OpenAI 的秘密“大蒜”模型、C3 AI 的转型计划、机器人技术与华盛顿特区 | 2025年12月4日

欢迎各位来到Informations TITB。

Welcome everyone to the Informations TITB.

我叫阿卡什·帕什里特沙。

My name is Akash Pashritsha.

今天是12月4日，星期四。

It is Thursday, December 4.

今天我们为大家准备了精彩的节目。

We have got a great lineup for you today.

首先，我们将深入探讨关于网页发布者的新报告，他们正困惑于网站流量激增究竟是来自机器人还是人类。

First up, we are diving deep into a new into new reporting on web publishers' newfound confusion around whether or not heightened traffic on their websites is coming from bots or from humans.

接着，我们将关注一些财报新闻，包括与C3 AI新任首席执行官的对话，该公司正面临收入下降和今年股价大幅下跌的困境。

We'll then turn to some earnings news including a conversation with the new CEO of C3 AI as the company faces falling revenue and a stock that has taken a steep hit this year.

我们还将联系我们的AI记者，了解OpenAI如何通过其正在开发的代号为“大蒜”的系统来应对谷歌的新一代Gemini三模型。

We'll also check-in with our AI reporter who is tracking OpenAI's efforts to counter Google's new Gemini three model with its own in development system code named garlic.

我不知道为什么它被叫做大蒜。

I don't know why it's called garlic.

这将是我们要问她的第一个问题。

That's gonna be the first question we ask her.

最后但同样重要的是，我们将以机器人领域综述结束本期节目，剖析最重要的新闻及其对这一领域快速增长的意义。

And last but not least, we will wrap up with a robotics roundup, breaking down the biggest headlines and news and what they mean for the sector's rapid growth.

这将是一场盛大的节目，所以我们马上开始。

It's gonna be a big show, so let's get right on into things.

聊天机器人正在给运营网站的公司带来混乱，因为这些公司无法确定他们看到的新流量在某些情况下是来自真实用户还是机器人。

Chatbots are wreaking havoc for companies operating websites because the companies behind those sites can't figure out whether or not this new traffic they're seeing is in some cases coming from real people or from bots.

现在邀请凯瑟琳·佩洛夫与我们讨论这一问题。

Joining me now to discuss this is Katherine Perloff.

凯瑟琳，欢迎来到节目。

Katherine, welcome to the show.

很高兴见到你。

It's great to see you.

嗨。

Hi.

那么这里到底发生了什么？

So what's going on here?

你是怎么发现这个问题的？流量激增，但人们却不知道来源在哪里？

How how did you even come across this issue of there being heightened traffic and people not knowing where it's coming from?

是的。

Yeah.

实际上，萨里，我们会深入讨论这个话题，因为我之前和一位出版商聊过，他说他们很难屏蔽Grok，因为无法分辨Grok——也就是xAI聊天机器人——到底是机器人还是真人，这让他们很难采取屏蔽措施。

Well, and actually, Sari, we'll get into this because I was talking to a publisher who was saying that they were having trouble blocking Grok, because they couldn't tell, you know, Grok being the x AI chatbot, because they couldn't tell whether it was a a bot or not, and it made it hard to block.

当我进一步调查这个问题时，我发现这其实是一个相当普遍的现象。

And then when I dug more into the issue, I saw that, this is a pretty widespread problem.

你知道，网络爬取和网站机器人早已存在，但在人工智能时代，这个问题变得尤为重要，因为网站方觉得，AI公司未经许可就拿走他们的内容，甚至用机器人取代了广告商愿意付费接触的访客。

You know, scraping has been going on for a while, crawlers, bots to websites, but, it's taken on sort of a new significance for websites in the age of AI when they feel, you know, that the AI companies are taking their content without paying for it or, you know, replacing the visitors that advertisers are paying them to reach.

嗯。

Mhmm.

