TITV One Year Anniversary Show

所以我相信他们能做到，你看搜索广告市场有两千五百亿美元，社交广告去年底达到两千六百三十亿，AI广告我觉得肯定会跟上。

So I'm I'm confident that they will, and I think you look at search ads, $250,000,000,000 market, social ads, two sixty three at the end of last year, AI ads will be, I think Right.

会大得多。

Much bigger.

Jeanne，我知道你离开Stripe已经有一段时间了，现在又加入了Shopify的董事会，这真是个令人兴奋的新机会。

Jeanne, I know it's been a while since you were at Stripe, and you just joined the board of Shopify, which is an exciting new opportunity.

总的来说，如果你要预测一下明年AI对电商会带来什么最大变化，你觉得会是什么？

I mean, broadly speaking, just if you were to make a prediction about AI and commerce for the next year, what do you think is going to change the most?

我们最近观察到一个很有趣的现象，其实是让长尾商品得到了更多曝光。

One of the really interesting dynamics we've been seeing is actually that it's exposing more of the long tail, right?

你可以在AI里做非常精准的搜索，一下子就能找到那些卖你想要的那种日本风纽扣衬衫的小店。

So you can sort of do like a very specific search within AI, and all of a sudden, it's getting you access to, you know, the sort of torso and tail of shops that are selling your specific Japanese button down, you know, shirt that you wanted.

从电商角度看，这确实是个好事，让那些没有上亿美元广告预算的小商家也能被看到。

So near run, that's been positive from a commerce perspective, is just getting more exposure to the folks that don't have the $100,000,000 you know, advertising budget to show up.

我觉得有意思的是，未来人们会不会干脆把购物决策完全交给AI，让它直接下单——很多人其实低估了，其实很多人买东西，是因为享受这个过程。

I think like what'll be interesting is where people want to hand over the reins on commerce and actually let AI go complete the purchase versus I think people sometimes underestimate the degree to which people shop because they enjoy

这让我想到，Brian Chesky 曾经说过，他不确定聊天机器人是不是购物和电商的合适方式，你知道的。

it and It want to go look at reminds me of I mean, look, Brian Chesky has said that he's not sure if a chatbot is the right medium for shopping and commerce and, you know

举个例子，我这周末住了一家Airbnb，如果用AI来搜索，我找房子的方式会跟现在用的筛选功能完全不同。

There are a lot of like, if you take I just was at an Airbnb this weekend, and I would have searched for what I wanted fundamentally differently with AI than with the filtering method that you have now.

这样我能更快地找到那十套我真正想看的房源。

So I think there could get me much more rapidly to the 10 house homes I actually want to look at.

我觉得AI在发现新东西方面肯定会特别厉害。

So I think it'll be like incredible for discovery for sure.

然后就是，人们到底愿意让AI直接完成交易，还是更想自己慢慢挑这十套房源，或者那五件裙子之类的。

And then, you know, degree to which people again go off and say, okay, now complete that transaction versus I actually do want to peruse those 10 different homes or, you know, those five dresses, etcetera.

非常感谢你们两位来参加我们的节目。

Well, I want to thank you both for joining us.

今天的对话非常精彩。

It was a great conversation.

今天我们请到的是来自Vercel的Jeanne Grosser，以及Theory Ventures的Tomasz Tunguz，欢迎来到TITV。

That is Jeanne Grosser from Vercel and Tomasz Tunguz from Theory Ventures here on TITV.

不管是GPT、Sonnet、DeepSeek还是Fable，AI模型的优势都是当前大家普遍关注的核心话题，不管 Adoption 进展到哪一步。

Whether it's GPT or Sonnet or DeepSeek or Fable, the strength of AI models is one of the big topics everyone is tracking irrespective of where adoption is at.

当OpenAI和Anthropic这类闭源模型背后的公司激烈竞争时，它们也都正面临越来越多对开源模型的关注。

And while the companies behind closed source models like OpenAI and Anthropic are battling it out, they are both brushing up against a lot of recent interest in open source models.

为了理清这场竞赛的现状，我想请来Decagon的CEO杰西·张。

To unpack the current state of that race, I wanna bring on Jesse Zhang, the CEO of Decagon.

杰西，欢迎来到节目。

Jesse, welcome to the show.

很高兴你能来。

It's great to have you here.

谢谢你的邀请。

Thanks for having me.

你最近在X上写了好多有意思的内容，发了不少好帖子，其中有一篇最近的特别吸引我。

So you've been doing a lot of interesting writing on X, a lot of great posts, and one of your most recent posts really intrigued me.

那篇讲的是开源模型，你提到：‘大家对开源AI在企业中的应用都理解错了。’

It was about open source models, and you wrote that everyone is wrong about open source AI in the enterprise.

他们错在哪儿？

What are they wrong about?

是的，简单来说，现在有个很强的说法，认为开源已经全面胜出了。

Yeah, I would say to put it simply, there's a really strong narrative now that open source has taken over.

它的采用率正在大幅上升。

It's getting a lot more adoption.

从各种数据都能看出来，比如Hugging Face上的模型使用量、推理服务商的支出，全都飙升了。

You see this in ton of metrics, the usage of the models on Hugging Face, the spend on the inference providers, they're all ripping.

这一点确实没错。

So that is true.

开源确实在快速发展，但如果你仔细看研究数据，会发现开源模型相比闭源模型的占比其实正在下降。

Open source is happening more, but then actually, if you look at the research, a lot of the stats are saying that the share of open source compared to closed source is actually shrinking.

所以我们一直在想，为什么会这样？毕竟我们自己是开源的坚定支持者，目前我们90%以上都是开源的。

And so we were just thinking why that is, because we are a big proponent of open source, so we are pretty much 90 plus percent open source at this point.

真没想到。

Wow.

你们一直都是90%开源的吗？不是，不是的，

Have you always been 90 No, no,

我们刚开始的时候，全是闭源的，因为你得先搞清楚使用场景，对吧？

When we started, it was all closed source, because you're just trying to figure out the use case, right?

对的。

Yeah.

但我的观点是，为什么开源在增长，它的占比却在下降呢？

But my thesis was, Hey, why is it the case that open source is growing, but then the share is shrinking?

简单来说，开源特别适合那些结构清晰、已经明确的规模化生产场景，你知道你的代理或模型调用该怎么设计。

That, simply put, open source is great for scale production use cases where you know the structure of the use case and you know the structure of your agent or the model calls.

其他所有情况，闭源反而更合适。

Everything else, closed source makes more sense.

所以你看企业用户，大多数应用场景还处在实验阶段，或者还没完全落地、没规模化，这就是为什么闭源方案依然这么流行。

And so if you look at the enterprise, most of the use cases are still in that experimentation phase or still in that sort of not fully productionized, scaled out phase, and that's why Close Force is still so popular.

所以我从你的帖子里理解到的是，当你为AI设定一个任务时，你希望用最先进的模型。

So the point that I sort of read from your post was this idea that when you come up with a task for AI to do, you want the cutting edge model.

你需要前沿模型，因为你不确定完成这个任务到底需要多大的算力。

You want the frontier model because you don't know how much power essentially you're gonna need to complete the task.

