think about because it gives us more things that we can control

但这是件值得思考的好事，因为它给了我们更多可以控制的东西

47:12

for and more things that are knowable.

以及更多可知的东西。

47:14

So kind of in parallel to the model evolution, there's also evolution of the user you've alluded

所以，某种程度上，和 model 的演化并行，用户也在演化，你提到过

47:22

to this a little bit about kind of your concept is that most AI users now are experts.

这一点，你的概念大概是：现在大多数 AI 用户都是专家。

47:31

Talk a little bit about the role of expertise.

稍微聊聊专业能力的作用。

47:34

I think this is also kind of echoing back to our conversation about DSPY and like prompt

我觉得这也有点像是在呼应我们之前关于 DSPY 和 prompt 的对话。

47:40

optimization.

optimization。

47:42

You've done some research into how folks are using these models and the role of AI fluency.

你做过一些研究，是关于人们如何使用这些 models，以及 AI fluency 所扮演的角色。

47:47

Tell us about that research.

跟我们讲讲那项研究吧。

47:49

Oh, yeah.

哦，对。

47:50

First I should say, so the distribution of users across expertise levels.

首先我得说，就是用户在不同专业水平上的分布。

47:56

So I guess the nuance picture I'd offer is that the people deriving a lot of value

所以我想，我给出的更细致的图景是，那些从 AI 中获得很多价值的人

48:00

from AI in the current moment tend to be experts.

在当下往往都是专家。

48:03

It must be the case that most users of AI are beginners just because the numbers are so

情况肯定是，大多数 AI 用户都是初学者，只是因为数字太...

48:08

large and expertise can't be that widely distributed yet.

大规模和专业能力还没法那么广泛地分布。

48:12

And that's a very interesting thing because I think probably most things are getting designed

而这非常有意思，因为我觉得大多数东西可能都是被设计成

48:16

for those experts implicitly or explicitly.

给那些专家用的，不管是有意还是无意。

48:19

But for the whole economic picture to work out, many more people need to derive value from

但要让整个经济图景成立，需要多得多的人从中获得价值，

48:24

this via one avenue or another.

通过这样或那样的途径。

48:27

And so that does shine a light on this expertise thing as a real factor.

所以这确实让专业能力这件事作为一个真正的因素凸显出来。

48:32

And the headline result there actually builds on something that anthropic did.

而那里的核心结果其实建立在 anthropic 做过的某件事之上。

48:35

They have this AI fluency index and their core observation in that work is that experts

他们有一个 AI fluency index，而他们那项工作的核心观察是，专家

48:41

display an augmentative style.

展现出一种 augmentative style。

48:43

They iterate with the AI.

他们会和 AI 反复迭代。

48:46

They push back.

他们会反驳。

48:47

They complain.

他们会抱怨。

48:48

They change their requirements.

他们会改变自己的需求。

48:50

It's a really collaborative mode, whereas novices low fluency users delegate.

这真的是一种协作模式，而新手、低熟练度用户则会采取委派模式。

48:56

So they trust in the AI.

所以他们信任 AI。

48:58

They let it do its thing.

他们让 AI 放手去做。

48:59

They accept the responses uncritically.

他们会不加批判地接受这些回应。

49:02

And our contributors just show that this is a causal factor in success with these products

而我们的贡献者恰恰表明，这是这些产品能否成功的一个 causal factor，

49:08

right now.

就在当下。

49:09

Experts can do harder things more reliably as a result of all that friction they introduce,

专家能更可靠地完成更难的事，正是因为所有这些他们引入的 friction，

49:16

all that pushback.

所有这些 pushback。

49:18

Whereas novice users, they accept, but they end up accepting the wrong thing and they're

而新手用户呢，他们会接受，但最终接受的是错误的东西，而且他们

49:22

not able to level up from the basic tasks that they think to start with.

没法从他们一开始以为要做的那些基础任务 level up 上去。

49:27

And so that's obviously significant.