因此，随着AI聊天机器人越来越流行，它们对网站——尤其是网络出版商——构成的威胁日益凸显，屏蔽机器人已成为网站的更高优先级任务。

So blocking bots has become a bigger priority for websites as these, AI chatbots become more popular and the sort of threat they pose to websites, especially web publishers.

但在某些

But in some

这些

of these

零售商变得更为重要。

retailers have become more important.

但实际阻止机器人是一个挑战。

But the actual blocking of the bots is sort of a challenge.

所以最终目标是阻止机器人。

So so the end goal is to block the bots.

但要做到这一点，我们首先得弄清楚访问者是人类还是机器人。

But to do that, we have to figure out if it's a human or a bot in the first place.

为什么很难分辨哪个是人类，哪个是机器人？

Why is it so hard to to to detect which is which?

是的。

Yeah.

但有一个注意事项是，并不是所有网站都希望阻止所有机器人，尤其是随着代理的兴起。

And one caveat is, you know, not all websites wanna block all bots, especially with the rise of, like, agents.

你知道，本质上，

You know, you essentially,

如果

if

如果你是一个电商网站，你可能希望代理访问你的网站，让用户完成购买。

you're an ecommerce site, you might want the agent to come to your site and let people make a purchase.

但确实很难，因为这在很大程度上取决于发送机器人公司的声明以及它们如何展示自己的机器人。

But, yeah, it's so hard because, you know, a lot of this kind of depends on what the company sending the bot, says and how they present their bot.

比如，OpenAI，有人告诉我，他们在机器人方面相当透明，会主动告知网站：

You know, OpenAI, for example, people told me was pretty good about, being pretty transparent about their bots and, you know, telling, websites, hey.

我来你的网站是为了抓取数据。

I'm coming to your site to scrape.

我来你的网站是因为用户在搜索中提出了一个问题。

I'm coming to your site because a user asked a question in search.

我带着我的代理访问你的网站。

I'm coming to your site with my agent.

你知道，他们还挺透明的，但并不是所有AI公司都这样。

You know, they're they're pretty, transparent, but not all AI companies are.

我认为整个代理问题让情况变得复杂，因为如果是一个人类让AI去做某事，那这算人类还是机器人？

I think the whole agentic, issue makes it complicated because, like, you know, if it's a human asking an AI to do something, does that make it a human or a bot?

但到底算人类访问者还是机器人访问者呢？

But, like, is it a human visitor or a bot visitor?

但没错，我认为很多AI公司也知道网站不希望他们出现。

But, yeah, I think also a lot of, AI companies also know that websites don't want them there.

所以如果他们的行为更像人类访问者，就不会被检测出来。

So if they look more like a human visit visitor, then they will not be detected.

他们不会被屏蔽，从而可以获取所需内容，用于训练他们的模型。

They won't be blocked, and then they can get the content they want, which can inform their models.

所以

So

那么，现在有哪些公司正在将此转化为他们的机会呢？

So who are the companies now that are making this an opportunity for them?

你提到了一些初创公司和几家大型科技公司。

You outlined a couple of the startups and a couple of the bigger tech companies.

现在帮助网站解决这个问题是一个大生意吗？

Is this a big business now helping websites figure this out?

是的。

Yeah.

你知道，这很有趣。

You know, it's interesting.

我的故事中主要聚焦了两家公司。

So the I mean, I kind of focus on two companies in my story, though.

有很多公司都在做这方面的工作。

There is many companies that work on this.

其中之一是Datadome，它专门从事阻止机器人的工作。

One is Datadome, which is kind of a specialist just in the blocking of bots.

一个是Cloudflare，这家公司更广为人知的是提供CDN服务，帮助网站连接服务器并分发内容。

One is Cloudflare, which is a company more known as CBN, or it sort of helps websites sort of just kinda connect with servers and deliver the content.

但它也提供反机器人软件。

But it also offers this bot blocking software.

我还特别提到了一家叫Dark Visitors的公司，它由一位Meta工程师创立，而他六个月前辞职，全职投入这家公司。

I actually also highlighted this company, Dark Visitors, which was started by a meta engineer, and then he quit six months ago to sort of work on this full time.