我理解你的意思是，我们之前遇到过很多这样的场景。

The way I read what you were saying is that, you know, we had a a lot of these use cases.

人们会想，嘿，这个我好像也能做。

People thought about, hey, I think I could do this.

那为什么不直接用前沿模型呢？

Why not use a frontier model?

但现在大家开始觉得，嗯，也许没必要用那么强的模型了。

Now we're at the point though where they're saying, okay, well, maybe I don't need that.

用开源模型就够了。

I can use the open source models.

不过我想问你的是，这项技术的应用前景难道不会继续扩展吗？

The question I have for you though is won't the applications of the technology continue to keep going?

我们总会发现新的用途，所以前沿模型的使用场景理论上还是会持续下去，对吧？

We're going to keep finding new use cases, so frontier model use conceivably will still continue, right?

是的，这正是我想表达的核心观点。

Yeah, that's my main point.

所以

So

并不是说开源模型现在更聪明了，就能完全取代前沿模型。

it's not that, Hey, open source is going to take over and because they're smarter now, you don't need the frontier models.

我觉得总会不断出现新的应用场景。

I would say there's always going to be new use cases.

另外我提到过，要让开源模型达到前沿模型的水平，其实需要投入大量工作。

And the other thing I wrote about is that the effort to get open source to perform at the same level as Frontier, it actually does take effort.

这并不是说这些努力不值得。

And that's not to say that effort's not worth it.

我们确实投入了很多资源，而且完全值得，比如在做AI客服时，一旦这个场景稳定下来并且在生产环境中规模化了，用开源模型就特别合适，因为投入训练这些模型是划算的——尤其是那些小型模型，能带来很大的收益。

I mean, again, we put a lot of effort in and it's super worth it because once We do AI customer service, and so once that use case is solidified and you're already scaled in production, then open source makes a lot of sense because it makes sense to put in that effort to train those models because those open source models, especially the small ones, you get a lot of benefit out of them.

而且延迟更低，这正是我们最看重的一点。

You get better latency, which was the main reason for us.

成本也更低。

You get better cost as well.

这些模型开箱即用的话，根本达不到要求。

Out of the box, those models are just not good enough.

你得花大量功夫，才能让它们够用，能在某些场景下替代前沿模型。

So you have to put in a lot of work to make them good enough and make them be able to sub in for these frontier models in certain spots.

一旦投入了这些工作，结果就必须能证明这些投入是值得的。

And once you put in that work, now the results have to justify the work.

只有当你在生产环境中大规模部署时，这些投入才真正说得通。

And they only justify the work if you're deployed in production at scale.

正如你所说，大多数应用场景都不是这样的。

Most of the use cases, to your point, are not like that.

所以人们还是会继续用前沿模型来应对这些情况。

So people are just gonna continue using frontier models for that.

你们在开源模型方面具体在用哪一个？

Which model specifically are you guys using in the open source category?

是多种多样的。

It's a mix.

我们主要用的是Genma 4，这是美国的主要模型。

So we have Genma four is the main American one.

中国的很多模型也很不错，比如Quinn和GLM。

A lot of the Chinese ones are quite good, Quinn, GLM.

我们有自己的评估体系，来判断哪些模型在特定场景下表现更好。

And we have our own evals to decide which ones perform better in certain scenarios.

对。

Right.

你所说的这场开源模型竞赛，是指中国和北美地区开发的开源模型之间的竞争，对吧？

How do you mean, the race here, the open source race here, I guess, between open source models coming out of China and those developed here in North America.

以前大家对这场竞赛非常关注。

At one point, this was a race that people were paying a lot of attention to.

现在大家普遍认为，这些模型确实很出色。

Now it's accepted, I think, that you know, those models are great.

人们愿意使用它们。

People are willing to use them.

我的意思是，如果你要预测一下这场竞赛会怎么发展，你觉得现在这个问题是不是已经过去了，我们可以随便用任何模型了？

I mean, if you were to sort of make a prediction as to how you see that race playing out, do you think that the issue of it is kind of behind us now and now we can just sort of use whatever we want?

我觉得这完全取决于个人的选择。

I think it's really up to people's choices.

我们接触过一些企业，他们的管理层很坦率地表示：我们只用美国训练的模型，因为即使是中国的开源模型，也都是托管在美国的。

We've talked to enterprises where, super fairly, their leadership has made a decision, Hey, we only wanna use models that are trained in The US, because even the Chinese open source models, they're all hosted in The US.

你的数据并不会外流，所以这并不是问题，但问题是，这些模型里可能带有某种偏见。

Your data's not going anywhere, so that's not a concern, but it's, Hey, maybe there's some bias in those models.

我们不希望这些模型进入我们的系统。

We don't want those models in our system.

这取决于每一家企业自己的决定。

And so that's up to each enterprise.

还有一些企业则对使用开源模型非常放心。

Other ones are very comfortable using open source models.

我认为中国采取了非常支持开源的路线，因为这正是他们竞争的方式。

I think China has taken a very open source forward approach because I think that's how they compete.

在美国，很多顶尖实验室如果站在他们的立场上，把最好的模型保留为闭源是合情合理的，毕竟这样能赚更多钱。

In The US, a lot of the leading labs, I think just if you're in their shoes, it makes sense to put your best models not open source because you make more money that way.

所以现在也听到中国方面说，也许他们已经达到了某些领域的前沿水平，接下来可能想停止开源了。

So there is kind of that balance, you also hear from China now that, okay, maybe they want to kind of now that they've reached frontier level in certain cases, maybe they want to stop doing open source.

谁知道呢？

Who knows?

我们拭目以待，但我认为提供更多选择对大家都更好。

We'll see, but I think having a lot of choices is just better for everyone.

你觉得Meta的决定怎么样？

What do you think of Meta's decision?

Meta最初推出的Llama系列模型都是开源的。

Meta, initially, the LAMA family of models, and that was open source.

Muse的话，除非我搞错了，应该是个闭源模型。

Muse, unless I'm missing something, that is a closed source model.

所以目前来看，这是在向非开源方向转变，这是一种选择。

So this is a transition away from open source for the moment, which is a choice.

你觉得是什么因素促成了这个决定？

What do you think make up that decision?

我觉得这很合理。

Again, I think it makes sense.

如果你站在美国公司的立场上，这其实是常态。

If you're in their shoes in The US, that's the norm.

所有新模型都是闭源的，一旦闭源，就能保留更多优势，还能靠用户付费来赚更多钱。

All of the new models are closed source, and if they're closed source, you maintain more of your alpha, and you get to make more money because people are paying for the model.

这形成了一个自我强化的循环。

That's like a self reinforcing loop.

用这些收入来覆盖模型训练的成本，自然就说得通了。

You take that revenue, that helps cover the cost of training the model, and so that makes sense.

但如果中国也这么做，美国那边根本没人会用这些模型，大家担心数据会流向中国。

And again, in China, if they did that, no one in The US would use those models because people are afraid of their data going to China.