所以这显然很重要。

49:29

And it feels so tantalizing because pushing back is a natural human behavior.

而且它之所以这么诱人，是因为反驳本来就是人的一种自然行为。

49:35

I feel like we could encourage everyone in the world to do this.

我觉得我们可以鼓励世界上每个人都这么做。

49:37

We probably need to get them out of the mode of thinking, it's a super intelligence.

我们可能得让他们摆脱“它是 super intelligence”这种思维模式。

49:41

You should just trust it.

你就该直接相信它。

49:43

That has been the narrative for a while.

这种叙事已经持续一阵子了。

49:45

What we're seeing in the current moment and possibly for the foreseeable future is that

我们现在看到的，以及可能在可预见的未来会看到的，是

49:49

you got to complain, collaborate, introduce yourself, pushback, all that stuff that I think

你得抱怨、协作、介绍自己、反驳，所有这些我觉得

49:54

we do.

我们都会做的事。

49:55

We take that for granted, right?

我们都把这当成理所当然的，对吧？

49:57

Yeah.

嗯。

49:58

And so from a methodology perspective, how did you approach exploring this?

那从 methodology 的角度来说，你是怎么着手探索这个的？

50:04

We built on the work that Anthropic did, which they set up a nice framework with some

我们基于 Anthropic 做的工作，他们搭了一个很好的框架，和一些

50:10

independent research who were doing this kind of usability stuff.

独立研究者一起，他们当时就在做这类 usability 的东西。

50:14

And we just have an annotation protocol.

而我们只是有一套 annotation protocol。

50:16

We can talk in detail if you want about this, but at BigSpin we have lots of these best

如果你想的话，我们可以详细聊这个，但在 BigSpin，我们有很多这类 best practices

50:20

practices around having language models essentially collaborate on annotation projects to kind

围绕让 language models 基本上在 annotation projects 上协作，来 kind

50:26

of triangulate on the truth and factor out their individual biases.

去 triangulate 真相，并把他们各自的 biases 排除掉。

50:30

So we do that stuff and we apply all these fluency markers.

所以我们就做这些，然后把这些 fluency markers 全都用上。

50:33

And then separately we do a thing of estimating task complexity and looking for signs of visible

然后另外，我们还会做一件事：估算 task complexity，并寻找 visible

50:38

and invisible failures.

和 invisible failures 的迹象。

50:41

And so it's the connection between the fluency markers and the task complexity success metrics

所以，正是 fluency markers 和 task complexity success metrics 之间的联系

50:47

that was our contribution there.

那才是我们在这方面的贡献。

50:50

And that's where you can see high fluency users are the ones doing harder tasks.

从这里就能看出，high fluency users 才是那些做更难任务的人。

50:54

Paradoxically, there's more signs of failure for them because they complain, they push back,

矛盾的是，他们反而有更多 failure 的迹象，因为他们会抱怨、会反驳，

50:58

they're trying harder things.

他们在尝试更困难的事情。

51:00

But as part of all that friction, they're successful with harder things as well.

但作为所有这些摩擦的一部分，他们在更困难的事情上也取得了成功。

51:05

And have you were to try to apply this insight from the perspective of someone in an organization

而如果你要从一个组织里某个人的视角，试着应用这个洞察

51:11

that's trying to help or guide their organization to be more successful with AI?

这个人正试图帮助或引导自己的组织，让它在 AI 上更成功？

51:18

Like what do you think are the key lessons of this fluency work?

比如，你觉得这项 fluency 工作的关键经验是什么？

51:21

If it's an organ that's just starting out and wants people to figure out how this could

如果是一个刚起步的组织，想让人们弄清楚这怎么才能

51:25

be part of the organization's mission, it would just be that pushback message.

成为组织使命的一部分，那就会是那个 pushback message。

51:30

And you could do an experiment where you interact with it about something where you're

而且你可以做一个实验，和它围绕某件你正在……的事情互动

51:33

a world expert.