它的主要功能是进行数据分析，以弄清楚所有机器人的行为。

And it's sort of more analytics for finding out what all the bots are.

但哪家公司是最好的？

But who's who's the who's the best company?

人们最常求助的是哪家？

Who who do people turn to most?

你知道，一些出版商曾推荐过Datadome。

You know, some publishers kinda did flag Datadome.

Datadome和另一家公司Human在一份Forrester报告中被列为机器人管理领域的领导者，因为它们更专业。

Datadome and other company, Human, was flagged in a Forrester report as, like, the leader in this bot management category because they're more specialist.

但你知道，其他出版商也很喜欢Cloudflare，而且Cloudflare在这方面一直非常公开。

But, you know, other publishers like Cloudflare too, and Cloudflare's kind of, you know, been very public about this.

他们在夏天发布了一则公告，表示将帮助出版商屏蔽机器人。

They made an announcement over the summer that, like, they were gonna help publishers block the bots.

这是一个巨大的机会吗？

You know, is it a big opportunity?

你还问过，我认为出版商表示，这曾经只是他们的IT人员偶尔会考虑的事情，没那么重要。

You also asked, like, I think for publishers, they've said that this is, like, a kind of, like, a thing that maybe their IT guys used to only think about, like, you know, and it wasn't that big a deal.

但现在，销售人员也开始关注这个问题了。

But now, like, the salespeople are, like, thinking about it.

现在这已经是一个商业问题了。

Now it's, like, a commercial issue.

有一位出版商告诉我，这甚至比我们的广告技术供应商还重要，而这些供应商可是相当关键的，毕竟广告是许多出版商的主要收入来源。

And, you know, one per one publisher told me, this is even more important than, like, our ad tech vendors, these vendors, which is kind of a big thing because ads are how a lot of publishers make money.

所以这个问题正变得越来越重要，但它会成为一个大生意吗？

So it is increasingly important, but is it a big business?

一些公司正在增长，我所接触的公司表示，他们已经从这些公司那里获得了业务。

Some these companies are growing, and they've talked about you know, the companies I talked to, they said they've been getting business from there.

但AI对他们的客户来说是一种生存威胁。

But, I mean, the AI is an existential threat to their customers.

所以你有一些客户在使用你，但他们能否长期存活下来成为付费客户，这是我一个疑问。

So it's sort of you have customers who are using you, but can they survive long enough to become a, you know, paying customers is kind of, a question I had.

那么，让我问你一个稍微不同的问题：根本问题之一是，网站能否让这些运行爬虫的公司长期为内容付费？

So Well, so and so let me ask you a slightly different question, which is that the the fundamental one of the fundamental issues is the idea of will websites get these crawlers, these companies operating these bots, will they get them to pay for the content in the long run?

至少，这是我们的希望。

That's the hope, at least.

你所做的这些报道，能帮助我们回答这个问题吗？

This reporting that you've done, does it help us answer that question at all?

还是说它只是突显了让这些公司为数据付费有多么困难？

Or is it really just highlighting how difficult it's going to be to get those companies to pay for the data at all?

我认为，进行这种屏蔽的部分目的，是为了能够量化数据。

I think, you know, I think part of the effort to do this blocking was to be able to quantify Mhmm.

AI公司对出版商内容的需求有多大。

How much the AI companies needed the content for publishers.

所以，如果我可以，比如去联系OpenAI，作为一个没有和OpenAI达成协议的出版商，或者即使我已经是他们的客户，但我想重新谈判协议。

So if I could, you know, go to OpenAI if I if I'm a publisher that doesn't have a deal with OpenAI, Or even if I am I am a publisher that has a deal with OpenAI, but I wanna renegotiate it.