所以在那儿采取开源策略是有道理的，因为这能让人使用你的产品

So having open source strategy there makes sense because it gets people to use your

给咱们说说你们公司最近的情况。

Give us an update on your company.

我们差不多一年前就请过你上节目了，实际上，哦，

We had you on the show almost a year ago now, actually, Oh,

还没那么久呢。

it hasn't been that long.

已经那么久了。

It's been that long.

你之前说，我们录节目前，你们公司现在有五百人了？

You were telling me before we started taping, you're at 500 people now as the company?

是的。

Yeah.

公司发展得可真快。

Companies run very quickly.

我们总部就在旧金山，不远，不过现在我们——

We're based here in SF, not too far but now we're-

你们的产品线扩展了不少吧？

Your product suite has expanded quite a bit, right?

是的。

Yes.

我们第一次见面的时候，大概是一年前吧，那时候公司规模只有现在的十分之一，而且我们公司才成立两年多一点。

So when we first met, I think, I mean, a year ago, were maybe a tenth the size, and we only started the company two, two and a years ago.

是指收入吗？

In terms of revenue or?

不是，是员工人数。

No, people.

是员工人数。

People.

明白了。

Okay.

大概五十人左右。

Like 50 people maybe.

一开始我们做的是AI客服助手。

Early on, was AI customer service agent.

这么说最能理解我们是做什么的。

That's the easiest way to think about us.

我觉得这依然是我们的核心，不过我们稍微拓宽了方向——Decagon的目标是帮助任何企业最大化客户体验。

I would say very much that's still our core focus, but we've expanded the aperture a little bit to be, Hey, our goal as Decagon is to help any business maximize their customer experience.

让客户能轻松地与企业互动，随时都能有个对话助手可以交流。

So make it really easy for customers to interact with them, have a conversational agent that they can talk to at any point.

这可以用于客户服务。

And it could be for customer service.

也可以是主动发起的。

It could be proactive as well.

比如主动联系客户，提升参与度，或者防止他们流失之类的。

You could proactively ping someone to get them engaged or make sure they don't churn or whatever.

所以我们把这个概念称为管家式服务。

And so we call that concept like a concierge.

这相当于为企业提供了一种全新的用户体验。

It's like a new UX for the business.

没错。

Right.

现在公司的营收规模有多大？

How big is the company now in terms of revenue?

我们不公开谈论营收，但增长非常快。

We don't talk about revenue publicly, but we've been growing very quickly.

去年我们的营收大概是原来的五倍。

Last year, we were roughly 5X.

今年的目标也是差不多的增幅。

This year, probably want to do roughly the same.

真厉害。

Wow.

2025年增长五倍，没错。

5X in 2025 and Yes.

有一类我们在这里经常提到的产品，就是客户支持工单系统，用户能直接编写解决方案。

Top There's a category of products that we've been written around here, customer support tickets where they can actually code up solutions.

我们在通讯简报里也写过这个功能。

That's something that we've written about in our newsletters.

你们有没有关注过这个方向？有没有尝试过？

That seemed like some Have you guys expanded into that, looked into that at all?

是的。

Yeah.

核心想法是，有个智能代理在和客户对话，这就是代理的主要工作。

So the concept there is, hey, you have an agent that is talking with the customer, so that's what that agent does.

不管是语音代理还是其他形式，对话的最终结果可能是——我们需要修复某个问题。

Maybe it's a voice agent or whatever, and maybe the outcome of that conversation is, Hey, we actually need to fix something.

这时候，可以再让一个更慢但更聪明的代理在后台配合，帮助优化或解决问题。

So then you can have a slower, maybe smarter agent that's sitting behind it that it works with to get better or to fix things.

这其实是我们今年最重要的发布之一，我们叫它Duet，因为就像两个智能体在协同工作。

And so that was actually one of our biggest launches this year, and so we call it Duet because it's like two agents working together.

你这边是个快速响应的智能体。

So you're a fast agent.

OpenAI和Anthropic最近推出的新模型，比如Opus 4A、Fable这些大型思考型模型，其实不太适合用在语音智能体上，因为它们反应太慢了。

A lot of the new releases from OpenAI Anthropic, like Opus 4A, Fable, these big thinking models are not actually be that useful in your voice agent because they're way too slow.

它们可能要十秒钟才回复，但正适合做第二个智能体。

They could take ten seconds to reply, but they're actually really good for the second agent.

第二个智能体的任务是观察第一个智能体的行动，分析这些对话，发现问题——哦，原来我们在某个环节做得不够好。

And so the second agent's goal is to look at what the first agent is doing, look at these conversations, your point, identify that, oh, yeah, actually, there is this area that we're not doing super well.

这是改进的方法。

Here's how we'd fix it.

我们会更新你的文档，或者调整我们写的工具之类的。

We will update your documentation or we'll change this tool that we wrote, etcetera.

它的角色，其实就是为未来打造一个类似Cloud Code的后台辅助系统。

And its job is to just kind of almost be like a Cloud Code type thing for the future.

然后实际去完成任务，同时把问题修好。

And actually do the task, but fix the problem.

这一点在Duet里你们确实实现了。

That is something with Duet that you have implemented.

对。

Yeah.

那大家平时都用Duet来做什么呢？

So what do people use Duet for?

他们可能会过来说：我想搭建一些新的工作流程。

They could come in and just be like, Hey, I want to build these new workflows.

我们这方面的文档还不太完善，但这里有上万条之前的对话记录。

We don't really have good documentation for them, but here, here's 10,000 previous conversations.

你去帮我们把这些流程写出来。

You go and write these workflows for us.

他们会进入Decagon系统，编写这些文档，我们叫它‘智能体操作手册’，用来指导主智能体该怎么做。

And so they'll go inside Decagon, they'll write these documents, we call them agent operating procedures, to program the main agent what to do.

等你上线运行后，它会自动为你生成测试用例。

And then once you're out in production, it'll actually create tests for you.

它会读取所有对话记录，每一条新来的 transcripts 都不放过，这种事人类根本做不到，数量太多了。

It will read those conversations, every transcript as they come in, like no human will ever be able to do that, there's way too many.

它还会告诉你：这个部分运行得不错，但那个地方出问题了。

And it'll tell you, Okay, this is going well, but this thing is broken.

只要有人提到这个话题，我们就明显卡住了。

Time someone talks about this topic, we're struggling.

所以我仔细翻看了所有对话，终于发现：原来问题出在这儿，这才是提升的关键。

So I've gone in and read all those conversations and I've come up with, Oh, actually, how you get better.

现在，我不但找到了改进的方法，还亲自去做了调整。

And so now, not only have I identified how to get better, I've gone in and made the changes.

所以，该动手的时候就动手干。

So like do the work.

那你又是怎么决定要进行这些评估的呢？

How do you then determine the I'm going to evals here, really.

现在很多人都在关注智能体的评估问题。

Evals with agents is a topic that I think people are really focused on these days.

比如在客服场景里，其实标准挺明确的。

I mean, with customer service, I guess it's sort of binary.

你要么解决了客户的问题，要么没解决。

You either answered the inquiry or you didn't.

但像你刚才说的这种创建工作流的任务，效果到底好不好，就成了一个问题。

In a task like you just described, creating workflows, the effectiveness of those workflows becomes a question.

