本集简介
双语字幕
仅展示文本字幕，不包含中文音频；想边听边看，请使用 Bayt 播客 App。
欢迎各位收看《信息》频道的特别节目TITV。
Welcome everyone to a special edition of the Information's TITV.
我是阿卡什·帕什里卡。
My name is Akash Pasricha.
今天是7月15日，星期三，我们正在门洛帕克的瑞银私人AI软件与互联网大会现场，接下来会采访几位业内人士，聊聊当前软件行业的情况。
It is Wednesday, July 15, and we are here on the ground in Menlo Park at UBS' Private AI Software and Internet Conference, where we're gonna be talking to a number of people here about the current state of the software sector.
我们还会探讨人工智能研究的现状。
We're gonna talk about the current state of AI research.
也会聊聊机器人技术的未来，看看它能带来什么。
We're gonna talk about the future of robotics and what that has to hold.
接下来会有几场非常精彩的对话，敬请期待。
We're gonna have some great conversations for you.
在正式开始之前，我先把话筒交给我们在旧金山的记者罗基特·德鲁。
Before we get started, I wanna turn it over to Rocket Drew who is in our San Francisco bureau.
他给大家带来一条简短的新闻更新。
He has a quick news update for you.
Rocket，你来吧。
Rocket, over to you.
谢谢，Akash。
Thanks, Akash.
《The Information》亚洲团队独家报道，DeepSeek 因收入强劲增长，正筹备第二轮融资，金额约74亿美元，并计划明年在上海上市。
The Information's Asia Bureau published exclusive reporting that strong revenue growth at DeepSeek is empowering the company to raise its second funding round of about 7,400,000,000 and plot a potential public offering in Shanghai next year.
他们最近的年化收入已达到4亿到5亿美元之间。
Their annualized revenue reached between $400,000,000 and $500,000,000 recently.
关于这件事，我们明天会带来亚洲团队的更多跟进报道。
We're going to have more from our Asia Bureau on that story tomorrow.
另一条消息：Stripe 和私募股权公司 Advent International 已提交收购 PayPal 的报价，金额超过530亿美元，据路透社报道。
Another piece of news: Stripe and PE firm Advent International have submitted an offer to buy PayPal for more than $53,000,000,000 according to reporting from Reuters.
报道称，PayPal 尚未对这一报价作出回应，而 Stripe 和 Advent 希望在未来几周内推动进一步谈判。
The outlet says PayPal has not responded to the offer and that Stripe and Advent want to advance discussions in the coming weeks.
我们会持续关注这件事的进展。
We'll keep our eye on that story.
与此同时，即将在今年被SpaceX以600亿美元收购的代码生成初创公司Cursor，正在规划一场重大的转型。
Meanwhile, Cursor, the coden startup set to be acquired by SpaceX for $60,000,000,000 later this year, has been mapping out a major transformation.
下面我们请同事Grace Kay来深入聊聊这个故事。
Let's bring in our colleague Grace Kay for more on this story.
欢迎再次来到节目，Grace。
Welcome back on the show, Grace.
嗨，Rocket，很高兴能来参加。
Hi Rocket, great to be here.
你听说了今年早些时候公司全员大会的事吧？
So you heard about a company wide meeting earlier this year.
当时Cursor的首席执行官迈克尔·特鲁尔对员工说了些什么？
What was it that Cursor CEO, Michael Truell, told his staff at that meeting?
是的。
Yeah.
迈克尔提出了一套全新的公司战略。
So Michael outlined kind of a new strategy for the company.
你知道吗，这家曾经以编程起家的公司，现在正转向通用人工智能模型。
You know, what has famously been a coding startup is now pivoting towards general AI models.
他定下了一些时间节点。
He put some deadlines out there.
他们希望到今年年底能跻身顶尖AI公司之列。
They wanna be one of the top AI companies by the end of the year.
他们计划到2027年拥有最多的算力，并在那时推动人工智能的边界。
They wanna have the most compute, you know, by 2027 and be pushing the boundaries of AI by then.
他还回应了员工对SpaceX收购的一些疑虑。
He also addressed some concerns from staff around the SpaceX acquisition.
可以说，这次会议对公司来说意义重大。
So, you know, it it was a pretty big meeting for the company.
这能解释为什么SpaceX一开始会对收购Cursor感兴趣吗？
Does this shed some light on why SpaceX AI was interested in acquiring Cursor in the first place?
对我来说，这件事一直有点让人捉摸不透。
I mean, there's always been a little bit of a mystery around that to me.
是的。
Yeah.
我觉得，一开始我心想，他们肯定是冲着编程数据来的，毕竟他们的Grok编程模型一直表现不佳。
I think, like, initially going into it, my thought was, you know, obviously, they want the coding data just because they've been struggling so much, you know, with their Grok coding model.
但迈克尔在宣布这项合作后不久的一次会议上提到，SpaceX之所以特别兴奋，是因为Cursor的品牌影响力，还有他们那些企业客户关系。
But one thing that Michael noted, in one of the meetings shortly after, you know, this partnership was announced, he talked about how SpaceX was really excited because of Cursor's brand and, you know, some of the enterprise relationships they have.
很多《财富》500强公司都是他们的客户。
A lot of Fortune 500 companies are, you know, their customers.
他们还有一个规模大得多的市场和销售团队，远超xAI的规模。
They also have a really large go to market, you know, sales, marketing team, much larger than x AI's.
挺有意思的。
It's funny.
xAI之前根本没什么品牌知名度。
Like, x AI didn't have a great brand going into this.
我真想不出，他们还会看上哪家公司，觉得对方的品牌比自己还差。
I it's hard to imagine they would look at any company and think that brand has a worse brand than ours.
但Cursor确实品牌不错，这么想也合理。
But I guess Cursor does have, like, a great brand, so it makes sense.
现在Cursor的员工对这次收购感觉怎么样？
How are staff feeling about the acquisition right now at Cursor?
是的。
Yeah.
我觉得大家看法不一。
I think feelings are mixed.
有些人可能挺兴奋的，毕竟能和SpaceX合作。
Think there are some people who are maybe excited to partner with SpaceX.
他们马上就要上市了。
They have the IPO.
现在是个大公司了。
It's a big company right now.
但也有不少人对这次收购感到担心。
But there are also a lot of people who are concerned about the acquisition.
我跟一些人聊过，他们提到埃隆·马斯克收购推特时，裁掉了公司四分之三的员工。
I talked to some people who They looked at Elon Musk's acquisition of Twitter when he cut three quarters of the company.
任何收购都难免让人担心裁员，这很正常。
There's obviously concerns about cuts, I think, with any acquisition.
但埃隆·马斯克一向以这种方式出名。
But with Elon Musk, he's he's kind of known for that.
所以这确实是个让人担忧的问题。
So that's that's definitely a concern.
是的。
Yeah.
他作为老板，本来就不好相处。
He's kind of hard to work for at best.
等他收购了你的公司，做法可能就更不留情面了。
And then when he acquires your company, it can be a a little ruthless, I guess.
现在Cursor在收购前有哪些变化？
How is Cursor changing ahead of the acquisition right now?
是的。
Yeah.
所以 Cursor 正在努力转型，想成为一个更通用的 AI 公司。
So Cursor is, you know, trying to pivot to become more of a generalized AI company.
他们正在开发一个对标 Claude 的助手，内部叫 Sand，已经让我们内部测试了。
They're working on this Claude coworker competitor, which is internally being called Sand and, you know, that they've rolled out in our testing.
他们也在研究一些聊天机器人。
They're also looking at some chatbots.
他们在探索各种方式，想突破编程这个领域。
They're looking at different ways they can push beyond coding.
迈克尔内部说，这其实是客户一直推动的方向，但有意思的是，他在员工会议上提到的一些目标，和埃隆·马斯克在 xAI 设定的方向非常相似。
Michael internally has kind of characterized this as something that customers have been pushing for, But it is interesting that, like, some of the goals he outlined in these meetings with staff are very similar to goals that Elon Musk has outlined at x AI.
所以现在看起来，他们俩虽然各自独立，但很多想法其实不谋而合。
So it's kinda similar how, like, right now they're in parallel, but, like, they're aligned on a lot of things.
Sand 这个代号挺有意思的，给这么一个重要的新产品起名叫‘土’，听起来有点搞笑，不过倒也挺有味道。
Sand is kind of a funny code name for your upcoming really exciting new prize, sort of like naming it like dirt, but I guess it's kinda cool.
人们总开玩笑说，AI芯片就是把沙子熔了，教会它思考。
It's like people joke about AI chips being just like sand that we melted down and taught how to think.
所以这个名字可能还真有点儿意思。
So maybe there's something kind of feeling about that.
是的。
Yeah.
我觉得公司内部代号都挺有意思的，比如以前有家公司叫‘大蒜’，Meta那边还用过‘牛油果’之类的。
Think the code names are always interesting because like, you know, there was garlic at some, you know, one of the companies that was avocado at meta, like.
就像一场饼干野餐。
It's like a cookie picnic.
对。
Yeah.
像这种即将被收购的公司，反垄断法规具体要求是什么？
So what do antitrust rules require for a company like this that's pending an acquisition?
他们现在挺尴尬的，明明想一起干很多事情。
They're in this awkward spot where they wanna do all this work together.
他们可能已经在做计划了，但交易还没最终完成。
They're probably making plans, but the deal hasn't totally gone through yet.
是的。
Yeah.
这真的很有趣。
It's really interesting.
通常情况下，他们得等到交易完成才能真正合作。
Ordinarily, they wouldn't really be able to work together until the deal went through.
但由于他们之前就有合作关系，涉及共同开发模型和其他AI产品，情况就变得特别复杂。
But because they have the separate partnership that includes working on models and other AI products together, it's very complicated.
我觉得现在他们尽量保持独立，但因为那个合作关系，实际上又在很多方面紧密合作。
I think right now they're, like, trying to keep separate, but there's also, you know, a lot of ways that they're working together very closely because of that other partnership.
埃隆在这次收购中扮演了重要角色吗？
Is Elon taking a big role in the acquisition himself?