因为很多这类协议——我几个月前报道过——AI公司和出版商之间的现有协议其实并没有基于太多数学计算或真正的商业考量，因为当时没人知道这种经济模式会是什么样子。

Because a lot of these deals, I reported this a couple months ago, they weren't really the the existing deals between AI AI companies and publishers weren't really based on a whole lot of math and a whole lot of, like, real, like, business considerations because no one really knew, what this would what this economy would look like.

所以他们只是随便定了些固定数字。

So they're kind of just flat numbers they could come up with.

因此，很多这种屏蔽行为的目标是，好吧。

So the the goal with a lot of this blocking is, okay.

如果我们能通过展示AI公司的爬虫访问我们网站的频率和数量，来量化AI公司对我们有多依赖，我们就可以把这些数据作为谈判条件，比如说。

If we can quantify how much the AI companies need us by showing how much how often their bots come to our sites, we can start using that as a term in deals and, you know, saying

这样就有了具体的数据了。

this actually have a number then.

我们有了筹码。

We have we have leverage.

我们有议价能力。

We have leverage.

有人提出了类似按抓取次数付费或按查询次数付费的模式。

And there's been there's been kind of models suggested, like, by cloud for, like, pay per crawl or by like, pay per query.

所以我认为，如果出版商无法量化这一点，谈判就会变得困难。

So I think that the problem is is if publishers can't quantify it, it makes it hard to negotiate.

现在，我来谈谈Grok。

Now, you know, I talk about Grok.

我认为我们不知道Grok是否与出版商达成了任何协议。

Like, I don't think we know of Grok having any deal with publishers.

因此，也许OpenAI是整个过程中最负责任的公司之一，它们也是最早进行授权协议的公司。

And so maybe, you know, like, OpenAI is being maybe among the best sort of corporate citizens in this whole endeavor, and they're also the ones doing the licensing deal.

所以，那些根本没出现的公司，本来也不会和你达成协议，因为它们根本不愿意。

So maybe the companies that, like, aren't showing up, you weren't gonna get a deal with anyway because they weren't gonna be willing.

对。

Right.

但我觉得更大的问题是，我们是否会进入一个拥有标准的世界，而这就是商业运作的常态？

But, yeah, I think the bigger question is, like, are we gonna get to a world where there are standards and that's just, like, the you know, I the kind of the way business is done?

或者，所有合法的AI公司都必须现身，说：嘿。

Or and then, you know, all AI companies to be legitimate are gonna have to, you know, show up and say, hey.

我是机器人，要让出版商知道，并获得他们的许可。

I'm a bot and let publishers know and give and give permission.

或者也很可能，这将变成一场全面混乱，出版商无法控制局面。

Or it's it's also very likely that it'll just become, a free for all, and publishers won't be able to control it.

所以

So

很好。

Great.

凯瑟琳，非常感谢你前来参加。

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

和往常一样，这是一次很棒的对话，我期待着很快再和你交谈。

It's a great conversation as always, and I look forward to talking with you again soon.

好的。

Okay.

很好。

Great.

谢谢，阿卡什。

Thanks, Akash.

C3AI 昨晚公布了季度业绩。

C3AI reported quarterly results last night.

与一年前相比，该公司收入下降了约20%。

The company's revenue shrank about 20% compared to a year ago.

这是连续第二个季度出现收缩。

It was the second straight quarter of contraction.

去年这个时候，公司的总收入还在接近30%的增长。

This time last year, the company's top line was growing nearly 30%.

这一切都为公司新任首席执行官史蒂文·埃希基安铺平了道路，他于今年九月刚刚上任，我想请他来帮助我们了解他将如何重振公司。

It all sets the stage for the company's new CEO, Steven Ehickian, who just took that position in September, and I wanna bring him on to help us understand how he is going to reenergize the company.

史蒂文，欢迎来到TI TV。

Steven, welcome to TI TV.

很高兴你能来。

It's great to have you here.

谢谢，阿卡什。

Thank you, Akash.

很高兴来到这里。

Nice to be here.

我想把这次讨论分成两部分。

So I wanna segment this discussion into two parts.