随着技术的发展，现在对智能体的评估是变难了，还是变简单了？

Are the evals with agents, are they getting harder or are they getting easier right now with the way the technology is going?

我觉得这得看具体领域。

I think it depends on the space.

在我们这个领域，评估标准其实比较清晰，因为一旦有人进来，我们心里就有数了——什么样的对话才算成功，就能据此设计评估标准。

I think in our space, evals are fairly clear because if someone comes in, have an idea of what is a successful conversation with, so you can set up evals against those.

所以我们产品里有个叫‘模拟’的功能，可以模拟各种类型的对话，在智能体上线前，就能先测试它表现如何。

And so we have a concept called simulations in our product where you can just simulate all sorts of different types of conversations, and even before you push an agent live, you can eval if it's working.

同样的，一旦上线了，你就可以对这些对话进行评估。

Same thing, once it's in production, you can evaluate those conversations.

所以这个还算简单。

So that's fairly simple.

有没有哪些方面变得越来越难了？

There any areas where it's getting more difficult?

现在评估这个话题确实很热门，可能在你们业务里不明显，但从我接触的人来看，大家现在都在为评估问题头疼。

I mean, eval is It's a topic right now that And it might not be in your business, but I guess from the people you talk to, I mean, evals, this is an area that people are really grappling with right now.

我觉得大家真正纠结的是如何评估底层模型本身。

I think the area that people are grappling with is more evalling the base models themselves.

比如，Fable 和 Muse Spark 哪个更好？这种问题很难说清楚。

It's like, Hey, is Fable better or is Muse Spark better or whatever?

真的很难下定论。

It's really hard to say.

所以大家才需要评估体系和基准测试，然后有人会说：其实这个基准并不靠谱，因为 XYZ 原因。

So that's why people have Eval's, they have benchmarks, and then people can come out and say, Actually, no, this benchmark's not good because of XYZ reason.

最难衡量的就是模型到底好不好，但在我们的场景里，目标非常明确。

That's the tough where it's really tough to measure if a model is good or not, but in our use case, it's so targeted.

就是一个代理，解决一个公司的一个具体问题，这家公司自己很清楚什么叫成功。

It's like one agent for one use case for one company, and so that company knows, Here's what success looks like.

所以在我们的场景里，评估起来简单多了，你懂吧。

So the evals are a lot simpler in our use case, if that makes sense.

如果现在平台上90%的工作都是基于开源模型做的，那你们用Fable和5.6具体是做什么任务？

So if 90% of the work that's happening on your platform right now is coming from open source models, What are the tasks then that you are using Fable for, 5.6 for?

你们到底拿这些模型来干什么？

What are you using those models for specifically?

Duet就是个很好的例子，对吧？

So Duet is a good example, right?

Duet是依赖前沿模型的。

So Duet relies on the frontier models.

那既然这样，我们当初为什么要大力推动用开源模型呢？

If rewind, why did we make a push to use open source in the first place?

因为在我们的语音助手或聊天助手里，延迟真的很重要。

It's because in our voice agent or our chat agent, latency does matter.

如果客户来找你，跟助手对话，助手能迅速响应、快速解决你的问题，体验就会好很多。

If you're a customer and you're coming in, you're talking to the agent, if it can be super snappy and just get to my problem faster, that's just a much better experience.

你可不想听到‘请稍等’，然后干等十秒钟。

You don't wanna be like, Okay, give me a second, and then it's just waiting for ten seconds.

所以我们转向了开源模型，因为它们体积更小，运行更快。

So that's why we moved to open source, because those are smaller models, they're faster.

这些模型刚出来时效果未必好，得经过微调和训练，才能完成我们要求的任务，但这么做是可行的——在语音助手的流程中，很多步骤其实不需要复杂模型，比如‘选择话题’这一步，根本用不着大模型。

Those models are not gonna be that good out of the box, and so you have to fine tune them, you have to train them so that they can perform the same tasks as you would need to, but it works because in this voice agent, most of the steps along the way, one of the steps could be choose the topic, that model doesn't need to be a big model.

做这个任务，不需要懂太多东西。

You don't need to know a lot to do that task.

你只需要特别擅长判断该选哪个话题就行。

All you have to do is be really good at choosing the topic.

没错。

Right.

所以开源在这里是行得通的。

So that's why open source works there.

但对于像 Duet 这种更通用的场景，你只是跟它说：‘帮我看看我所有的对话，做点事吧。’

But for things like Duet where it's much more general purpose, you're just asking it, Hey, go read all my conversations and do some work for me.

在这种情况下，用智能就没太大意义了。

Using the intelligence doesn't make sense there.

这就是其中的权衡。

So that's the trade off.

我想问你几个关于这里技术细节的问题。

I want to ask you about a couple of the more technical elements here.

持续学习、自我迭代改进，你们可是深陷这个领域啊。

So continual learning, recursive self improvement, I mean, you're deep in the space.

我们上周刚开完 ICML，所有研究者都在讨论这些话题。

We just had ICML last week, and all the researchers were talking all about these topics.

你对持续学习这个问题怎么看？实现它的机会有多大？到底在推动 AI 普及上，它有多关键？

I mean, what's your view on the continual learning problem here, the opportunity to get to there, the extent to which you actually need it to drive AI adoption?

你觉得呢？

What do you think?

我觉得这显然非常强大。

I think it's obviously very powerful.

毕竟，有大量证据表明它确实有效。

I mean, there's so much evidence that it works.

而且，如果你看看很多实验室的研究，就会发现他们其实面临数据不足的问题。

Also, if you look at the work that a lot of labs are doing, there is a bit of a data limitation problem.

我们已经用尽了所有能拿到的数据来预训练这些模型，那接下来该往哪儿走？

We've taken all the data that we can to pre train these models, well, where do you go from there?

持续学习就很棒，或者干脆直接根据模型的实际表现来学习。

Well, the continual learning is great, or just being able to kind of learn from actually the model's performance.

这样一来，你就打开了通往全新类型数据的大门。

Then you kind of open up your aperture to like a whole world of new type of data.

你甚至可以说，人类的学习方式也是这样的，对吧？

And you could argue that's how humans learn as well, right?

就像小时候，你总会有人告诉你该怎么开始，但之后这一生，你主要是靠自己的经历和错误来学习。

Like when you're little, you kind of get instruction of like, Okay, here's how you get started and whatever, but then from there, throughout life, you're kind of learning through your own experience and your own mistakes.

所以你可以说，这本来就是很自然的发展方式。

So you could argue that that's sort of naturally what should happen.

这是从模型层面来说的。

That's at the model level.

对我们来说，我们可能处在应用栈更高的一个层级，目标是整体提升智能代理的表现。

For us, we were maybe like one level higher in the app stack, where we are trying to improve the entire agent as a whole.

所以这种学习的重点，不在于改进模型本身。

And so the type of learning there is less about improving the models.

而在于，通过我们的智能代理为某个企业处理数百万次客户服务互动，从中发现哪些地方可以改进，怎样才能变得更好。

It's more about how can you, through our agent having millions of customer service interactions for a certain business, through those interactions, figure out what can be improved and how to get better.