我最初听说，在XAI那边，埃隆·马斯克确实和迈克尔见过面。
So initially, I'd heard at XAI, you know, Elon Musk was meeting with Michael.
他一直在跟阿曼见面，就是Cursor的联合创始人之一。
He was meeting with Aman, you know, the the co founders of Cursor.
而且，有人看到迈克尔出入过XAI的办公室。
And, you know, Michael has been seen at XAI's office.
但在Cursor这边，大家见到埃隆·马斯克的次数就少多了。
But on the Cursor side, you know, they're not seeing Elon Musk as much.
他还没怎么跟普通员工接触过，之类的。
He hasn't really met with rank and file or anything like that.
他倒是见过Cursor的一些高层。
He has met with some higher ups at Cursor.
所以现在看来，埃隆的参与度没那么高，估计得等到交易正式完成才会更深入。
So it seems like right now, like, Elon isn't as involved, you know, and it probably won't be until the deal goes through.
明白了。
Okay.
我们有没有听说Cursor打算怎么融入XAI的整体布局？
Do we know yet how Cursor hopes to fit into the bigger picture at x AI?
我的意思是，如果它自己在开发所有模型和产品，那Grok还有空间吗？
I mean, if it's making all of the models itself, it's making these products, does that leave any room for Grok?
是的。
Yeah.
我觉得这个还得再看看，等后续发展吧。
I I think that's something we'll kind of, like, wait to be seen.
目前来看，Cursor正在大力扩充销售团队，努力把自己打造成企业级解决方案，自四月以来，他们的销售团队规模已经扩大了60%。
Right now, it seems like Cursor is really building up their sales, you know, team and and trying to, like, sell themselves on being the enterprise, bringing in they they've, like, added 60% to their GTM team since April.
看来他们现在最重视的就是这块。
So that seems to be something they're really bulking up on.
我觉得他们肯定会在编程工具这条线上发挥关键作用，毕竟他们手上有独特的数据和专业能力。
I also think, you know, they're gonna be a huge part of the coding push, obviously, with their data that they're coming in with and their expertise that way.
没错。
For sure.
确实，即使Cursor的业务范围扩展到其他模型和产品，专注于编程工具依然是很合理的方向。
Well, it makes sense that Cursor would maintain a strong focus on coding even as its purview expands to include other kinds of models and products.
谢谢你能来参加节目，Grace，为我们做了这么清晰的解析。
So thanks for coming on, Grace, and breaking that down for us.
不客气。
Thanks.
有独家报道指出，在当前AI算力紧缺的背景下，越来越多初创公司开始转向新兴的云服务商，以获取所需的NVIDIA显卡。
The information published exclusive reporting that more startups are turning to newer cloud providers to secure the Nvidia GPUs they need during the ongoing AI compute crunch.
我们的亚马逊记者凯瑟琳·佩洛夫撰文分析了初创公司为何选择这些新兴服务商，而不是AWS。
Our Amazon reporter, Katherine Perloff, wrote about why startups are choosing these newer providers over AWS.
现在我们请凯瑟琳来详细说说具体情况。
Kathryn joins me now with the details.
欢迎来到节目，凯瑟琳。
Welcome on the show, Kathryn.
你好，罗克特。
Hi, Rocket.
嗨。
Hey.
你那篇文章开头提到一个开源AI开发者RC，他们原本承诺向AWS投入800万美元，却根本拿不到所需的NVIDIA芯片。
So you opened up your piece talking about an open source AI developer, RC, which had committed $8,000,000 to AWS but couldn't get the Nvidia chips they needed.
现在AI初创公司遇到这种问题有多普遍？
How common is this issue for AI startups right now?
我觉得这在行业内其实挺常见的，而且不光是AWS这边有问题。
I think, you know, it is a it's not an uncommon problem throughout the industry, and it's, you know, not just with AWS.
NVIDIA的芯片本来就难抢，整个行业都存在算力不足的情况，就连一些新兴云服务商那儿也一样紧张。
The, you know, Nvidia chips are hard to come by, and there are capacity issues everywhere, even in at some of the neo clouds.
但我接触的那些初创公司都说，在AWS上根本拿不到他们需要的NVIDIA芯片。
But, you know, the the startups I talked to found that it was just really hard to get the Nvidia chips they needed at AWS.
有时候价格太高，有时候又只能整批购买，要求的承诺量远超他们的实际需求，他们更想要灵活一点的方案。
Sometimes that they were just too expensive or they were sold in they were only available in sort of bigger chunks or bigger commitments than they needed, and they wanted a kind of more flexible arrangement.
当然，还有一些其他原因，让公司觉得新兴云服务商或别的云初创企业更适合自己。
And there are some other reasons that companies might find, you know, a neo cloud or another type of cloud startup more suitable for their needs.
但说到底，现在抢算力的时候，AWS也不总是能供得上。
But yeah, in the hunt for capacity, AWS doesn't always have it.
原来如此。
I see.
所以这些初创公司并不是简单地被拒之门外，被告知完全没有资源。
So it's not always that these startups are getting turned away at the door and being told there's literally no capacity.
有时候他们说，我们有资源，但如果你想要，就得接受这些可能并不符合你需求的条件。
Sometimes we say, Well, we have capacity, but if you want it, you're going to have to accept these terms that maybe aren't what you're looking for.
是的，我觉得有时候情况是这样：有资源吗？
Yeah, I think it's like, you know, sometimes it's like, Okay, is there capacity?
答案是没有，但你明天再看看，也许会有，只是价格太高了。
And it's like, No, but you look tomorrow, maybe there's something, but it's too expensive.
或者，如果你非要 exactly 需要的配置，就得一次性买一年，可你又没那么多钱。
Or it's like, If you want exactly what you need, you have to buy it for like a year and you don't have the money for that.
所以这其实是多种因素叠加的结果。
So it's kind of a combination of things.
你知道，AWS 说他们没有最低消费要求。
You know, AWS says like they don't have a minimum.
如果你想买一块芯片，也可以单独买一块，或者按块租用、购买计算资源。
And if you want to buy one chip, you can buy one chip or, know, rent, rent, buy compute for one chip.
当然可以。
You can.
我觉得这些选择确实存在，但实际中，它们未必符合初创公司的真正需求。
So I think, you know, like I think the options exist, but you know, in practice, they're not always what a startup might be looking for.
是的。
Yeah.
你提到过，Nebius 这类新兴云服务商，有 75% 的业务来自那些已经在大厂用满配额的初创公司。
You wrote that Nebius, one of these neo clouds, is getting 75% of its business from startups who have already maxed out at the big cloud firms.
是这样吗？
Is that right?
那从 AWS 这样的大厂转向新兴云服务商，有哪些优缺点呢？
So what are the pros and cons of switching to a neo cloud from a big provider like AWS?
我觉得，没错。
I think, you know yeah.
这挺有意思的。
It's interesting.
所以在这种情况下，你说有75%的公司已经试过微软、AWS和谷歌，但再也拿不到更多算力了，只好转向Nebius。
So the in that kind of, like, context, you said 75%, you know, they've already tried to use Microsoft, AWS, Google, they can't get any more capacity, so they go to Nebius.
我觉得，像Nebius这样的新兴云服务商，还有RunPod这类基础设施提供商，虽然不一定都叫‘新兴云’，但它们确实都在这个圈子里。
I think, you know, the the pro of a Neo Cloud or sort of like these infrastructure providers that don't know if you'd quite call them Neo Clouds, these together and RunPod.
它们也都在参与这场竞争。
They're also sort of in the conversation.
它们更多提供的是推理即服务，但用户还是会把工作负载跑在这些公司上。好处是，有时候在那里更容易找到算力，而且它们往往更灵活，更愿意临时跟初创公司谈合作。
They sort of offer more like inference as a service, but people are still running workloads on those companies instead of The advantage is, know, sometimes it's easier to find capacity there and sometimes they can be a lot more flexible in what they can offer or they're more likely to sort of do a deal with the startup on the fly.
我听说的另一个优势就是这种灵活性。我接触过一家专门做训练后处理的初创公司。
Another advantage I've heard is that on that flexibility, I talked to one startup that kind of specializes in post training.
他们说，现有的云架构根本不适合他们，但和RunPod合作后，他们能一起打造出更适合自己的环境。
And they said, you know, a lot of the current cloud architecture doesn't really work for us, but we were able to kind of like work with run podded together to kind of create an environment that was better for us.
所以我觉得，这些平台在定制化和灵活性上确实更有优势。
So I think there's just like a bit more customizability, and flexibility.
话虽如此，NeoCloud也不是万能解方，英伟达的芯片现在很难搞到。
You know, having said that, you know, the NeoClouds are not a panacea, and the Nvidia chips are hard to come by.
所以有时候，NeoCloud的费用反而更高。
So, you know, some sometimes the Neo Clouds can be more expensive.
它们并不总是更便宜。
They're not always cheaper.
有时候也会遇到容量不足的问题。
Sometimes they have capacity issues.
不过我觉得，关键还是灵活性，还有有时候价格更低，这往往是因为不用提前租太多资源。
So, you know, but I I think it's more just sort of flexibility and, you know, at times, like, a cheaper price, which can also just come from being able to rent less upfront.
是的。
Yeah.
即便没有现在这种算力紧缺、大厂拒之门外的情况，这种灵活性对很多初创公司来说依然非常有吸引力。
Still, that flexibility is probably really appealing to a lot of startups even if there wasn't this dramatic compute crunch where they were getting turned away from larger providers.
不过你提到AWS，他们对这事怎么看？
But you asked AWS, of course, what do they think about this?
他们给了你这么个说法，说实话，我觉得Red有点防御性，但他们说，几个例子并不能说明趋势。
And they gave you this statement that, like, frankly, to me, Red is a little bit defensive, but they said that, you know, a few anecdotes does not make a trend.
所以他们否认了这里存在某种趋势的说法。
So they pushed back on the idea that there was a trend happening here.
你怎么看这个说法？
What did you think of that?
在你看来，这算是一种趋势吗？
Is it a trend in your opinion?
我觉得，很明显AWS的表现很不错。
You know, I think that, you know, it's undeniable that AWS is doing well.