一个世界级专家。

51:34

We're all an expert in something.

我们每个人都在某个方面是专家。

51:36

Engage in a discourse with one of the best models about something you're an expert in

就你擅长的某个领域，跟最好的模型之一展开一场讨论，

51:39

and see how often you feel you have to push back.

看看你有多经常觉得自己必须反驳它。

51:42

And this could be a kind of a lesson.

这可能是一种教训。

51:44

For other spheres where I don't know the answer, it might be just as errorful.

对于那些我不知道答案的其他领域，它可能也一样容易出错。

51:48

That could be a good visceral thing.

那可能会是一种很直观的切身感受。

51:50

If the org is very far along, I think the main thing to do right now is to have a team

如果这个组织已经走得很远了，我觉得现在最主要的事情就是有一个团队

51:54

of these LMs interacting to improve things.

这些 LMs 互相交互，把事情做得更好。

51:57

For example, at Big Spin, I didn't set this up.

比如，在 Big Spin，这不是我搞起来的。

52:00

Our founding engineer is very future forward on agents and he's incredible at this.

我们的创始工程师对 agents 非常前瞻，而且他在这方面特别厉害。

52:06

And when we do PRs now, the first round of review is the agents all interacting,

而现在我们做 PRs 的时候，第一轮 review 是 agents 全都在互相交互，

52:12

collaborating, disagreeing.

协作、意见不合。

52:14

They do the first round of comments.

他们做第一轮 comments。

52:15

They do the first round of code updates.

他们做第一轮 code updates。

52:17

Only after they resolve things do we look at a PR.

只有等他们把问题解决之后，我们才会看 PR。

52:20

So the final human stage should be very high value and the agents did all that work.

所以最后那个人类环节应该价值非常高，而 agents 把那些活儿都干了。

52:25

But when you have one agent do it, they often just reinforce themselves and you don't get

但如果你只让一个 agent 来做，它往往只是在自我强化，你得不到好的结果。

52:29

good outcomes.

真正有变革性的，就是那种 team of rivals 的机制。

52:30

It's that team of rivals thing that is transformative.

你知道，我觉得这挺有意思，因为一方面，这当然说得通。

52:34

You know, I think it's interesting because on the one hand, of course, that makes sense.

但另一方面，你知道，这里面有个点：它也意味着，在那些你没有足够专业能力去评估答案的领域，你就不该用这些东西；可偏偏那些地方，你才最需要帮助和支持。

52:40

But on the other hand, there's something, you know, it also implies that you shouldn't

我们其实都待在自己那条非常窄的垄里，甚至都没真正意识到。

52:46

be using these things in areas where you don't have enough expertise to evaluate the

谁知道这个花园里，外面还有什么呢。

52:52

answer, yet that's where you most need the assistance, the support.

这玩意儿真的很酷。

53:00

So.

所以。

53:01

And again, and this is a little bit worrisome about the overall narrative around AI.

而且再次，这对围绕AI的整体叙事来说有点令人担忧。

53:06

The place where we can get around this is with software development because let's say

我们能绕开这个问题的地方是软件开发，因为假设

53:11

that I'm trying to accomplish something in a language that I don't know how to code in.

我想用一种我不会编程的语言来完成某件事。

53:15

I can have the agent do work for me because probably in the end, I can run the program

我可以让agent替我干活，因为很可能到最后，我可以运行这个程序

53:20

and look at the results.

然后看结果。

53:22

And that's what mattered to me is that I run the results and I see.

而对我来说重要的是，我运行结果，然后我看到了。

53:25

And if I don't see what I want and I can complain and we can iterate, that verification

如果我没看到我想要的，我可以抱怨，我们可以迭代，那个verification

53:31

step that doesn't imply I have comprehensive knowledge, it just implies that I know what

这一步并不意味着我有全面的知识，它只意味着我知道

53:34

I want to see in the end is so critical.