对。

Right.

你觉得AI研究人员的工作，有多大程度能被自动化？

What's your view on the extent to which the role of AI researchers can be automated?

我觉得离那还远着呢。

I think we're kind of far from that.

特别远。

Far

远远达不到。

from it.

这才是让人焦虑的地方。

This is the anxiety.

这正是我担心的，我只是在跟你说说AI研究到底是怎么回事。

That's the kind of concern I have, it's just I'm telling you what the AI research is.

总有更多事情可以做。

I mean, there's always more to do.

我们现在有个相当庞大的研究团队，成本很高，但人都是顶尖的。原因很简单：如果我们不在这上面投入，不投资训练开源模型，我们的产品就不可能达到顶尖水平——延迟问题、各种性能瓶颈，只要我们能把数据收集好、训练到位，这些问题都能解决。

We have a pretty sizable research team at this point, and it's a very expensive team, but we have some great people, the reason is that, hey, we decided, unless we invest in these things, unless we invest in training open source models, our product is not gonna perform state of the art because we're gonna have latency issues, we're gonna have all these things that could be done better if we actually compile the data and train it.

所以我觉得，研究人员能发挥的空间还多得很。

So I think there's just gonna be so many use cases for researchers to help with.

我想问你个问题。

I wanna ask you a question.

Salesforce 收购了 Fin。

So Salesforce bought Fin.

你们收到过收购要约吗？

Have you guys gotten any acquisition offers?

我们也不谈这个，但目前真不是我们最关心的事。

We don't talk about those either, but it's not really top of mind for us.

我们才成立两年多。

We're only two and a years

老了。

old.

那融资呢？

What about fundraising?

有的。

Yeah.

融资方面，我们一直挺积极的。

Fundraising, we've been very active.

到目前为止，我们已经完成了五轮融资。

We've done five rounds so far.

第六轮是不是正在

Is there a sixth happening

进行中？

right now?

目前还没有，不过很快就会有。

Right now, no, but at some point soon.

我们刚完成上一轮融资，时间也不长。

We just did our last round not too long ago.

你们是怎么看待公司增长的？

How do you think about growth of the company?

你提到过要把产品系列扩展到客服代表之外。

You talked about expansion of the product suite beyond just the customer service agents.

三年后，Decagon的扩张会是什么样子？

Three years from now, where does Decagon What does expansion look like?

是进入新市场吗？

Is it new markets?

是推出横向产品吗？

Is it horizontal products?

你们最终的愿景是什么？

What's the ultimate ambition for you?

我们的目标，正如我之前说的，是打造一个或一组智能代理，彻底改变企业与客户互动的方式。

So ambition for us, as I mentioned, big vision here is can we produce an agent or a set of agents that can just improve how every business interacts with their customer.

我们关注的核心，就是这个互动层面。

So that interaction layer is what we care about.

你可以这样想，过去人们只有网站，后来手机出现了，大家都开始用手机。

And you can almost think of it as, Hey, in the past, people had websites and then phones came out and everyone was using them.

于是，移动应用就成了新的交互界面。

So then a new UI was created, which is the mobile app.

现在，紧随其后又出现了一种新的交互方式，那就是对话式智能代理。

Now there's almost a new UI after that, which is a conversational agent.

这是一种与企业互动的新途径。

It's a new way to interface with the business.

这并不意味着手机应用和网站会消失，但它确实提供了一种全新的互动方式，而这正是我们努力的方向。

That doesn't mean mobile apps and websites are going away, but it's a new way to interface with the business, and so that's the vision that we're building towards.

那这到底意味着什么呢？

So what does that mean?

首先，我们需要推出多种类型的对话场景，不只是被动的客户服务，对吧？

Well, one, we have to launch a lot of different types of conversations that an agent could have, not just inbound customer service, right?

目前我们的产品开发重心就在这里。

So that's where a lot of our product development is.

围绕提升互动体验、深入理解互动内容、实现自动学习这些方向，还有很多可以打造的产品，就像你刚才说的那样。

There's a lot of other products you can build in this theme of making that interaction better, being able to understand the interactions well, being able to auto learn, as you were saying, right?

我们和Duet合作推出的一个重要功能，就是Duet自动导航——它不需要你主动去查询，每天早上会自动为你生成一份报告。

One of our big launches with Duet was this concept of Duet autopilot where instead of you even having to go in and ask at things, it just prepares a report every morning.

你早上一醒，登录系统，Duet就会告诉你：昨天的所有对话我都看过了，这是整理好的情况报告。

You just wake up, you log in, and Duet tells you, Yeah, I read all the conversations from last day, and I put together this report of what's going on.

这些就是我们正在推出的功能，我们觉得这个愿景实在太宏大了，要做的事情多得是，短期内根本不用担心没东西可做。

So those are the things that we're launching, and we just think that that vision is just so large, and there's so much to do there that we're not gonna run out of things to build anytime soon.

杰西，非常感谢你今天来做客。

Well, Jesse, I want to thank you for coming on.

这位是Decagon的首席执行官杰西·张，欢迎来到TI TV。

That is Jesse Zhang, the CEO of Decagon here on TI TV.

顶尖研究人员之间的AI人才争夺战，未来几个月很可能仍是主流话题。

The AI talent wars over top researchers is likely to be another dominant story for many months to come.

根本原因在于，AI研究人员还有很多关键问题亟待解决。

The core reason, of course, is that there are still some meaty questions AI researchers still need to solve.

软件能降低推理成本吗？

Can software cut inference costs?

智能代理自我提升的速度能有多快？

How quickly can agents improve themselves?

递归式自我改进真能实现吗？

Will recursive self improvement become a reality?

接下来我想请我们的AI与机器人记者罗克特·德鲁来聊聊这些话题。

I wanna bring on our AI and robotics reporter, Rocket Drew, for a conversation about all of those topics.

罗克特，欢迎来到节目。

Rocket, welcome to the show.

很高兴你能来。

It's great to have you here.

非常感谢。

Thanks so much.

能亲自到场真的很棒。

It's great to be here in person.

太有意思了。

So fun.

我得说一句，其实我才是来访的客人。

I should say, I mean, I'm the visitor.

你这儿常驻。

You are based here.

没错。

That's true.

确实如此。

That's true.

你开麦了。

You're on

这是你的地盘。

my home turf.

好了。

There we go.

行。

Okay.

我们刚才跟张杰聊了开源模型的事。

So we just talked to Jesse Zhang about open source models.

明白吗？

Okay?

我们稍微聊了点关于递归自改进的内容。

We touched a little bit on recursive self improvement.

持续学习这个话题，我们节目一直想搞清楚它到底是什么。

This topic of continual learning is one that we've been trying to wrap our heads around on the show.

你能帮我们理清一下，持续学习具体包括哪些内容？目前研究进展到什么阶段了？

Can you help us understand what is included in continual learning, and where is the research at right now?

是的。

Yeah.

当然可以。

Absolutely.