他们的增长直接提速到了28%，比上个季度高出四个百分点。
Like, they just they accelerated their growth to 28%, four percentage points from the prior quarter.
有时候，这些新兴公司的崛起反而可能帮了AWS，比如像Vercel和Render这样的云服务初创公司，其实都是跑在AWS上的。
And, you know, sometimes, like, some of the rise of some of these upstarts might benefit AWS because, like, some of the sort of infrastructure as a service provider, some other kind of cloud startups like Vercel and Render run on AWS.
所以，这事儿没那么简单。
So, you know, it's a complicated story.
很多顶尖初创公司都在用AWS，比如OpenAI、Anthropic，不管你是否把它们算作初创公司，它们确实都在用。
And a lot of top startups do use AWS, obviously OpenAI, Anthropic, if you want to call them startups, you know, they do.
所以我觉得情况挺复杂的，但不可否认的是，这些新兴云服务商和基础设施公司确实在快速增长。
So I think it's a complicated picture, but I also think it's undeniable that a lot of these neo clouds and, you know, infrastructure providers are growing.
我们报道它们的融资轮次和营收情况，我一些同事也这么做。
We report on their, you know, funding rounds and revenue, and some of my colleagues do.
我觉得AI基础设施现在可能是风投领域最热门的方向之一。
I feel like AI infrastructure might be one of the hottest areas of venture investing right now.
如果这些公司都在增长，背后有投资逻辑，那一定有什么在推动它们的发展。
So if these companies are growing, if there's a thesis behind them, something has to be propelling that growth.
我接触过很多初创公司，它们觉得这些新兴云平台至少对部分工作负载来说是可行的替代方案。
And I talk to a lot of startups they're seeing these as viable alternatives for at least some of their workloads.
说得对。
Great.
看来计算资源紧张至少还会持续一段时间。
Well, it seems like the compute crunch is here at least for a while longer.
所以我觉得我们还得继续了解这些动态。
So I think we're going to continue to have to learn about these dynamics.
谢谢你跟我们详细分析了继续用 AWS 还是转向新兴云平台的利弊。
So thanks for explaining the trade offs to us about sticking with AWS or going to a Neo Cloud.
这对我们帮助很大。
It's very helpful.
再次感谢，凯瑟琳。
Thanks again, Kathryn.
谢谢。
Thank you.
好的，阿卡什。
Alright, Akash.
现在交给你了。
Back over to you.
我现在和瑞银的首席人工智能官丹尼尔·马加岑尼在一起。
I'm here with Daniele Magazzeni, Chief AI Officer at UBS.
Daniele，非常感谢你邀请我们。
Daniele, thank you so much for having us.
真的非常感激。
Really appreciate it.
谢谢你们邀请我。
Thanks for having me.
在加入瑞银之前，你是一位人工智能教授。
So before you were at UBS, you were a professor of AI.
所以我很想和你深入聊聊人工智能的一些技术话题。
And so I wanted to get into some of the technical topics of AI with you.
我的第一个问题是，目前人工智能研究者还没解决的技术难题有哪些？
My first question is, what are the technical challenges that AI researchers have yet to solve right now?
这是个好问题。
Yeah, it's a good question.
我认为，对于将人工智能应用到关键流程中，最关键的是要能信任这个模型。
A lot of number, I would say what is critical for the actual implementation of this AI in critical processes is the ability to trust this model.
因此，评估模型的可靠性和准确性非常重要。
Therefore, being able to assess the reliability and the accuracy of this model is very important.
从技术角度看，目前有很多研究在做评估测试，你可以把它们理解为一组测试用例，通过不同问题来检验模型的回答质量。
And from a technical point of view, there is a lot of research going on evals, which you can think about them as a test cases that given different questions, assess the answers.
不过，正如我们所知，AI本质上是概率性的，属于统计评估，这个问题确实很难，但并非无法解决，目前已有大量研究在推进。
However, given that this AI, as we know, is probabilistic, it's type of statistical assessment, you know, it's a hard problem, but it's doable and there is a lot of research going on.
另外，我们其实可以通过数学方法证明，智能体正在执行它该做的事。
The other thing is that you can actually mathematically prove that the agents are doing what they are supposed to do.
这并不容易，难度很大，但还是可以做到的。
It's not easy, it's hard, but it's doable.
我相信，朝这个方向的研究会非常、非常有价值。
And I believe that that research in that direction will be very, very valuable and very valuable.
那我们接下来就逐一探讨这两个挑战。
So let's get into both of these challenges then one by one.
关于评估测试，也就是模型自身的有效性。
So evals, I mean, this is the effectiveness of the models themselves.
我们有这些基准测试。
We have these benchmarks.
现在有个问题，到底哪个基准最好？我们该怎么衡量模型之间的差异？这其中有哪些难点？
There's a question around which benchmark is the best, how do we even benchmark the models against What each are the challenges with that?
就拿我们之前提到的信息来说，比如当模型知道自己正在被评估时，它们可能会装得比实际更好。
I mean, here at the information we've written about, for example, when the models know that they're being evaluated, they can pretend to be better than they are, I guess.
我以前从来没做过评估。
I've never done an eval before.
除了这些，关于评估和模型，还有哪些其他类型的挑战正在出现？
What are the other flavor of challenges with respect to evals and models that are coming up?
我觉得其中一个难点是，在AI的商业应用中，评估模型的能力其实取决于人的经验和主观判断。
Look, I would say one challenge is that when it comes to business use of AI, the knowledge to evaluate the models is actually in the human experience, human Subjective.
没错。
Correct.
而这一切，又取决于从事这项工作的人在过去一年里积累的专业知识和经验。
And based on, again, the expertise and the experience that humans doing the job, you know, had during the year.
所以问题在于，怎么把人的知识和经验转化成机器能理解的格式，让模型能用它来评估自己。
So the challenge is how do you translate the human knowledge and human expertise into a machine readable format so that they can actually use to assess the model.
这确实是个难题。
That's definitely one challenge.
就像你说的，人能骗模型的方式，跟骗人是不一样的。
As you said, you can fool the model in a different way than you fool humans.
虽然难，但特别重要。
So it's hard, but it's very important.
没错。
Yeah.
而且这类工作，真的需要AI研究者和业务领域的专家紧密合作。
And for these, I guess you really need AI researchers working very closely with the domain expert in the business.
还有就是智能体的有效性，得有数学证明它们真能完成你期望的任务。
And then the effectiveness of the agents, the mathematical proof that they're doing what you hope for them to do.
这不也是一种评估方式吗？
I mean, is that not just another flavor of eval?
我是说，关于智能体的有效性，现在都冒出哪些问题了？
I mean, what are the questions that are coming up with respect to agent effectiveness?
我觉得其中一个区别是，评估通常只针对一个大模型在特定任务上的表现，而数学证明则想考察一个智能体在它所承担的多个任务中的整体行为。
I guess one difference is that in eval, you assess one LLM into a specific task whereby in the mathematical proof, really want to assess the behavior of an agent across multiple tasks that this agent is tasked to.
这确实是个难题。
Again, it's a hard problem.
不容易，但也不是做不到。
It's not easy, but it's doable.
所以相关研究会越来越多。
So more and more research.
为了让我更清楚，你的意思是，这个结果不可能是靠运气碰出来的吧？
And just so I understand this, I mean, is this the idea that chance alone wouldn't have been able to produce that result?
你所说的数学证明，指的就是这个意思吗？
Is that what you mean by mathematical proof or
对
what
对。
Right.
你是说
Do you
统计评估和正确性证明之间有区别吗？
So is there a difference between a statistical assessment versus a proof of correctness?
顺便说一句，验证人工智能系统的正确性，早在大语言模型出现之前就是学术界的老话题了。
And by the way, verifying the correctness of AI systems is old topic in academia and in the research before the LLMs.
所以，一个可能的方向是借鉴以往的研究成果，看看它们能否以及如何被调整，以应对这种新型的生成式人工智能。
Therefore, I guess one opportunity is to leverage what has been done before and see how, if and how can be adapted to deal with this new generative AI.
我想问你关于安全的问题。
I'm gonna ask you about safety.
是的。
Yeah.
现在安全措施跟得上人工智能模型的创新步伐吗？
Is safety keeping pace with the innovation and AI models right now?
我觉得对安全的关注依然非常高。
I guess the focus on safety remains extremely high.
比如在瑞银，信任是至关重要的，是我们工作的核心。
Also for UBS, for example, trust is paramount, is the key element of how we work.
所以在使用人工智能时，我们也要确保能信赖所用的AI系统，这样我们的客户才能信赖我们使用AI的方式。
So also when it comes to AI, we want to make sure that we can trust the AI we use and therefore our clients can trust the way we use AI.
当然，如你所说，技术发展速度非常快，我们必须跟上步伐，但正因如此，我们才更想加速推进，同时绝不牺牲安全性。
Of course, as you mentioned, pace is very, you know, rapid evolving, so we need to catch up, but also that's why we want to go faster, we want to accelerate, but without compromising in safety.
所以，外部的技术和银行实际能上线部署的之间，总会有一定的差距。
So that's where there is always a gap between what you have outside versus what we can deploy in production in a bank.
没错。
Right.
但这正是因为我们在乎安全。
But that's because we care about safety.
所以在这里，瑞银你们是银行，属于金融服务机构。
And so here at UBS, I mean, you guys are a bank, financial services, financial institutions.
这是监管最严格的行业之一，你们在用前沿模型吗？
This is one of the most regulated industries that Are there you using the frontier models?
你们会用开源模型吗？
Are you using open source models?
你们得等上一年才能用这些模型吗？
Do you have to wait a year to work with these models?
你们的流程是怎样的？
What's your process?
首先，我们非常注重多样性。
So first of all, we really want to diversify.
我们既用前沿模型，也用一些小型语言模型，它们在越来越多的任务上表现很好。
So we use frontier models as well as small language models, which are good for a growing number of tasks.
我觉得，不是所有问题都需要用前沿模型来解决。
My view is that you do not need frontier model for a non frontier problem.