我想先谈谈你在过去90天研究公司过程中发现的问题，然后再谈谈你计划实施的解决方案。

I wanna talk about the problems that you've high that you've identified as you've studied the company over the course of the last ninety days, and then I wanna turn to talking about the solutions that you plan to implement.

你知道，昨晚的电话会议上，有人问你，为什么增长停滞，为什么连续两个季度出现收缩，而去年同期却不是这样。

You know, last night on the call, you were asked about why growth or why contraction had really persisted for these two straight quarters compared to this time last year.

你提到销售执行是其中一个根本原因。

And you mentioned sales execution as one of the fundamental reasons.

稍微解释一下。

Unpack that for a little bit.

你所说的销售执行指的是什么？

What did you mean by sales execution?

是的。

Yeah.

当你有一个由汤姆·西贝尔领导的企业，而他是这家公司的CEO，今年早些时候他公开了健康问题，当这样一位核心领导者被 sidelined 时，对任何组织来说都是沉重的打击。

So when you have a business that's led by Tom Siebel, who's the CEO of this company, and he had health issues earlier this year that he disclosed, that is a hard knock against any organization, when you take a leader like that and sidelined.

他是这家公司的核心与灵魂。

He was the heart and soul of this company.

因此，当他被 sidelined 时，确实严重拖慢了销售方面的执行进度。

So when sidelined, he it really did slow down the execution across sales.

这到底意味着什么？

What does that actually mean?

我们的业务主要是落地我们所说的初始生产部署，即IPD。

Our business is about landing what we call initial production deployments, IPDs.

这些是与客户合作的初始用例。

These are initial use cases with customers.

这些是核心运营中的高价值用例。

These are high value use cases across core operations.

当我们交付真正的经济效益时，这些用例就会转化并增长。

When we deliver real economic value, those convert and grow.

不幸的是，在第一季度，我们缺乏强有力的领导来推动问责制，以确保这些IPD的执行。

Unfortunately, in Q1, we didn't have the strong leadership to drive accountability to actually ensure the execution of those IPDs.

因此，我认为这是未达成目标的首要原因。

So I would say that was the number one reason for the miss on that.

好的。

Okay.

我想再具体一点，对于那些没有在销售团队工作过、不太了解组织中发生这种情况意味着什么的人来说。

And and I just wanna get a little more granular here, you know, for for those of us who haven't been on sales teams and don't really understand, you know, what it's like when something like that happens in an organization.

我的意思是，团队没有得到关于如何定位产品的指导吗？

I mean, is this direction on how to position the product that the teams weren't getting?

是定价问题吗？

Was it issues around pricing?

你能帮我们具体了解一下，现实中这种情况是什么样子的吗？

You know, was it sort of help us understand really, you know, on the ground what that looked like.

你知道，企业销售不是那种简单的即插即用方案，比如你有个小零件就能大量销售。

You know, so enterprise sales is not a quick turnkey solution when you have a little widget and you just do high, high volume.

这是一种高度依赖关系的执行方式，你需要应对复杂的使用场景，并与我们的内部团队和客户紧密协作推动执行。

This is a highly relationship based execution where you have a complex use case, and you really have to drive the execution with our internal teams and our customers.

这些使用场景包括供应链优化、资产绩效管理到需求预测等。

These are use cases from like supply chain optimization, to asset performance, to demand forecasting.

我强调执行在这里至关重要，因为当我们签下客户时，我们与客户就希望交付的价值达成了共识。

I highlight the execution is so critical here, because when we close an account, we have an alignment with the customer of the value we want to deliver.

如果我们不能及时交付这些价值，这些公司就不会转化，而这需要我们公司全力以赴去协助实现。

If we can't deliver that in a timely basis, those companies don't convert, and that requires the full scale and effort of our company to go help deliver that.

所以当你失去领导者，比如你的CEO被边缘化时，你会失去很多。

So when you lose your leader, your CEO, and he's sidelined, you lose a lot.