那是个

That's a

很好的切入点，因为持续学习现在正是人工智能研究前沿的热门话题。

great place to start because continual learning is a super hot topic on the frontier of AI research right now.

基本想法是，如果你有一个能持续学习的AI，它就能在工作中学习。

The basic idea is that if you have an AI that can continually learn, then it learns on the job.

它会在执行任务的过程中，从经验里吸收新知识、新信息和新技能。

It picks up new knowledge, new information, new skills from experience just as it goes about working and completing tasks for you.

我们现在说的是模型，还是智能体？

And we're talking about the models or the agents right now?

这是个好问题。

Well, that's a good question.

我们就先说模型吧。

Let's say the models.

咱们假设自己是个纯粹派。

Let's say we're being kind of purists.

我们真正希望的是，模型能自己适应，灵活地获取新信息。

We really want the model to be able to adapt and flexibly pick up new information.

目前其实有一些变通方法，能让智能体实现持续学习，我稍后解释。

Now there are kind of workarounds right now to get your agent to sort of continually learn, and I'll explain that in a second.

但我觉得有些人可能会对这个想法有反应，比如：聊天机器人不是已经能持续学习了吗？

But I think some people might have a reaction to this idea that's like, well, can't chatbots already continually learn?

ChatGPT 不是用得越多就越了解我吗？

Doesn't ChatGPT learn more about me as I use it?

没错。

Right.

在某种程度上确实如此。

That's true to some extent.

但别忘了，现在AI模型的训练方式是：先经历一个大规模的训练阶段，这期间是独立的，之后就直接使用模型了。

But as a reminder, the way AI models are trained these days is you have one big phase of training, and that's kind of a discrete period, and then you just use the model.

在使用过程中，模型的参数——也就是它的‘大脑’——其实是被锁定的。

And basically the weights of the model, the brain of the model, are frozen while you're using it.

我觉得这就像第一阶段你打造工具，第二阶段你才开始用这个工具。

I think of it like in stage one, you're forging the tool, and in stage two, you're using the tool.

持续学习则重新思考了这种模式。

Continual learning reimagines this.

它说的是，如果这个工具在你使用的过程中能不断调整优化，变得越来越擅长手头的工作，那会怎么样？

It says, what if the tool could adapt and change as you use it to become better and better at the jobs?

现在确实有一些变通办法，但说实话，都挺粗糙的。

Now there are some workarounds, but they're very they're kinda janky, frankly.

我觉得这正是关键所在。

I think that's the way to

没错。

put it.

比如当ChatGPT随着时间越来越了解你时，它其实是在给自己留下一些备忘，方便以后参考。

Like when ChatGPT learns more about you over time, it's leaving little notes to itself that it can reference at a later date.

就像你提到的，一些智能代理也有类似的临时方案。

And like you mentioned, some of the agents have workarounds as well.

它们也能给自己留备忘记录。

They can leave notes to themselves

对。

Right.

为了提醒自己该怎么完成任务。

To remind themselves how to solve a task.

但这种方式一点都不优雅。

But it doesn't feel elegant.

这跟人类根据经验即时学习新东西的方式完全不一样。

It's not like the way humans pick up new knowledge on the fly and and from experience.

就像你雇了个实习生，他干了一天。

It's like you hired an intern, and the intern works for a day.

到下班时，他稍微熟练了一点，于是给自己留了张便条，结果第二天醒来却全忘了，像《土拨鼠之日》一样，重新到办公室，看别人留的笔记，努力回想之前该怎么做。

And by the end of the day, it's a little bit better at its job, but then it leaves a note to itself, and it wakes up the next day with amnesia, like Groundhog Day, and comes into the office and reads the notes that someone else left for it and tries to remember how to

工作，这根本算不上真正意义上的即时学习和灵活掌握新技能。

work, it doesn't really feel like the true solution to learning on the fly and flexibly incorporating new skills.

那我们现在关于持续学习的情况到底怎么样？

So where are we at then with continual learning?

我们至少还能留这些笔记。

We can leave these notes.

现在有没有什么应用已经实现了研究人员所期望的持续学习方式，还是说这仍然是我们正在努力的方向？

Are there applications today that incorporate continual learning the way researchers want, or is this something that we are working towards?

毫无疑问，我们还在朝着这个目标前进。

We're still working towards it for sure.

我觉得目前我们只是用一些临时的解决办法，而且这些办法会变得越来越好。

I think right now we have these little workarounds, and they will continue to get better and better.

有些人认为，我们真得从头再来，重新设计AI模型的架构和训练方式，才能从根本上提升它们持续学习的能力。

Some people think we really do need to go back to the drawing board and rethink the way AI models are designed and the ways that they're trained in order to really fundamentally improve their ability to continually learn.

我们确实有一些有希望的研究方向，但眼下所有尝试持续学习的方法都存在明显的局限性。

We have some some promising directions to explore, but there are big limitations right now with all of the ways that we try to do continual learning.

比如说，你可以收集AI为你完成任务时产生的大量数据，然后用这些数据来微调模型。

So for example, you could take a bunch of data from the AI doing jobs for you, and then you could fine tune on that data.

但这种方式实在是太粗暴了。

But this is a really blunt way to go about it.

你只是让AI去模仿它之前做过的一切，然后原封不动地复述出来。

You're just getting the AI to imitate everything that it's done, and just sort of regurgitate it.

而且你还面临

And you also run the

AI会忘记它以前知道的东西的风险。

risk of the AI forgetting things that it used to know.

你说的‘模仿’是什么意思？

When you say what do you mean by imitate?

就是简单地重复它以前做过的事情，而不是灵活地把新技能融入到它的运作方式中。

In just like repeat, the the things that it has done before, rather than, like, flexibly incorporate new skills into the way

它的工作方式。

it works.

比如，想象一下，这就像假装在持续学习。

So for example, think of It's like pretend pretend continual.

没错。

Exactly.

没错。

Exactly.

你可以这么想：我能把国际象棋特级大师的每一步都记下来，然后针对这些步法做微调，但这并不会让我自己变成特级大师。

Think about it as like, I can write down all of the moves that a chess grandmaster makes, and I can sort of fine tune on those moves, it's but not going to make me a chess grandmaster myself.

我得真正理解特级大师走每一步背后的原理和原因。

I have to learn the principles behind the moves that the chess grandmaster did and learn the reasons they took those moves.

这才是实现持续学习真正需要的东西。

That's what it would really take to have continual learning.

明白了。

Okay.

那我们现在回头看看递归式自我改进。

So now let's go back to recursive self improvement.

这算不算持续学习的一部分？

Is that a part of continual learning?

超级持续学习？它又该归到哪一类？

Is super continual Where does that fall into it?

我觉得咱们先把它俩分开来看。

I would think about them separately to start.

它们之间确实有联系，但我建议先分开来看。

There's ways that they're related, but I would think about them separately.

我的意思是，即使我们还没解决持续学习的问题，只要AI能从新经验中学习、从实践中成长，依然可以实现递归式自我改进——也就是让一个AI模型独立创造出下一代AI，全程无需人类干预。

What I mean by that is even if we don't solve continual learning, AI that can learn new experiences, learn from its experiences and learn on the job, you could still have recursive self improvement, which is the idea that you could have one AI model create its successor, so train the next generation of AI, without any humans involved.