所以UBS的一个明确策略就是，确保员工为不同的问题或任务选择合适的模型。
Therefore, one clear strategy we have at UBS is to make sure that our colleagues use the right model for the right questions or for the right tasks they want to solve.
这些模型会为客户生成建议吗？
And is this generating recommendations for clients?
你们内部会用它们来写自己的工具吗？
Is this coding your own tools internally?
你们最终主要用这些模型做什么？
What do you end up using these models for?
有几个应用场景我很乐意分享，比如我们确实会用AI生成的洞察来更好地服务客户。
Well, a couple of use cases I'm happy to share is that when, for sure, we have a number of AI generated insights to better serve our clients.
我们的核心始终是如何利用AI更有效地服务客户。
Our focus is always how we can use also AI to better serve our clients.
归根结底，这还是一个靠信任运转的行业，但AI能帮助人们提供更深入的洞察。
It's still a business around trust, but AI can help people to provide better insights.
当然，在提升开发效率方面，我们大量使用了AI。
Of course, we use AI when it comes to developer productivity, we are using it a lot.
在研究方面，AI极大地提升了我们获取和产出信息的方式。
And when it comes to research, the way we consume information, we produce information is really enhanced by the AI.
我们看待AI的方式，就是如何让人工作得更聪明，更好地服务客户。
The way we see AI is really how to empower people to work in an even smarter way to better serve our clients.
这正是我们的重点。
That's our focus.
你刚才提到了小型语言模型。
And you mentioned small language models.
对。
Yeah.
在降低AI成本方面，这也是主要的做法吗？
Is that the dominant way for how you optimize for costs of AI as well?
是的，从成本角度，还有可持续性的角度看，这都很重要。
Yeah, look, it's important from a cost perspective, as well as for a sustainability point of view.
对于特别简单的问题，根本没必要用最新的大模型。
You really don't want to use the latest model for very, very simple questions.
所以得根据问题的复杂度，选用合适的模型。我们的策略是建立一个模型库，部署多种模型，再用一个AI系统来智能地把问题分派给最匹配的那个模型。
So you want to leverage, you know, the appropriate model for the question you have, and to do so, the strategy is to have, you know, a proper model garden where you host a number of models, and then to have an AI that can help route the right question to the right model.
我们并不指望每个人自己去选模型，而是会帮他们以最优化的方式搞定。
We don't expect everyone to select the models, but we want to do them for them in a very optimized way.
我很好奇，在你推动瑞银内部逐步采纳AI的过程中，有没有什么特别有效的诀窍？
I wonder, as you've led the process of getting UBS internally to adopt AI more and more, I mean, what's been a cheat code for you to adoption?
我们经常听到，要让AI真正落地，得先改变文化，教大家怎么用，建立信任。
We hear a lot about the cultural changes that have to happen in terms of using AI, teaching people how to use them, building trust.
那具体你是怎么做的呢？
How do you actually do that?
是靠办线上讲座吗？
Mean, is this running webinars?
是靠手把手带人吗？
Is this people holding your hand?
会不会派工程师直接到一线，坐在业务同事旁边，提醒他们：‘其实你可以试试这么操作’？
You know, do you have forward deployed engineers on the ground sitting next to associates saying, Hey, maybe you should do it this way, that way.
你们到底怎么推进的？
What do you do?
是的
Yeah.
这是个好问题。
First of all, it's a good question.
我总是说，这首先是个关于人的挑战和机遇。
And I always say it's people first challenge and opportunity.
无论技术多么重要和强大，真正关键的是我们的同事、还有我们自己，是否愿意去学习如何使用这项技术。
As important and powerful the technology is, is really the willingness of our colleagues, ourselves to want to willing to learn how to use this technology.
我发现有几个因素会拖慢大家的接受速度：首先，有些人可能一年前试过这些模型，那时候效果不太好，但现在模型进步了很多，他们需要重新回来试试。
I believe a few things we I noticed in terms of what is that slow down adoption is, first of all, maybe people tried models a year ago, weren't so good, but now models got better over time, so they really need to go back.
另外，我们每个人都得花时间去学习如何使用人工智能。
The other thing people need to, we all need to invest time to learn how to use this AI.
我经常看到一种模式：只要有人找到了一种特别高效使用AI的方法，他就再也停不下来了。
A pattern I've seen constantly is that as soon as anyone finds a very productive way of using AI, they cannot stop using it.
所以这周我们启动了一个叫‘AI高效一小时’的计划，目标是让每一位瑞银员工每周都能抽出一小时，通过实际操作来学习AI。
So one initiative we started this week actually is what we call AI Power Hour, whereby we really want every UBS employee to be able to invest one hour a day a week to learn about AI by doing.
我们还准备了不少资源，包括一些智能助手，能教你如何使用AI。
And we have a number of resources, including agents that can teach you how to use AI.
花时间学习AI，对公司和个人都有好处，因为大家能学到东西，但关键还是得有主动想学的意愿。
So the idea of investing time to learn about AI is something which is very good for the company, as well as for each individual in the company, because they learn, but it's really about the willingness of people to want it to learn.
这现在已经不容商量了，我觉得。
It's non negotiable anymore, I guess.
没错。
Right.
我再问你最后一个问题，还是回到你当教授时的经历。
Let me ask you one last question again, back to your background as a professor.
是的。
Yeah.
自我迭代提升。
Recursive self improvement.
这让你感到害怕吗？
Does it scare you?
这让你感到兴奋吗？
Does it excite you?
你是怎么看的？
Where do you land on it?
我觉得两者都有，既让人兴奋，也让人担心。
Look, it's both, I would say, exciting and scaring.
我相信，只要在安全机制、控制和验证上足够重视，兴奋点就会压过规模扩张带来的顾虑。
I do believe that given the great focus around guardrails, controls, validation, The exciting part is wins over the scaling part.
没错。
Right.
但我完全同意，保持谨慎、清楚地意识到局限和风险，是至关重要的。
But I totally agree that being remaining careful and cautious, also aware of the limitation and the risk is absolutely critical.
很好。
Great.
Daniele，非常感谢你参与这次对话。
Well, Daniele, I wanna thank you for joining
我们。
us.
非常感谢。
Thank you very much.
这位是来自瑞银的首席人工智能官丹尼尔·马加岑尼，正在TITV节目中。
That is Daniele Magazzeni, the Chief AI Officer at UBS here on TITV.
我身边这位是瑞银人工智能与软件研究主管卡尔·凯斯特德。
I'm here with Karl Keirstead, Head of AI and Software Research at UBS.
卡尔，非常感谢你邀请我们做客。
Karl, thank you so much for having us.
不客气，欢迎你们来。
Yeah, thank you coming.
卡尔，我想先聊聊整体的软件生态。
Karl, I want to start with the software landscape, broadly speaking.
今天早上我看了我平时跟踪的一组大约80家软件公司。
I was looking at a basket of about 80 software companies that I track regularly this morning.
过去一年，这个指数下跌了大约26%。
That index is down about 26% in the last year.
不过最近三个月又涨了26%。
It's up 26% in the last three months, though.
是的。
Yeah.
所以基于这个情况，你给我们做个整体的分析吧。
And so with that in mind, give us the overview here.
目前软件行业到底处于什么状态？
Where is the software sector at right now?
我们还在经历SaaS的寒冬吗？
Are we still amidst the SaaS apocalypse?
具体情况是怎样的？
What's the story?
确实还在。
We are.
我觉得，至少在SaaS或应用软件股票方面，投资者情绪依然非常低迷。
I would say investor sentiment on at least the SaaS or application software stocks remains very depressed.
这些股票最近有小幅反弹，从四月底的低点涨了大约20%左右。
The stocks have had a bit of a rebound, I'd say 20 ish percent off the bottoms in late April.
但我认为这主要是因为半导体板块热度下降，资金整体转向了更防御性、更便宜的股票。
But I think that's primarily a function of a fade in the semis trade and a broader portfolio rotation into defensive cheaper stocks.
我不觉得这是投资者发现了应用软件领域基本面出现明显改善的信号。
I don't think it's because investors are picking up evidence of a fundamental improvement in the application software space.
所以，我们团队的看法是，既然应用软件的环境依然严峻，采购部门都在尽量控制开支。
So at least the view of our team is that given that the environment for application software still remains tough, procurement officers are trying to limit their spend.
我有点担心，接下来SaaS领域的财报可能会表现疲软，那波反弹也可能就此消退。
I worry a little bit that the upcoming results across the SaaS space might be a little bit soft, and that rally could fade.
为什么会这样？
Why is that?
说实话，我觉得部分原因跟人工智能有关，就拿今天早上提前发布的IBM业绩来说，情况就是这样。
Well, I think it's partly a function of AI, to be honest with you, where even in the IBM results that were preannounced this morning Right.
我们看到，财富500强企业对AI的投入大幅增加，已经开始挤压其他支出。
We're seeing pretty strong evidence that rising AI spend inside Fortune 500 enterprises is beginning to crowd out spend.
挤占了传统软件的预算。
It's crowding out traditional software spend.
也挤占了IT服务的开支。
It's crowding out IT services spend.
最近Accenture和印度几家公司的财报就清楚反映了这一点。
You can see that in the results of Accenture recently in the Indian firms.
再加上今天早些时候IBM的预先公告，非AI硬件的支出也开始被挤压。
And as per the IBM preannouncement this morning, it's beginning to crowd out non AI hardware spend.
这已经是一个非常突出的趋势。
So this is a fairly dominant theme.
所以我认为这应该是最主要的因素。
So I think that's probably the main one.
其次，很多传统的SaaS公司所服务的市场现在都已趋于成熟。
And then I think the secondary one is that a lot of the traditional SaaS firms have fairly mature end markets now.
不过，软件领域里还是有一些细分板块表现得不错。
Now, there are pockets of the software space that are doing relatively well.
就像你都知道的，Snowflake、Datadog 这些公司。
So as you know, Snowflake, Datadog, those things.
任何底层技术似乎都挺有前景，我们这档节目也聊过这个观点：如果你能自己搭建后端，再用 AI 做出应用层，这似乎成了大家普遍在走的路，就是
Anything back end seems to, you know, we've been talking about this on the show, this idea that if you can build the back end yourself and create your own application layer using AI, that seems to be the formula that everyone is going for, which
换句话说，就是技术栈的底层部分。
is Or put good another way, the bottom of the tech stack.