这就像一个暂时失去方向的组织。

It's kind like a rudderless organization for a while.

所以对于任何没有做过企业销售的人来说，这对一个组织来说是巨大的打击。

And so for anybody who hasn't run enterprise sales before, that's a huge knock against an organization.

所以

So

也就是说，根本不知道该优先跟进哪些客户，也不知道到底该往哪个方向推动增长，诸如此类的问题。

that's- Basically, basically not knowing which accounts to prioritize, which, you know, what direction really you really wanna take the growth, stuff like that.

Akash，这影响到组织的每一个维度、每一个相关环节。

Akash, it's across every dimension, every part of the organization that touches.

那么，你该从哪里开始？该瞄准哪些客户？

So how do you get started, which Snows target?

如何提升推进速度？

How to drive velocity?

一切都要从高层开始。

All starts from the top.

我确实想谈谈你们计划中的解决方案，但我想先聊聊一个引起我注意的客户细分领域。

I do want to talk about the solutions that you're planning here, but I wanted to talk about one customer segment that caught my eye.

制造业在你们业务中的占比显著下降了。

Manufacturing as a share of your business declined considerably.

我想知道为什么偏偏是这个细分领域。

I wondered why that segment in particular.

我可以告诉你们的是，从长远来看，这个细分领域目前对我们来说正在快速增长。

What I can tell you going forward is that that segment is actually a fast growing Sydney right now for us.

当你想到离散制造时，我想到了我们为交付哥伦比亚级核潜艇所开展的工作。

And when you think of discrete manufacturing, I look at the work we're doing to deliver the nuclear class submarine, the Columbia class.

这完全关乎重振美国工业，特别是海洋工业基础。

This is all about re industrializing America, specifically the maritime industrial base.

我们在这个细分领域以及半导体领域都看到了显著增长。

We're seeing big growth on that segment, as well as semiconductor.

我之前提到过，AMD 是我们从第二季度开始的客户。

I kind of mentioned AMD is a customer of ours from Q2.

这再次聚焦于离散制造，这是我们关注的一个领域。

This is, again, focusing on discrete, you know, manufacturing, and that is an area of focus for us.

那么，史蒂文，你们接下来的计划是什么？

So what's the plan from here now for you, Steven?

你们打算如何让这家公司重回增长，并最终成为我们所听说的那些AI领域中的领军企业之一？

How do you plan to bring this company back to growth and ultimately, hopefully, becoming, you know, one of the booming names in AI that we've heard so much about?

是的。

Yeah.

所以，阿卡什，我与客户、潜在客户和合作伙伴进行了无数次会议。

So Akash, I've spent countless meetings with customers, prospects, and partners.

我想说的是，C3的机会比我预期的要大得多，主要是因为企业AI市场增长得太快了。

And I will say this, the observation is that opportunity for C3 is bigger than I could have anticipated, primarily because the enterprise AI market is growing so fast.

我认为我们的产品真正具有差异化，我们不仅仅是聊天机器人，而是在为大型制造企业、能源、医疗保健以及联邦、州和地方政府推动核心运营。

And I think we have a truly differentiated product here, where we are not just a chatbot, we are driving core operations for large manufacturing companies, energy, healthcare, and federal and state and local.

因此，我的重点是第一，执行。

So my focus is on, one, execution.

我每天深入销售团队，推动执行工作，最终为客户创造经济价值。

I'm in the weeds day to day with our sales team, driving these execution and ultimately driving economic value for our customers.

如果我做到了这一点，你就会看到转化效果。

If I do that, you will see the conversion.

如果我们专注于产品，我们拥有一个出色的平台，积累了十六年的开发成果，具备丰富的功能。

And if we focus our product, we have an amazing platform, sixteen years of development, plenty of capabilities.

我们将专注于正确的细分市场。

We'll be focused on the right segments.

我提到了联邦、州和地方政府、能源、制造、医疗保健以及其他一些选定领域，我们在这些领域都能获胜。

I mentioned federal, state and local, energy, manufacturing, healthcare, and select others, we win.