整个过程完全不需要人类参与。

No humans in the loop at all.

整个研究循环都由AI自己完成。

The entire research loop is done just by AI.

在当前的框架下，你照样能做到这一点：依然是一轮训练，然后把AI模型部署到现实世界中。

You could do that under the current paradigm, where we still have one training phase, and then you put the AI models out into the world.

它们的关联在于，两者都是通向超级智能AI的路径。

The way that they're related is they're sort of both paths to creating a super intelligent AI.

你可以想象，与其依赖递归式自我改进，不如让大量AI模型分布在世界各地。

As in you can imagine that instead of recursive self improvement, you could have a bunch of AI models out there in the world.

它们各自服务于不同的领域。

They're working with different enterprises.

它们在不同的机构工作，不断掌握各种新技能。

They're working with different organizations, and they're picking up all sorts of new skills.

然后，这些知识 somehow 会被汇总起来，反馈给某个更中心的模型。

And then somehow they're accumulating that knowledge and feeding it back into some more central model.

即使人类还在某种程度上参与训练过程，这种方式也能让模型迅速变得非常强大。

That's a way that models could also get very good very quickly, even if humans are still involved in the loop of training them to some extent.

那我想问问你，关于持续学习和递归自我提升，这里有哪些权衡？我的意思是，我们昨天刚和斯蒂芬妮聊过。

So I want to ask you then about some of the trade offs here of having continuing learning, achieving recursive self improvement, because mean, look, we talked with Stephanie yesterday.

她刚参加完 ICML 回来，跟我们说 AI 研究人员现在开始紧张了——以前是程序员担心 AI 会取代他们，现在连研究人员自己也开始担心了。

She came back from ICML, and she was telling us about how AI researchers are getting nervous that, Hey, I mean, first it was AI coders that were coders that were nervous that AI could do their job.

他们说，也许 AI 也能做我的工作。

Now researchers are saying, Maybe they could do my job as well.

所以这里确实存在一些焦虑情绪。

So you have a bit of anxiety there.

但另一方面，大家也对 AI 的普及感到兴奋，觉得持续学习这类技术进步，或许能推动企业更广泛地采用 AI。

On the other hand, you have the issue of AI adoption and some excitement that, well, maybe there's an opportunity here for continual learning and stuff like that, these advancements, to improve adoption in the enterprise.

你觉得实现持续学习能在多大程度上推动AI的普及，又在多大程度上可能让AI研究人员失业？

To what extent do you think that achieving continual learning will help improve adoption, and to what extent do you think it might put AI researchers out of a job?

我的意思是，这两个问题都很有分量。

I mean, I think- Two meaty questions.

确实如此。

Definitely.

我觉得AI研究人员现在有点紧张是有道理的，毕竟他们的工作可能被自动化取代。

I think AI researchers are right to be a little bit nervous right now about being automated.

你说得对，写代码只是研究人员日常工作中的一小部分。

To your point, coding is just one piece of what a researcher does in their day.

他们还要设计实验。

They also design experiments.

提出假设。

They come up with hypotheses.

执行实验。

They run experiments.

他们还要解读结果。

They interpret the results.

他们提出新的假设。

They generate new hypotheses.

我的意思是，这正是科学方法的全过程。

I mean, that's the scientific method.

对吧？

Right?

目前，整个流程里最被自动化取代的部分就是写代码。

Coding is the piece of that that's most automated right now.

你看到Anthropic最近宣布了，Claude已经能写他们80%的代码了。

You saw Anthropic recently announced that Claude is writing 80% of their code.

OpenAI也刚说，他们内部用于研究的计算资源里，过去六个月花在写代码上的开销增长了整整一百倍。

OpenAI recently said that of their research budget, their the compute that goes to research internally, the amount spent on coding has increased a 100 times over the past six months.

天啊。

Wow.

所以编程绝对是当前的前沿，但人工智能正在越来越多地参与整个研究过程。

So coding is definitely the frontier here, but AI can do more and more of the whole research process.

我觉得人工智能研究人员有点紧张也是有道理的。

I think AI researchers are right to be a little bit nervous.

不过换个角度看，他们能完成的工作量会大得多。

Another way to think of this though is they'll be able to get a lot more done.

工作效率会比以往任何时候都高。

Be more productive than ever.

他们每花一小时工作，就能完成相当于过去一个人一千小时的研究量。

For every one hour of time that they spend at work, they're getting a thousand hours of research done, or what used to take a a person a thousand hours.

那企业里的应用情况呢？

What about what about adoption in the enterprise?

应用规模非常大。

Adoption is huge.

这才是关键问题。

This is the big question.

这确实是关键问题。

It is it is the big question.

我的意思是，持续学习带来的收益，会远远超过我们迄今为止看到的任何AI应用。

I mean, continual learning would be so insanely profitable compared to what we've seen for AI so far.

原因还得回到我之前打的那个实习生的比方。

The reason goes back to that analogy I gave you about an intern.

对吧？

Right?

想象一下，现在的AI就像个实习生。

Imagine AI right now is like an intern.

你也许能教会它做某一项具体任务，它也能基本做好。

Maybe you can teach it to do one specific task, and it kind of mostly does that task right.

但实习生的用处其实很有限。

But interns are not that useful.

对吧？

Right?

在公司里，实习生其实挺让人头疼的。

In in a company, interns are kind of annoying.

你雇实习生，

You hire interns

不是因为他们有多出色，你也没抱太大期望。

not because they're Rocket you're not being very

不好意思。

I'm sorry.

我以前也当过实习生。

I've been an intern.

我知道自己那时候是什么样子。

I I know what I was like.

你当时是个不错的实习生。

Just You were a good intern.

我是说真的。

I'm telling.

也许吧。

Maybe.

也许吧。

Maybe.

我当过好几次实习生。

I've been an intern a few times.

我觉得我那时候占了不少资源。

I think I I took a lot of work.

但你招实习生，是因为他们最终会变成正式员工。

But you hire interns because eventually they become employees.

他们慢慢熟悉了工作，以后就能派上用场了。

Eventually they learn the ropes, and they can be productive later on.

如果现在的AI就像实习生，那只有持续学习才能让他们有朝一日成为正式员工。

If AI right now is like an intern, it's continual learning that would let them become an employee someday.

如果没有持续学习，他们可能就永远困在实习生的状态里了。

Without continual learning, the thought is maybe they'll be just stuck in intern mode forever.

所以你可以自动化工作流程中的某些环节，但还是得有人在一旁帮忙、绕着它们干活。

So you can automate specific pieces of workflows, but you'll still have to have people hand holding them and working around them.

没错。

Right.

那我再问你一个问题。

So let me ask you one more question then.

我们一直在讨论前沿模型的优势，以及到底需要多大程度上依赖这些模型的全部能力。

We talk about the strength of frontier models and the extent to which you need the full power that these frontier models have.

但另一方面，其实只要模型能达到某个基本门槛就够了。

Then you have, well, we only need the models to hit a certain threshold, essentially.