没错。
Right.
基础设施、数据和网络安全这几个领域表现都还不错。
Infrastructure, data, cybersecurity is doing fairly well.
而上面的应用层，则一直不太顺利。
The layer on top, the application layer has struggled.
这其实也是我们团队过去一年的核心观点：看好数据和安全，对应用层保持谨慎。
And fundamentally, that's been the call of our team for the last year, to summarize long in for data security, cautious apps.
展开剩余字幕（还有 304 条）
那投资者现在到底在找什么呢？
So what are investors looking for then?
只是看增长率是不是在上升吗？
Is it just growth rates going up?
这是他们最看重的指标吗？
Is that the number one priority?
是的，当我问那些纯多头的大型投资者和对冲基金时，他们给我的回答都很一致：你们到底在看什么？
Yeah, least when I talk to large long only investors, hedge funds, and I ask them the question you just asked me, What do you look?
他们一致说，我们需要看到增长曲线出现拐点。
The consistent answer is we need to see a bending of the growth curve.
所以，大家都在等这个信号。
So that's what we're all waiting for.
我想跟你聊聊几个具体的名字。
I wanna ask you about a couple specific names.
微软的股价在过去一年里跌了大约20%。
Microsoft, stock is down about 20% in the past year.
上周Xbox部门确实进行了裁员，这一点我们都知道。
We obviously had the layoffs last week in the Xbox unit.
那投资者对这件事最关心的是什么？
What are investors looking for there with that story?
是的，这只股票挺有意思的。
Yeah, this is an interesting stock.
目前半导体板块和超大规模云服务商之间的表现差异可以说是前所未有的。
We're probably seeing the widest divergence between the long semis trade and the hyperscalers.
两者都跟AI相关，但一个已经创了新高，而你提到的微软、亚马逊、谷歌、甲骨文和Core Weave这些公司，表现普遍落后。
Both are exposed to AI, but one is at all time highs, and as you point out, Microsoft, Amazon, Google, Oracle, Core Weave have generally lagged.
这让我特别感兴趣。
So that's interesting to me.
总的来说，我对云服务商的前景还是持乐观态度的。
So I've generally got a constructive call on the hyperscaler level.
微软很快就要发布财报了，这个问题问得正好。
Now Microsoft is going to report soon, so it's a good question.
我们需要看到什么，才能扭转这种说法，缩小超大规模云服务商股票和半导体股票之间的差距？
What do we need to see to bend that narrative and to close the gap between the hyperscaler stocks and the semi stocks?
我觉得你可能需要看到更多资本支出预测的下调修订。
I think you probably need to see more reserved CapEx estimate revisions.
我觉得微软股价里已经隐含了一种担忧：当公司发布财报、透露资本支出情况时，实际数字会远超市场预期。
I think what's embedded in Microsoft shares is a concern that when the company reports and gives color on CapEx, that it'll massively exceed street estimates.
还会更多。
Even more.
比他们目前给出的还要多。
Even more than they've provided so far.
所以这是一个担忧。
So that's one concern.
我觉得另一个需要解决的问题是，市场普遍担心OpenAI和Anthropic正在切入知识型软件领域，而这个领域的巨头正是微软。
I think the other issue that needs to get resolved is that there's this background fear that OpenAI and Anthropic are going after the knowledge work software space, and the giant in the knowledge work software space is Microsoft.
没错。
Right.
那微软要怎么在Office Copilot这个业务上保持竞争力呢？
So how can Microsoft remain competitive in the Office co pilot franchise?
这个目前还不好说。
That's That's a a TBD.
这两个问题可能是最关键的。
Those are probably two of the most critical questions.
现在大家对微软的看法确实有点变化，我不禁想，到底要怎么做才能扭转这种印象？就像你说的，OpenAI和Anthropic给人的感觉是瞄准了同样的客户群。
It does feel like the perception around Microsoft, mean, I wonder what it takes to bring that back, because like you said, OpenAI Anthropic, there's a perception that they're targeting the same customer base.
现在Copilot已经掺和进来了。
You've got Copilot in the mix here.
大家也在担心它的采用速度到底有多快。
There are questions around how quickly it's being adopted.
他们还在调整新的定价模式。
They're fiddling with new pricing models as well.
还有游戏业务摆在那儿。
You've got the gaming business.
我不知道。
I don't know.
你觉得微软会把游戏业务拆分出去吗？
Do you think the gaming business gets spun off?
我的意思是，你怎么看？
I mean, what do you think here?
我不觉得会拆分，但说实话，这对股价影响不大。我觉得市场对游戏部门最近裁员的解读其实是正面的，认为这是在合理控制人力成本，有助于维持华尔街的好感。
I don't think it does, but I don't think it matters for the stock, to be honest I with think the street interpretation of the recent layoffs in the gaming space was constructive, good headcount management to That's maintain the Wall Street perception.
不过，微软和OpenAI之间的竞争问题确实很值得探讨。
But yeah, this question of Microsoft's competitiveness with OpenAI in particular is a fascinating one.
很明显，微软持有OpenAI百分之二十五到三十的股份。
Obviously, Microsoft owns 25 to 30% of OpenAI.
是的。
Yeah.
微软和甲骨文是OpenAI最大的算力供应商，但未来五年，微软在生产力软件领域的竞争力显然会越来越强。
They're, along with Oracle, the largest provider of compute for OpenAI, and yet it seems fairly evident that over the next five years, they're going to become increasingly competitive in the productivity software space.
所以我觉得市场得相信微软能赢得这场较量。
So I think the street needs to believe that Microsoft can win that battle.
我想再回过头聊聊你刚才说的芯片股和超大规模云服务商股票之间的差异。
I want to go back to the delta that you were talking about here between the chip stocks and the hyperscaler stocks.
如果我没理解错的话，你说现在芯片股特别受追捧。
So if I understood you correctly, you said the chip stocks are getting a lot of love right now.
但云服务商股票却不是这样。
Hyperscalers are not.
你觉得现在造成这种差距的核心原因是什么？
What's the core reason for that delta right now, do you think?
我觉得目前半导体股票被普遍视为简单的多头机会。
Well, I think the semi stocks are believed to be easy longs right now.
很多公司的盈利预期都在上调。
There's a lot of upward estimate revisions.
你知道的，DRAM和内存的成本上涨，带来了明显的收入增长空间。
There's cost inflation, you're aware, with DRAM and memory that's creating significant revenue upside.
但说到云服务商的股票，一般来说，现金流预期正在下降，因为资本支出在上升，而且竞争也更加激烈。
And yet on the hyperscaler stocks, generally speaking, cash flow estimates are going down as CapEx goes up and there's heightened competition.
所以我觉得这在很大程度上解释了原因。
So I think that probably explains a lot of it.
因此我们的看法是，在未来几个月甚至几个季度里，这种差距可能会逐渐缩小。
So our call is that we might, over the course of the next several months, several quarters, see that gap start to close a little bit.
你读过萨提·纳德拉在X上的那篇帖子吗？
Did you read Sethi Nadella's post I
当然读了。
certainly did.
是在X上吗？
On X?
对。
Mhmm.
你对那篇文章有什么看法？
What was your reaction to it?
是的，萨提亚已经发了两篇帖子。
Yeah, so there's been two of Satya's posts.
这两篇内容很相似。
Both are similar.
我觉得微软想表达的是，要让人工智能真正成功，光有世界一流的前沿模型还不够，还需要一个更广泛的前沿生态系统。
I think what Microsoft is getting at is to make AI successful, you need not only access to world class frontier models, but you need that broader, call it frontier ecosystem.
你需要数据，还需要完整的安全和治理体系。
You need the data, all the security and all the governance.
所以我认为萨提亚说得对，企业要想在AI领域成功，确实需要这些东西，而微软正好能提供这些。
So I think Satya is correct to point out that that's what enterprises need to be successful in AI and Microsoft can bring that to the table.
有些投资者感到困惑的地方在于，似乎微软和前沿实验室之间正在酝酿一点矛盾。
I think the confusing part for some investors is this notion of sort of creating a little it sounds like a little bit of conflict is brewing between Microsoft and the Frontier Labs.
对。
Right.
但这不只是微软的问题。
Well, it's not just Microsoft.
我的意思是，比如说
I mean, like,
Palantir的罗斯
Ross of Palantir
Salesforce那边也有动作。
has Salesforce.
也看到一些迹象表明
Seen something to
我的意思是，马克·贝尼奥夫虽然没点名，但很明显。
I mean, Mark Benioff, he didn't call them out, but Yeah.
他发过一篇关于零数据留存的文章。
He had that post about zero data retention.
是的。
Yeah.
所以现在确实能感觉到一些变化正在发生。
And so it it does feel like there is there's movement right now.
现在的情况就像是所有人联合起来对抗AI实验室，至少这些CEO们是这么在塑造这个论调的。
It's everyone against the AI labs, or at least that's the way these CEOs are trying to position the argument.
你觉得AI实验室现在面临信任危机吗？
Do you think that the AI labs have a trust issue right now?
我不觉得它们的信任问题有某些软件公司CEO们说的那么严重。
I don't think it has they have as big a trust issue as perhaps some of these software CEO companies are suggesting.
作为我们研究的一部分，我们经常跟企业IT高管交流。
I'll admit, as part of our research process, we are talking to enterprise IT executives all the time.
我敢说， Fortune 500企业的IT高管中，真正担心前沿模型会窃取公司机密的，一只手都数得过来。
I can probably count on a couple of fingers the number of times a Fortune 500 enterprise IT executive has expressed a worry that the frontier models are basically going to ingest their corporate IP and This diminish their
塞思刚才的意思是，你如果不把数据给它，就别指望能从中获得价值。
is what Seth just said, is that you can't get usefulness out of it unless you
给它。
give it.
但现实里，这种担忧远没有博客里说得那么普遍。
Don't hear that as much as the blog would suggest.
有意思。
Interesting.
所以从更大的角度看，未来五年可能会出现显著的重叠。
So I think bigger picture, what might be happening is over the next five years, we have significant overlap developing.