我最终想看到什么，这一点非常关键。

53:37

And I think this is a causal factor in models being so good at coding because it's like

而且我觉得，这是 models 在 coding 上如此擅长的一个 causal factor，因为它就像

53:41

the ultimate verifiable domain for them.

对它们来说，这就是终极的 verifiable domain。

53:44

But as soon as we leave that and go even into something like the legal realm where the

但一旦我们离开那个领域，甚至进入像法律领域这样的地方，那里的

53:49

requirements are strict, but they're not codified in code and they have ambiguity about

要求很严格，但它们并没有被 codified in code，而且它们还有

53:54

them.

关于它们的 ambiguity。

53:55

This whole picture falls apart and you then are back at what you just said, which is this

这整个图景就崩了，然后你又回到你刚才说的那一点，也就是这个

54:00

awful kind of paradox is like, yeah, use AI, but in the end, unless you're expert enough

那种糟糕的悖论就像是，是啊，用 AI，但到最后，除非你足够专业

54:05

to evaluate every single one of its responses, you might be in real trouble.

能评估它的每一个回应，否则你可能真会陷入大麻烦。

54:10

I don't know how to get out of this because the verification step is like, we go to trial,

我不知道怎么摆脱这一点，因为 verification step 就像是，我们得上法庭，

54:15

but this is very consequential.

但这后果非常严重。

54:17

That's expensive.

那很贵。

54:18

Yeah, that's funny.

是啊，这挺好笑。

54:20

I mean, it does make me think a little bit about, you know, some of the types of errors

我是说，这确实让我稍微想到，你知道，某些类型的错误

54:26

that we're trying to avoid are factuality and, you know, there is a temptation to say,

我们正试图避免的，是 factuality，而且，你知道，会有一种冲动想说，

54:33

well, let's just throw more tokens at it.

好吧，那我们就干脆多给它扔点 tokens 吧。

54:34

Like, I'll have a critic model that evaluates everything that the, you know, is generated

就像，我会有一个 critic model，来评估所有那些，你知道，被生成出来的东西

54:42

by the primary model.

由 primary model 生成的。

54:44

But then you go back to my observation that these models tend to correlate in their responses

但然后你再回到我那个观察，就是这些 models 在它们的回答上往往会 correlate

54:50

as well.

也一样。

54:51

Yeah, it's super interesting.

是啊，这特别有意思。

54:55

That's a good point.

这个观点很好。

54:56

Yeah, for my picture, we want real diversity of perspectives.

是啊，就我设想的图景而言，我们想要的是真正多元的视角。

54:59

This is just like, you know, red teaming for humans.

这就像，你知道的，对人类做 red teaming 一样。

55:02

This is most successful when you have a really diverse team of people who think creatively

当你有一个真正多元化的团队，成员都能有创意地思考时，这种方式最成功

55:06

and differently.

而且他们的想法还各不相同。

55:07

And if every one of the members of that team is thinking in a homogeneous way, they miss

而如果那个团队里的每个成员都用同质化的方式思考，他们就会漏掉

55:11

all of the crucial things.

所有关键的东西。

55:13

Same exact issue of all the code review agents are biased in the same way.

完全相同的问题：所有 code review agents 都以同样的方式有 bias。

55:16

They will miss exactly the same class of bugs.

它们会漏掉完全同一类 bugs。

55:19

And then we're all sunk.

然后我们就全完蛋了。

55:20

Yeah, I don't know how you'd encourage this diversity in the ecosystem or probably, as you

是啊，我不知道你要怎么在 ecosystem 里鼓励这种多样性，或者，可能就像你

55:24

say, converging towards some kind of one model.

说的，最终会收敛到某种单一的 model。

55:27

But I think for my picture, we need diversity.

但我觉得，按我的设想，我们需要多样性。

55:29

Yeah, we got to keep those open weights models going or something because they're the weird

是啊，我们得让那些 open weights models 继续跑下去还是怎样，因为它们确实是这个领域里怪异的

55:33

players in the space for sure, for sure.