因此，这有两个方面：一是每天推动执行，二是将我们的资源聚焦于正确的市场和市场重点。

So there's two components, driving execution day to day, but also focusing our engine on the right markets and market focus.

那么如何

And how do

你如何将这一点与现实调和呢？正如你所说，企业——作为目前每家企业软件公司的主要客户群体——正面临让他们购买AI的挑战？

you square that then with the reality that enterprises, as you said, which are a big customer group for every enterprise software company right now is getting them to buy AI?

很慢。

It's slow.

我的意思是，这些企业说，我得找到合适的人选。

I mean, these enterprises are saying, look, I got to find the right people.

我得找到预算。

I got to find the budget.

我得找到投资回报率。

I got to find the ROI.

你知道吗，我们这周甚至看到一些数据，表明在某些情况下，大型公司的AI采用率已经开始趋于平稳。

You know, how do you square that with the fact we we even saw numbers this week that AI adoption in some cases for the largest companies is starting to plateau.

这不是我的意思。

That's not what I'm saying.

坦率地说，我们在这些行业——比如工业领域，再次强调制造业、医疗保健和能源——看到了强劲的宏观顺风。

And honestly, where the focus for us on these on these industries, like industrial industries, again, manufacturing, healthcare, energy, they're seeing strong macro tailwinds.

我们所解决的问题是消除一线工作中日常繁重的琐事。

And the problem we're solving is eliminating a lot of the drudgery from day to day of the work on the ground.

再想想这个，这并非一个横向用例。

Think of this, again, this is not a horizontal use case.

这些都是非常具体的应用，用于供应链、资产性能管理，以最小化停机时间。

These are very discrete applications for supply chains, for asset performance, to minimize downtime.

想想空军中的应用，确保最大数量的机队在空中飞行；对壳牌公司而言，确保石油生产按计划进行。

Think of the application across the Air Force to make sure you have the maximum number of fleet in the air to Shell, to make sure oil production is on schedule.

从半导体到疫苗的需求预测，这些正是我们正在影响的领域。

Demand forecasting, semiconductors to vaccines, these are what we're influencing.

所以从我的角度来看，这正是我们与这些超大规模企业和系统集成商建立合作伙伴关系的地方，我们能够将垂直技术应用到他们的横向平台上，从而推向市场。

So from my perspective, and this is where we're finding partnerships with these hyperscalers and system integrators, we are able to apply these vertical technology on top of their horizontal platforms to go to market.

我

I

想问你一个稍微私人一点的问题，就是你担任这个职位已经九十天了。

wanna ask you a slightly more personal question, which is you've been in the role now for ninety days.

我想知道，在你担任这个职位之前的九十天里——我不确定你是什么时候知道自己将成为CEO的。

I wonder, in the ninety days before you took the role, and I don't know when you knew that you were gonna be CEO.

大概在那个时期的某个时候吧。

Presumably, it was at some point in that period.

但你是如何为接手这样一个扭亏为盈的职位做准备的呢？

But how do you prepare for coming into a a turnaround role like this?

你都在和谁交谈？

Who are you talking to?

你有研究过那些经历过类似转型的公司的案例吗？

Are you studying playbooks of companies that have gone through this?

你是如何准备的？

How did you prepare?

嗯，有两个方面可能对我有帮助。

Well, two areas would probably help me.

第一，我曾经是C3的客户。

One, I was a customer of C3 first.

在加入C3之前，我负责管理美国总务署（GSA），当时正在推动政府效率提升，并评估人工智能解决方案以帮助实现这一目标。

So before C3, was running the GSA, the General Service Administration, and I was in process of making the government more efficient and evaluating AI solutions to help us do so.

我看到了整个市场的全貌，遇到了C3，并且看到了那些令人眼前一亮的演示。

And I saw the full swath of the market, came across C3, and I saw the demos, like, that jumped off the page.