前沿实验室可能会面临更激烈的竞争，因为很多公司都在转向更便宜、更小的模型。
The Frontier Labs are probably facing a more competitive model market, as a lot of companies essentially down tier to cheaper, smaller models.
他们的应对方式很可能是向上游软件领域整合，而那里正是许多软件公司的主战场。
And their reaction is going to be to vertically integrate up into the software space where a lot of these software companies have their homes.
在未来几年，前沿实验室和现有软件巨头之间的重叠会越来越明显。
So we have a significant overlap between the Frontier Labs and incumbent software firms coming in the next several years.
所以，看到双方开始出现摩擦，我一点都不意外。
So it's not shocking to me that you get some measure of conflict beginning to brew.
既然他们不像萨提亚说的那样，频繁担心数据控制权的问题，那他们到底在担心什么？
So if they're not talking as much as Satya points out about giving up control over their data, what are the concerns that they're talking about?
他们是不是更在意模型的使用成本？
Are they concerned about cost to the models, I presume?
很多
A lot of
主要是成本问题，所以现在技术圈里‘令牌优化’成了热门话题。我注意到，今年年初大力投入AI的企业，现在发现他们的计算和令牌开销远远超出了预算。
it is cost, and that's why we're seeing this notion of token optimization, which is becoming dominant subject in tech circles, and I would say the concern is that enterprises that have leaned in aggressively year to date around AI are finding their compute token costs far exceeding what they budgeted for at the beginning of the year.
没错。
Right.
他们开始削减使用量了。
And they're starting to throttle it back.
科技投资者担心的是，如果企业这么削减，那依赖OpenAI和Anthropic的整个科技生态又该怎么应对？
And the concern that tech investors have is, as they do so, how can that be good for any of the tech ecosystem that is dependent upon OpenAI and Anthropic?
换句话说，如果你减少令牌使用，转而选择更便宜的模型，这对前沿实验室未必是好事。
In other words, if you pull back on your token consumption and you potentially down tier to cheaper models, that might not be good for the Frontier Labs.
如果前沿实验室受影响，那整个科技行业都会遭殃。
And if it's not good for the Frontier Labs, it's negative for most of the tech trade.
现在市场正在消化这个现实。
That's what the Street is absorbing right now.
换句话说，我从你这里听到的是，也许这个思路有点反其道而行之：如果萨提亚建议用开源来控制你的知识产权和数据的话。
So in other words, what I'm hearing from you is, and maybe this is a bit backwards, but if Satya is saying use open source, you know, as a way of controlling your IP, your data.
或者甚至用微软自己的模型。
Or maybe even Microsoft's own models.
对，就是微软自己的模型。
Right, Microsoft's own models.
我的意思是，核心观点是别那么依赖前沿模型。
I mean, the idea here is that don't use the frontier models as much.
顺便说一句，其实大家也负担不起像预期那样频繁使用这些模型，但反过来，这可能会伤到云服务商，因为如果前沿模型没人用得起那么多算力，最终还是会反噬他们，对吧？
By the way, people can't really afford to use it as much as they might hope, but that in turn could end up hurting the hyperscalers because of the fact that the frontier models, if they can't afford all the compute, then that comes back to bite them, no?
确实有可能。
It could.
这确实是一个担忧。
That's one concern.
我觉得在讨论之前，得先说明一点：云服务商的股价现在压力挺大的。
I would argue that we preface this conversation with the notion that the hyperscaler stocks are feeling a little bit heavy.
是的
Yeah.
这可能确实是一个原因。
This might in fact be one reason.
换句话说，如果企业从像 Anthropic Cloud Opus 4.8 这样的高端前沿模型转向开源的中国模型，比如。
In other words, if there's a down tiering away from the premium frontier models like Anthropic Cloud Opus 4.8 to open source Chinese models, for instance.
这对前沿实验室不利，对微软、甲骨文以及整个基础设施层也不利。
That's not great for the frontier labs, and therefore that's not great for Microsoft, Oracle, and the whole infrastructure layer.
但这里也有一个抵消因素，那就是所谓的杰文斯悖论——如果 AI 成本降低，比如转向更便宜的模型，或者当 OpenAI 和 Anthropic 推出基于新一代英伟达芯片、更高效利用 token 的新模型时，AI 的成本下降，企业如瑞银就会更积极地投入 AI，这样一来，对云服务商反而是好事。
But there's an offset, and that is the hopeful idea of Jevon's paradox, which is that if you make AI cheaper, which occurs when you down tier to a cheaper model, or when OpenAI and Anthropic launch their next models based on next generation Nvidia chips, and they are much more token efficient, that you're lowering the cost of AI, and by doing so, enterprises like UBS will lean in even more to AI, and if they do that, that's good for the hyperscalers.
对，没错。
Right, right.
我们再快速过几个其他公司。
Let's go through a couple of other names quickly.
比如甲骨文，最近几个月，大家一直担心他们的数据中心建设延迟。
So Oracle, look, last couple months, there have been a lot of concerns about the data center delays.
这现在还是推动股价的主要因素吗？
Is that still a driving story for the stock?
你最近看到什么情况？
What are you seeing?
我觉得这部分故事已经淡下去了。
I'd say that part of the story has faded.
你应该知道，甲骨文最近很积极地反驳了关于延迟的说法，上一次财报电话会上，他们花了大量时间详细说明了五个主要数据中心的建设进度，明确列出每个项目预计上线的时间，就是为了澄清这些质疑。
I think you're aware Oracle has come out pretty aggressively pushing back on the notion that there are delays, and they spent quite a bit of time on the last earnings call going through each of their five major data center builds and laying out exactly when they are expected to go live in an effort to push back against that.
我觉得更大的问题是，市场情绪一直在反复波动。
I think the bigger issue is that the street has gone through waves.
你应该也注意到，过去几年里，有时候资本支出增加是好事，因为它代表需求强劲。
I think you're aware in the last couple of years where sometimes more CapEx is good because it's a positive demand signal.
但有时候资本支出增加反而是坏事，因为会压低自由现金流，引发对过度建设的担忧。
Sometimes more CapEx is bad because it depresses free cash flow and there's concerns around overbuilding.
现在我们又回到了‘资本支出增加是坏事’的阶段。
We are back into one of those more CapEx is bad period.
换句话说，我觉得市场又开始担心这些AI资本支出的回报率了，这不仅拖累了甲骨文的股价，也影响了微软。
So to put it in a different way, I think the Street's back to being worried about what the return on all this AI CapEx is, and I think that's weighing on Oracle shares in addition to Microsoft.
没错。
Right.
我想问问你对Salesforce几周前收购Finn这件事怎么看。
I want to ask you about the Salesforce acquisition of Finn that we just saw a couple weeks ago.
你认为还有哪些领域的AI公司可能会成为收购目标？
What other pockets of AI companies do you expect to see acquisitions in?
你们所关注的那些大型软件公司，除了自己研发，还会在哪些领域选择收购而不是自建？
Where else will the big software companies that you cover, where else will they look to buy things rather than build?
我觉得他们其实应该多买一些。
I think they probably should buy more.
如果我是家老牌软件公司，正面临AI这种代际技术变革，我会通过收购来加快步伐。
Think if I'm an incumbent software firm and I'm facing a generational technology shift in AI, I'm going to move faster by acquiring more.
所以我非常赞赏这些瞄准未来趋势的收购行动。
So I'm actually applauding these efforts to acquire where the puck is going.
但你是觉得未来会更多看到应用层的收购，还是底层AI基础设施的收购？
But are expecting more application layer acquisitions or more of the underlying AI infrastructure acquisitions?
我的意思是，现在哪些领域最热门，
I mean, what are the hot pockets here that
我觉得应用层会多得多。
you'd I'd say it'll be far more at the apps layer.
因为很多现有的软件公司本身就是做应用软件的，面临增长放缓的压力，必须做出应对，所以这些
So I think just because more incumbent software firms are application software companies that are facing the prospect of decelerating growth, they need to react, and so And these
他们收购的都是横向产品，之后可以向客户展示。
are horizontal product acquisitions that they can then point to their customers.
要么是横向的，要么在某些情况下是垂直领域的原生AI公司，事实上，接下来两天我们舞台上就会有不少这样的公司。
Either horizontal or in some cases vertical AI native companies, and we frankly have a lot of them on stage in the next two days.
我再问你一个问题，关于企业软件公司常提到的客户粘性——他们说，别担心SaaS会崩溃，因为我们有成千上万的客户。
Let me ask you one more question about the stickiness that enterprise software companies like to point to, the idea that, hey, the SaaS apocalypse is not a real concern because of the fact that we have tens of thousands of customers.
要彻底替换掉一个系统，换别的东西，真的非常困难。
It's very difficult for us to rip out a system entirely and swap to something else.
这个说法在当前市场上站得住脚吗？
How is that argument holding up in the market right now?
站不住。
It's not.
我觉得现在的情况是，
I would say the
所以人们现在愿意把用惯的软件彻底换掉——
So people are willing to tear out softwares that they have-
从今年一月到四月，应用软件股票的定价已经反映出一种认知转变：过去那种软件粘性强、难以替代的说法已经不成立了。
I'd say the perception that got priced in application software stocks January through April is that that historical argument of durability stickiness is flawed.
之所以不成立，是因为二十年来第一次，客户能够利用不断进步的AI模型，自己搭建出替代传统软件的方案。
It's flawed because for the first time in twenty years, customers now can harness these ever improving AI models to custom build alternatives to incumbent software.
当然，投资者并没有过度推演这个观点，觉得像瑞银这样的公司很快就会用Anthropic的Claude来复刻他们运行全行的SAP系统。
Now, software investors are not extrapolating too much with that argument, thinking that a firm like UBS anytime soon is going to harness Anthropic Claude to replicate our SAP system that runs the bank.
没有哪个认真的投资者真会这么想。
No investor, serious investor, really thinks that's going to happen.
所以从这个角度看，确实还是有粘性的。
So in that sense, yes, there's stickiness.