参与者，肯定的，肯定的。

55:36

So we've talked about efficiency, interpretability, tokenomics, fluency.

所以我们已经聊了 efficiency、interpretability、tokenomics、fluency。

55:48

You're involved in a lot of different research directions, excellent, excellent, what's next

你参与了很多不同的研究方向，太棒了，太棒了，你接下来

55:52

for you?

要做什么？

55:53

Where do you see either where do you see this all going kind of externally, but also like

你觉得——或者说，你觉得这一切从外部来看会走向哪里，但同时也像是

55:57

where is your research going?

你的研究要往哪个方向走？

55:59

Yeah, this is great.

对，这太棒了。

56:01

And as I said before, we're trying to think in weird and creative ways about what the

而且就像我之前说的，我们正试着用古怪又有创意的方式去思考

56:05

future could hold.

未来可能会是什么样。

56:06

And I encourage my students to do this.

我也鼓励我的学生这么做。

56:08

And they're smart.

而且他们很聪明。

56:09

So they say, all right, Chris, I'll think along those lines, but what's your answer to

所以他们会说，好吧，Chris，我会按这个思路去想，但你对……的答案是什么呢？

56:12

this question?

这个问题？

56:13

So I do have an answer.

所以我还真有个答案。

56:14

And it's really shooting for the moon here, which would be what about the architectural

而且这里真的是在冲着月亮去，那会是什么呢，就是关于 architectural

56:18

innovation that would upend the whole story around the stack transformer and the way we

innovation，它会彻底颠覆围绕 stack transformer 的整个故事，以及我们

56:24

need to do data center build out to even get incremental gains in performance.

为了哪怕在 performance 上拿到一点增量提升，需要怎么做 data center build out。

56:28

That could be upended and it would come from some very innovative thing around maybe

那可能会被颠覆，而且它会来自某个非常创新的东西，也许是关于

56:33

recursive use of the building blocks that we've got.

对我们手里已有的 building blocks 做 recursive use。

56:37

So architectures, we should think.

所以 architectures，我们应该想一想。

56:38

And when people say, oh no, we don't need more architectures.

而当人们说，哦不，我们不需要更多 architectures 时，

56:40

The transformer is good enough.

transformer 已经够好了。

56:42

That's where we should push back as academics doing something more clever and more scrappy

这就是我们作为学者应该顶回去的地方——去做一些更聪明、更敢闯的事情，

56:47

that could change the world.

那可能会改变世界。

56:49

And the other one is thinking in the interps space, much more about data.

另一个方向是在 interps 领域里思考，更多地去关注 data。

56:54

And that's just because I want to tell the true story of how we go from data to model

而这只是因为我想讲一个真实的故事：我们怎么从 data 到 model

56:58

capabilities, but it also checks a box for me on connecting interpretability to safety.

capabilities，但它也帮我在把 interpretability 和 safety 连起来这件事上，算是完成了一个目标。

57:06

It has been hard for me to connect those two things.

对我来说，把这两件事连起来一直很难。

57:08

We have found some ways to do it, but it's not a slam dunk as a narrative, even though

我们已经找到了一些实现方法，但作为一套叙事，它并不是十拿九稳，尽管

57:12

it's a dominant narrative.

它是主流叙事。

57:13

But I will say that when we get into things like data poisoning from innocuous examples,

但我要说，当我们谈到用看似无害的样本进行 data poisoning 这类事情时，

57:19

this is probably a growing societal concern.

这很可能是一个日益严重的社会担忧。

57:22

There is evidence that with very few examples planted in a pre-training data set, you can

有证据表明，只要在 pre-training data set 里植入极少数样本，你就能

57:27

have a significant influence on the outlook and preferences and quirks of the final model.

对 final model 的观点、偏好和一些怪癖产生显著影响。

57:34

So can we detect those examples?

那么，我们能检测出那些样本吗？

57:36

What's the nature of those attacks?

那些 attacks 的本质是什么？

57:38

How well hidden could they be?