这其实也牵涉到开源模型的整个讨论。

This is the whole conversation around open source models too.

我们真的需要前沿模型吗？

Do we need the frontier model?

我的问题是，如果看我们希望AI完成的全部工作量，我们是否真的需要持续学习才能应对所有这些任务？

My question is, if you look at the aggregate volume of work that we want AI to do, do we need continual learning to do all of these tasks?

实习生做的那些事，也没必要每个人都得是火箭专家才能在职场成功，对吧？

The intern tasks, it's not like everyone needs to be a rocket scientist to succeed in the workplace, right?

没错。

That's right.

我也是这么想的。

That's how I'd think about it too.

有些工作其实规模很小，完全可以交给AI模型来处理。

There are intern sized tasks that you could have AI models handle.

对于其中一些任务，其实现在连顶尖模型都不需要。

And for some of those, you don't even need the frontier models today.

用一个更小的开源模型，稍微微调一下，专门应对这个任务就够了。

You can make do with a smaller open source model that's maybe fine tuned to be better at that specific task.

但总觉得我们还能做得更好。

But it feels like we should be able to do better.

我觉得完全有可能让AI模型像员工一样灵活工作，独立完成长期任务，而且做得很好。

It feels like there should be room to have an AI model that acts like an employee flexibly and can work autonomously on long range tasks and perform them very well.

所以，这种观点可以归为一种自我安慰或乐观的看法，认为达到那个阶段并不需要持续学习。

So put this take in sort of the bucket of cope or hopeful takes that continual learning won't be necessary to get to that point.

意思是，实际上，大多数任务根本没那么需要高智商。

It's like, well, maybe in practice, most tasks just didn't require that much intelligence.

没错。

Right.

你可以把整个工作流程拆分成一个个独立的小任务，然后交给不同的模型去处理。

And you can basically carve up a workflow into these different discrete chunks and pass them off to different models.

另一个属于这种自我安慰类观点的是，认为单纯扩大规模就足够了。

Something else that's in this bucket of kind of cope like takes about continual learning, I think, is that scaling will just be enough.

到时候模型会足够聪明，每天上班时，哪怕它昨天的事全忘了，也不是实习生水平，而是员工水平。

That you'll get models that are smart enough in general that when it clocks in in the morning, even if it has amnesia, it's not intern level, it's employee level.

所以它记不住也没关系。

So it's okay that it's forgetting.

或者我们就把实习期间的所有经验全塞进上下文窗口里。

Or we'll just take all of the experiences of the internship, and we'll just cram them in a context window.

我们的上下文窗口会越来越长，所以它其实没必要更新权重。

We'll have a context window that's longer and longer, so it doesn't really need to update its weights.

它不需要改变自己的大脑。

It doesn't need to change its brain.

它只需要把过去所有操作的笔记都读一遍，靠硬生生地堆数据，就能足够胜任工作。

It will just be basically reading all of the notes from everything that it's done and just brute force it will know enough to be able to work effectively.

我觉得这些都不是持续学习的真正解决方案，但我们确实有办法绕开这个问题，根本不需要一个真正的解决办法。

I don't think any of these really feel like the solution to continual learning, but there are ways that we could get around that problem and not really need a true solution.

很好。

Great.

Rocket，非常感谢你今天来做客。

Well, Rocket, I wanna thank you for coming on.

这位是Rocket Drew，我们《The Information》的AI与机器人记者。

That is Rocket Drew, our AI and robotics reporter here at The Information.

目前，很多AI公司都在把增长重心放在物理AI上，不管是做世界模型的软件公司，还是想为预期中的机器人浪潮提供算力的芯片公司。

Physical AI is where a lot of AI companies are focused their growth efforts right now, whether it's software companies building world models or chip companies hoping to power the wave of robotics that everyone is anticipating.

阿德里安·麦克尼尔是机器人公司Foxglove的联合创始人兼首席执行官，他正处在这一故事的中心。

Adrian Macneil, the co founder and CEO of robotics company Foxglove sits at the center of that story.

我想请他来聊聊这个领域，以及当下正在发生什么。

I wanna bring him on for a conversation about that sector and what is going on in the current moment.

阿德里安，欢迎来到节目。

Adrian, welcome to the show.

很高兴你能亲自到场。

It's great to have you here in person.

谢谢。

Thank you.

感谢邀请我来。

Thanks for having me.

我昨天刚去过The Florist，看到了Foxglove。

I was at The Florist just yesterday, and I saw Fox glove.

你看到了啊。

You did.

是的。

Yes.

这是一种有毒的

It is a poisonous

花。

flower.

数字化。

Digitality.

没错。

Yeah.

毛地黄。

Digitalis.

花店老板告诉我的。

Florist told me.

我几乎可以肯定，公司名字就是这个。

Almost sure it was that as the company name.

我跟花店老板说我在采访你，你们公司叫Foxglove，她一听就说，哦，她 probably 是因为Digitalis才取这个名字的，毕竟那是一种毒花，结果你们居然做机器人公司。

And when I told The Florist that I was interviewing you and your company was was called Foxglove, she said, oh, she probably named it that and did a robotics company because it was called Digitalis.

是的。

Yeah.

对。

Yeah.

其实Digitalis这个名字更好，但域名被占了，只好改用Foxglove。

That would have been a better name, but the domain wasn't available, so we went with Foxglove.

就是这样。

There you go.

明白了。

Okay.

我对公司命名有个看法——名字不该直接绑定在你们做的产品上，因为公司做的事是会变的。

I have this thing with company names, right, where they should not be tied to what it is that you create, because what you create as a company evolves over time.

你们会调整方向，会转型，但名字可以自己创造，然后慢慢赋予它意义。

You pivot, you shift, but names that you can kind of create and you just give meaning to them.

对吧？

Right?

亚马逊到底是什么呢？

Like, what is what is Amazon?

苹果又是什么呢？

What is Apple?

这些名字本身跟业务没关系，但你反复说多了，自然就赋予了它意义。

These these names that you just start repeating so much, and you can you can give meaning to a name that is that is not directly related.

那你们公司的名字代表什么呢？

So what is the meaning of your company?

再跟我们说说吧。

Remind us.

你之前上过我们的节目。

We had you on the show before.

你们从贝塞默那里融了四千万美元。

You raised $40,000,000 from Bessemer.

简单说说重点就行。

Just give us the quick highlights.

这是给机器人用的软件。

So this is software for robots.

对，没错。

Yeah, Exactly.

我们打造了一个面向物理人工智能的数据平台和数据基础设施。

So we build a data platform, data infrastructure for physical AI.

它能帮你完成整个数据闭环：从现场机器人收集日志数据，传回来进行分析、理解、筛选出关键事件，排查问题、分类处理，最终训练出更优的机器人模型。

So this helps you with the whole data flywheel, your logging journey from robots that are out there in the production, bringing that data back, evaluating it, understanding it, sifting through that data and finding interesting relevant events, debugging, triaging, and then ultimately building better models for your robots.

那你们是自己开发模型，还是用别人的？

And and so you build your own models or whose models

模型是别人提供的。

are the models.

不是。

No.