但我认为，当前股价中反映的更现实的看空观点是：很多辅助性应用和推动增长的附加服务，可能会被定制化的AI软件取代。
But I think the more realistic bear case that's been priced into these stocks is that a lot of the ancillary applications, the upsells that drive growth, that might get replicated by custom built AI software.
没错，这正是我想理清的：表面上看，这种担忧似乎非常强烈，但实际情况是，这种替代还没真正发生。
Well, and that's what I'm trying to sort of figure out is the perception is there, is that the reality of what's happening on the ground sounds like it's the perception is as strong as ever, but the reality is it's not happening just yet.
没错。
Correct.
我觉得这两种观察都是对的。
I'd say both observations are accurate.
这种潜在威胁如今确实已经被充分反映在股价里了。
The threat that it may happen is absolutely getting priced into the stocks today.
现在真的在发生吗？
Is it happening today?
还没有。
No.
但我们确实会和财富500强企业讨论，未来两三年世界可能会变成什么样。
But we do talk to Fortune five and enterprises about what the world might look like in two to three years' time.
我可以肯定地告诉你，因为我已经和很多人聊过这个话题，他们明确表示：考虑到这些AI模型的性能提升，以及自建替代方案的可能性，他们希望未来三年内对XYZ软件公司的支出减少30%。
And I can assure you, because I've had many of these conversations, they absolutely are articulating a view that given the performance improvements in these AI models and the ability to custom build alternatives, that they would like their spending with software company XYZ to be down 30% over the next three years.
这种说法我确实听到了。
I do hear this.
那我们回到刚才的话题，重新看看估值倍数，看看现在股价的位置。
And so let's bring it full circle now, back to the multiples, back to where the stocks are trading at right now.
一年后，软件公司股价会涨吗？
A year from now, are software companies, are they trading higher?
会有一些反弹吗？
Is there a little more of a recovery?
你觉得股价会维持现状吗？
Do you expect them to stay the same?
你怎么看？
What do you see?
我更倾向于认为市盈率会保持稳定。
I'm a little bit more in the multiple staying the same camp.
明白了。
Okay.
至少在我的观察范围内，
So least in my
所以先扩大收入，其他的事以后再解决。
So grow the top line and figure the rest out.
是的。
Yeah.
在我所覆盖的大型应用软件公司里，根本没有一只股票值得买入。
At least in my large cap application software coverage universe, I don't have a single buy.
所以我相信，未来十二个月会是一段艰难的时期。
So I'm in the camp that it's going to be a rocky ride for the next twelve months.
而且，我们团队在基础设施和数据安全领域更倾向于长期看好，这和我们之前聊过的理由一致。
And as a team, we have a much stronger bias long at the infrastructure data security levels for the reasons you and I talked about earlier.
对。
Right.
卡尔，非常感谢你邀请我们来做客。
Well, Karl, I wanna thank you so much for having us.
这位是来自瑞银的卡尔·基尔斯特德，正在TITV节目上。
That is Karl Keirstead from UBS here on TITV.
我旁边这位是Locus Robotics的首席财务官达斯汀·彼得森。
I'm here with Dustin Pederson, CFO of Locus Robotics.
达斯汀，欢迎来到TITV。
Dustin, welcome to TITV.
很高兴能来。
It's great to
你太客气了。
have you here.
谢谢你们邀请我。
Thank you for having me.
Locus Robotics，能给我们讲讲你们公司是做什么的吗？
Locus Robotics, tell us what the company does.
是的。
Yep.
我们是一家专注于自动化仓储和配送中心的自主移动机器人与软件公司。
So we're a we're an autonomous mobile robotics and software company focused on automating fulfillment and distribution warehouses.
简单来说，就是用机器人代替人工在仓库里搬运货物，让员工能更高效地拣选商品。
So effectively making people more efficient by using robots to do a lot of the movement around a warehouse so people can pick more efficiently and effectively.
我们最近还推出了一款新方案，连拣选都能自动完成，核心目标就是全面提升仓储运营效率。
And we just launched a new solution that also does the picking as well, so it's all about making these operations more efficient.
那这个系统……它是一个平台吗？
So this is the it's is it just a is it a platform?
它有机械臂或者夹爪吗？
Does it have arms, claws?
它到底长什么样？
What what does it what does it look like?
是的
Yep.
它是一种自主移动机器人，能在仓库里自由穿梭，完全不需要地面上的导轨或标记，想去哪儿就去哪儿。
So it's a it's a it's a autonomous vehicle that can drive around the warehouse completely, you know, free of any grids or tracks in the floor that just goes to where it needs to.
你可以想象成一个装了轮子的机器人，自己在仓库里到处走动。
So think of a a robot on wheels that can drive around.
我们有一种型号，需要人工通过平板电脑操作来拣货；还有一种升级版，机器人自带机械臂，能自己取出货箱、拣选商品，然后放进目标货篮，全程自动完成拣选。
We have a certain version that in which the human actually touches the tablet and does the picking, and a different version where which actually has an arm on it so it can extract the tote, pick the item, and put it in a destination bin, and do the picking for itself.
你们是把产品卖给运营仓库的公司吗？
And you sell to companies operating warehouses?
对。
Is that Yep.
我们服务的主要是仓储和配送中心，大约一半的客户是第三方物流公司。
Fulfillment distribution centers, about half of our business is third party logistics companies.
也就是那些专门做仓储配送服务的大公司；大约三分之一是零售和电商企业，剩下的则是医疗和工业领域的客户。
So think of large companies that are focused on doing these operations for a living, and then about a third of our business is retailer and e commerce companies, and the remainder is healthcare and industrial companies.
这些公司都有仓库，货架上摆着各种货品，需要从货箱里取出来，运到发货区，或者从装卸区搬上货架。
Basically, all these companies have warehouses that have things on the shelves in bins that need to move be moved out of the bins to the shipping area or from the loading dock onto the shelf.
现在仓库里一共有多少台机器人？
How many how many robots do you have in warehouses today?
是的。
Yep.
目前我们已经在超过三百五十个仓库部署了一万五千多台机器人。
So we've got north of 15,000 robots in warehouses today across over three fifty facilities.
那你们的定价模式是怎样的？
And so how does pricing work?
一台机器人多少钱？
How much does a robot cost?
是的。
Yep.
我们采用的是订阅制模式，这在行业内算是比较特别的。
So we do we have a subscription model, so we're a little unique for the industry.
这在行业里并不常见。
So we That's not common.
对吧？
Right?
这确实挺特别的。
That's that's that's pretty
是的。
Yeah.
这一点还挺新颖的。
That's a little bit novel.
我们从一开始就这么做了，这样能让客户把成本算作运营支出，而不是资本支出，同时也更有灵活性。
We've doing it since day one, and what that allows customers to do is it allows for them to be operational expenses versus CapEx and for them to have flexibility.
我们通常签三年的订阅合同，客户可以按需在旺季增加机器人数量。
So we do three year deals in which customers subscribe to the robots, but they can also add additional bots for peak seasons.
他们只需另签一份订单，就能在几乎一夜之间扩大运力，淡季时再缩减回来。
They sign a separate order form, and they can effectively scale up their operations almost overnight to handle more volume and then reduce it back down.
所以我们来管理整个机器人车队，客户就能获得与仓库需求相匹配的技术能力，从而获得良好的投资回报。
So we effectively manage the fleet, and the customer gets what they need in terms of technology capacity to match the demand going through the warehouse and get a good return on their investment.
那这个订阅服务是怎么处理设备更新的呢？
So how does this subscription work then in terms of refreshing the fleet?
你们大概多久会给订阅的客户更新一批新机器人？
How often are you adding new robots to a subscription?
说实话，我也不太清楚一台机器人的使用寿命有多长。
You know, I I don't I don't know what the lifespan of a of a robot is.
这个机制具体是怎么运作的？
How does that work?
没错。
Yep.
这些机器人实际承受的机械磨损很小，客户订阅后，会定期收到软件更新，从而让机器人能应对新的任务场景，走位更灵活，性能持续提升。本质上，只要客户需要，就可以一直用同一台机器人，随时增加新设备；到了续约时，也可以随时减少数量。
So the robots actually have very little mechanical stress and strain, and what a customer gets when they when they subscribe to the robot is they get updates periodically with new software so they can handle new use cases, they can navigate more tightly and continue to improve over time, and effectively, they keep the same robots for as long as they need them, and they can add in any given point in time, And at renewal, they can always reduce.
如果业务量下降，客户可以减少机器人数量，这完全没问题，我们可以回收这些设备，翻新后再部署到其他仓库。
If their volumes are lower, they can reduce the number of robots, and that's fine because we can take those back, refurbish them, and and send them out to a different facility.
明白了。
Right.
我想和你聊聊当前机器人领域的一些更宏观的话题。
I wanna talk to you about some of the broader topics in robotics right now.
这周在节目里，我们一直在讨论机器人中的软件部分，还有目前面临的资金限制问题。
You know, on the show this week, we've been talking about, the software component to robotics, some of the capital constraints right now.
在你自己的研发和整个行业的发展中，你觉得最大的瓶颈在哪里？
Where are you seeing a bottleneck in terms of your own development and the development for the industry as a whole?
是的。
Yep.
就我们自身而言，我们刚推出了抓取机器人，这段时间我们花了很多精力收集实际场景中的抓取数据，比如让机械臂从货箱里抓取物品，再放到目标货箱里。
So in terms of our own development, we just launched our picking bot, so we've been doing a lot of work to gather real world data in terms of picking, so think of being able to grab an arm, being able to grab things out of a bin and move them to a destination bin.
这方面我们还在不断学习。
So that's an area where we're learning.
多年来，我们在导航和车队调度方面已经积累了丰富经验，比如机器人如何始终协同工作，如何在仓库里实现自然流畅的移动。
We've obviously, over the years, we've gotten a lot of experience in terms of navigation as well as fleet orchestration, so how do the robots work as a team all the time, every time, as well as navigate the facility really, really intuitively and and smoothly?
是的
Yeah.
这方面我们积累了很多经验，因为实际场景中的特殊情况太多了。
So that's an area where we've built up a lot of learnings because there's a lot of edge cases out there.
那训练呢？
What about training?