它们能藏得有多隐蔽？

57:39

What's the smallest number of examples?

最少需要多少个 examples？

57:41

And why does it happen?

那为什么会发生？

57:43

These are all going to be very pressing questions.

这些都会是非常紧迫的问题。

57:46

And so again, it's just a data-oriented question that's very alive for me in the current moment.

所以还是那句话，这只是一个 data-oriented 的问题，而此刻它对我来说非常鲜活。

57:51

On the architecture front, is there research that you're seeing or doing that is as yet

在 architecture 方面，有没有你看到或正在做的研究，目前还

58:02

under the radar that you think is promising and are underappreciated?

没被太多人注意到，但你觉得很有前景、也被低估了？

58:08

I think you had my student, Julie Colledion, and she is an advocate for bite-level models,

我记得你邀请过我的学生 Julie Colledion，她是 bite-level models 的倡导者，

58:13

essentially tokenizer free models.

本质上就是 tokenizer free 的模型。

58:15

I think that's a big part of the future.

我觉得这会是未来的很大一部分。

58:17

It's a critical thing if you want to have truly multilingual models that are also equitable

如果你想拥有真正 multilingual 的模型，而且这些模型还要公平，

58:21

in terms of how many tokens they charge us for, getting back to that earlier theme.

也就是从它们按多少 tokens 向我们收费这个角度来说，回到之前那个主题。

58:25

But also, Julie's perspective is that this is speculative, but I think there's something

但另外，Julie 的观点是，这还是推测性的，不过我觉得这里面有点

58:29

to this, that it's a kind of inference-time scaling, because you do more compute at test time,

道理，它是一种 inference-time scaling，因为你在 test time 会做更多 compute，

58:36

because you have more tokens, and there are four more opportunities to build on interesting

因为你有更多 tokens，而且还有另外四个机会，可以在有趣的

58:41

things.

东西上继续发展。

58:42

So that could be a big part of the future, and the other one would be recursive architectures,

所以那可能会是未来很重要的一部分，另一个就是 recursive architectures，

58:47

as I said.

就像我说的。

58:48

But if you want to go all the way out, you could think, why do we always assume we're going

但如果你想彻底往远处想，你可以想想，为什么我们总是假设我们会做 gradient-based learning？

58:52

to do gradient-based learning?

那有很多替代方案，但没人去探索，因为所有人都把它当成不言自明的事。

58:53

There are lots of alternatives to that, and nobody is exploring them, because everyone

我们都待在自己非常窄的那一垄里，甚至都没真正意识到这一点。

58:58

takes it as a truism.

谁知道这个花园里，外面还有什么呢。

58:59

We're all in our very narrow row here without even really realizing it.

我们其实都待在自己那一窄垄里，甚至都没真正意识到。

59:03

Who knows what's outside in this garden.

谁知道这个花园外面有什么。

59:06

It's very risky as a research bet, because only one in a thousand of these ideas will

从研究下注的角度看，这风险非常大，因为这类想法里只有千分之一会真正有回报。

59:11

pay off.

但如果你不打算冒这种风险，那做学术研究者又有什么意义呢？

59:13

But what's the point of being an academic researcher if you're not going to take that kind of risk?

这正是我们的定位所在。

59:18

That's what we're positioned to do.

嗯，Chris，非常感谢你来参加，并分享了一点你正在做的事情。

59:20

Well, Chris, thanks so much for jumping on and sharing a bit about what you're working

这真的很酷。

59:24

on.

谢谢。

59:25

It's very cool stuff.

这真是很酷的东西。

59:26

Thank you.

谢谢。

59:27

What a wonderful conversation.

真是一次精彩的对话。

59:28

It gave me lots of new things to think about.

它让我有很多新的东西可以思考。

59:30

Awesome.

太棒了。

59:31

Yeah.

是啊。

Do AI Tokenomics Matter More Than Model Benchmarks? with Chris Potts - #776

Play Queue