关于训练，我们既用从客户现场收集的数据，也用一些现成的模型和数据，这方面整个行业都面临挑战。
In terms of training, so we're using both the data that we collect from our customer sites, but we're also using off the shelf, we'll call it models and and data, and that's an area from an industry standpoint.
我觉得现在合成数据用得很多。
I think there's a lot of synthetic data.
模型确实不少，但说到物理AI和真实应用场景，要应对那些特殊情况，真实世界的数据是必不可少的。
There are a lot of models, but when it comes to physical AI, when has real world world applications, real world data is really, really needed to get those edge cases.
想想特斯拉或者Waymo这样的公司，如果不能解决所有特殊情况，后果会很严重。
If you think of, like, companies like Tesla or Waymo, obviously, there's huge implications to not solving for all the edge cases.
在仓储机器人领域，从安全角度看，风险没那么高，但依然很重要——毕竟，这跟通用大语言模型不一样，后者偶尔出现幻觉还能接受。
In warehouse robotics, it's a little little lower stakes from a safety perspective, but still very important relative to think of your average LLM where you can you can deal with a hallucination from time to time.
在机器人领域，系统必须能正常运行，而且得一直稳定运行，同时还得时刻考虑工人的安全。
In robotics, it just has to work, and it has to work all the time, and you have to also keep worker safety in mind.
但我们还是回到资本限制这个话题吧。
But let's go back to, I mean, capital constraints here.
一方面，问题是获取所需数据的成本是不是太高了？
Mean, on one hand, there's the question of is it too expensive to get the data you need?
另一方面，这些数据本身是否存在呢？
On the other hand, it's does the data even exist?
是的。
Yep.
对这个行业来说，这两个问题中哪个更严重？
Which of those two is a bigger problem for the sector
现在？
right now?
我觉得对这个行业而言，更大的问题其实是数据根本就不存在，说实话。
I think for the sector, it's probably does the big day exist, honestly.
因为有些东西其实可以通过模拟来解决。
Because, I mean, some of this stuff is I mean, you can simulate it.
你可以用世界模型，我觉得它们能找到办法进入真实世界。
You can World models, I think, find ways to go into the world
然后拍下各种数据，对吧？
and videotape stuff, right?
对。
Yep.
当然可以这么做，现在也有很多公司在这么做。但我觉得，过去一两年里，整个机器人生态系统的投资力度很大，资金相当充足。
You can certainly do that, and there's a lot of companies that are doing that, but I think because there's been a lot of investment in the broader, we'll call it robotics ecosystem, especially over the last, two, one to two years, there's lots of capital.
但归根结底，你还是得靠真实的机器人去执行任务，才能收集到真正的现实世界数据。
But at the end of the day, you need real live robots doing things to collect that real world data.
所以我觉得，更大的瓶颈在于数据是否真的存在。
So I I would say the the constraint is more in the existence
以及数据获取的难度。
of and the egg situation.
这就像是你需要机器人去收集数据，但又没法先造出机器人来
It's like you need the robot to get the data, but you can't make the robot in
根本就造不出来。
the first place.
你得让机器人真正部署在现实环境中才行。
And you need robots deployed in a real world environment Yeah.
总不能只在仓库角落或者受控的训练环境里吧？那种方式有点用，但真正高质量的数据，还得靠机器人在真实世界里跑才行。
Versus, you know, maybe in the back of a warehouse or or in a training that helps to some extent, but you really need it out in the wild in order to get real real good data.
那在机器人领域，基础模型的情况怎么样？
What what about the the the foundation models here in in robotics?
我们现在所处的AI世界里，已经有几个大型实验室成了主导力量。
I mean, we we live in this world here where in AI I mean, we've got these big labs that have become, the the you know, there's a couple big players.
明白了。
Okay?
它们掌握了这些模型。
And they have the models.
每个人都想拿到这些模型。
Everyone wants to get their their hands on them.
对。
Mhmm.
机器人领域也是同样的情况吗？
Is that same story playing out in robotics?
会不会出现几家大型机器人物理AI模型公司，最终成为人人都得向他们购买的巨头？
Are there a few big robotics, physical AI model companies that you suspect will become the giants that everyone will need to buy from?
是的。
Yeah.
有可能。
Potentially.
有几家公司在外面很活跃，融资很多，也经常上新闻，我们稍微了解一点。
There's a couple of companies that are out there that are very have raised a lot of capital that are very much in the news, and we've got a little bit of
我们应该关注哪些公司？
Which name should we be paying attention to?
一些做实体智能、熟练型AI的公司，现在挺热门的。
Physical intelligence, skilled AI, or a couple that are out there that are buzzy.
我对哪家更优没有特别独特的见解，但我们的思路是，未来要成为所有类型硬件的机器人大脑。
I don't have any unique insight into one versus the other, but they've got an approach where we'll be the robotics brains for every type hardware over time.
也许确实会这样，但还有些公司采取更精细的方式，专注于解决非常具体的客户问题，以极有限的范围切入，让商业模型成立，客户能明显感受到价值，自下而上地逐步推进——这正是我们的做法。
That may be the case, and then you've got other companies that have a little more nuanced approach to solving very discrete customer problems in a very finite way such that the business case works, the customer sees a lot of value, and kind of building from the bottoms up in that way, and that's been our approach.
那么，你认为机器人领域会继续维持这种少数巨头主导、其余分散的格局，还是说长期来看会变得更加碎片化？
So do you anticipate that robotics will stay on this trend of there being a couple big model companies and then the rest, or do you think it ends up more fragmented in in the long run?
是的。
Yeah.
我觉得机器人领域比纯软件的大语言模型更复杂，因为就像我之前说的，它必须和硬件紧密结合。
I think robotics is a it's it's a little bit harder than just pure software LLMs because, as I mentioned before, there's the integration with the hardware.
每种硬件、每个零部件都有所不同，而且还有一个关键点：系统必须安全可靠，必须始终稳定运行。
So every hardware all different components of hardware are little bit different, and then the the second aspect to it is it has to be safe, and it has to work all the time in an integrated fashion.
理想情况下，你希望有一个统一的‘机器人大脑’来控制一切，但你必须做到99.99%的时间都万无一失。
So, you know, ideally, you'd have some sort of, we'll call it, robotics brain that can control everything, but you have to be able to do it 99.99% of the time.
必须安全可靠。
It's gotta be safe.
不能出故障，得真正解决客户的问题，才有价值。
It can't get hung up, and it's gotta solve the customer problem for it to really have value.
和Claude或者OpenAI这样的公司相比，我们还有很长的路要走，他们那边出现点幻觉或者错误，用户还能接受。
So there's a long road ahead relative to, we'll call it, Claude or OpenAI, where you can deal with some we'll call it I'm gonna call them hallucinations, but some some errors and exceptions, and people are just you can deal with that.
但在现实世界里，这种问题你根本没法轻易放过。
Whereas in the real world, you it's a little bit you can't quite deal with the same exception.
所以是分散发展，还是巨头垄断？
So so fragmented or or Giants.
会是分散的，但关键问题是，这个过程要多久？
It'll be fragmented, but at some point in time, the question is is what's the time scale?
肯定会有巨头出现，但我说的是长远来看。
There will be giants, but I think that's we're talking Right.
可能是十年以后的事，不太可能是短短几年内。
Tens of years out, probably not single digit number of years.
对。
Right.
在你走之前，我问你几个简单的问题。
Let me ask you a couple quick questions before I let you go.
现在芯片的情况，我们之前聊过的内存芯片短缺，对你们的业务有影响吗？
The chips story right now, I mean, the memory chip shortage we've been talking about, is that affecting your business?
确实有一点影响。
It is affecting a little bit.
我们自己有服务器，所以服务器的交货周期和价格都涨了，因为它们里面用到了内存，我们这边确实感受到了一些，但还没到苹果那些大公司那么严重的程度。
We have on-site servers, so we've seen the lead times and the cost of servers go up because they have memory in them, and so we've certainly seen that a little bit, but not to the extent of probably folks like Apple and some of the other folks in the news.
好的，再快速问一下，机器人领域里开源和闭源模型的争论，情况是不是也一样？
Okay, and very quickly, I mean, source versus closed source models in robotics.
是的。
Yep.
同样的争论也在上演吗？
Is the same debate playing out?
你更偏向哪一种呢？
Do you prefer one versus the other?
对。
Yep.
目前我们还没形成明确的看法，但我觉得随着行业逐渐成熟，这个问题会越来越受关注。
Right now, we don't really have probably an opinion on that, but I think that that will be a debate that plays out more and more as the industry matures.
我觉得机器人领域整体上机会巨大，但相比人们日常使用的那些AI模型，它还处在非常早期的阶段。
I think this is robotics in general has a massive opportunity, but it's much more in the early days relative to, we'll call it, the AI models that people use in their day to day life.
咱们就等着看吧。
So we'll see how that plays out.
现阶段我们还没什么坚定的立场，不过几年后再看吧，到时候再聊。
Don't really have a strong opinion one way or another at this point in time, but I guess we'll see and talk to you in a couple years.
很好。
Great.
德斯汀，非常感谢你今天来参加访谈。
Well, Dustin, I wanna thank you for coming on.
这里是来自Locus Robotics的Dustin Pederson，正在TIT节目中。
That is Dustin Pederson from Locus Robotics here on TIT.
今天的节目就到这里了。
That does it for today's show.
感谢大家的收听，也特别感谢高盛今天在现场为我们提供支持。
I wanna thank you all for joining us, and I wanna thank UBS for having us here on-site at the event today.
明天周四上午10点太平洋时间，下午1点东部时间，我们照常播出。
We're gonna be back to our regularly scheduled programming tomorrow on Thursday at 10AM Pacific, 1PM eastern.
明天我会从我们在旧金山的演播室为大家带来节目。
I will be coming to you from our San Francisco Bureau.
非常感谢你们的收看。
Thank you so much for tuning in.
明天见。
We'll see you tomorrow.
暂时再见了。
Bye bye for now.
关于 Bayt 播客
Bayt 提供中文+原文双语音频和字幕，帮助你打破语言障碍，轻松听懂全球优质播客。