我当时就想，这是什么？这就是我的初次接触。

I'm like, what is And so that was my introduction.

作为客户，我花了很多时间评估这项技术。

I spent a lot of time evaluating the technology as a customer.

所以这是我的第一段经历。

So that was my first.

其次，我过去曾创办过两家人工智能公司。

Second, I've built AI companies in the past.

两家公司都很幸运地被Salesforce收购了。

Two companies, was fortunate enough for Salesforce to acquire us.

因此，我了解技术的发展方向和能力。

So I understand where the technology is going and the capability.

我

I'm

抱歉，我不是故意打断你。

sorry, don't mean to interrupt you.

我问的是更哲学一点的问题，比如，你知道的，转型。

I'm asking a little more philosophical, like, you know, turnarounds.

转型真的很难，对吧？

They're really hard, Right?

历史上，至少在科技领域，成功的大型转型案例寥寥无几。

And and there's there's only been a couple of of of big turnarounds in history, at least in in technology.

我想知道，你有没有研究过这些公司？

I wonder, did you did you study any of these companies?

你有没有和那些设计过这些转型策略的领导者交流过？

Did you talk to any leaders who had sort of architected these playbooks?

是的，我花了很多时间研究。

Yes, I've spent a lot of time.

首先，我非常了解汤姆·西贝尔，也深入了解了现有董事会的战略，以及我们需要达成的目标。

One, I got to know Tom Siebel extremely well, got to know the boards of the existing strategy, and also knew where we needed to get to.

此外，根据我的背景，我身边还有一些在大型企业转型方面有经验的导师。

I also have, from my background, mentors in this space that done turnarounds of large enterprise companies.

所以我花了很多时间与他们相处。

So I spent a lot of time with them.

他们在过去帮助我构建硅谷的企业公司。

And they've been helpful in my previous past building enterprise companies in Silicon Valley.

所以是的，我研究过这些操作手册，并理解了其中的内容。

So yes, study playbooks, understood it.

但最重要的是，每家公司都是独特的。

But most importantly, every company is unique.

我花最多时间去理解的是与执行董事长汤姆·西贝尔及董事会的协同，了解他们的优势、哪些领域有资格获胜，以及哪些行业和产品。

And the alignment with the executive chairman, Tom Siebel, and the board was where I spent the most amount of time understanding their strengths, where they have the deserve, the right to win, which industries, which products.

这让我至少能形成一个有根据的判断，明确前九十天应该聚焦的方向。

And that allowed me, at least gave me an informed opinion of where to focus my first ninety days.

对。

Right.

斯蒂芬，非常感谢你参加我们的节目。

Well, Steven, thank you so much for coming on the show.

我们非常感激。

We really appreciate it.

我们期待看到你在未来几个月如何构建你的行动方案。

And we look forward to seeing how you architect your playbook in the months to come.

谢谢你，阿卡什。

Thank you, Akash.

非常感谢。

Appreciate it.

好的。

Okay.

这周围绕OpenAI的新闻层出不穷。

It has been a busy week of news with OpenAI.

其中一条被略微忽视的消息是我的同事斯蒂芬妮·帕拉佐洛的报道，称该公司正在开发一款新模型，有望重新向谷歌的Gemini三号模型发起挑战，而后者最近备受关注。

And one story that has been a little overlooked is my colleague Stephanie Palazzolo's reporting that the company is actually working on a new model that should hopefully pose a renewed challenge to Google's Gemini three model, which has been getting a lot of buzz.

我想请Stephanie来帮我们理解Sam Altman对这个新模型的想法。

I wanna bring on Stephanie to help us understand how Sam Altman is thinking about the new model.

Steph，欢迎再次做客节目。

Steph, welcome back to the show.

很高兴你来到这里。

It's great to have you here.

谢谢你们邀请我。

Thanks for having me.

这周对你来说

It's been

很忙吧？

a busy week for you, isn't it?

天啊。

Gosh.

你现在在圣地亚哥。

You're you're in San Diego.

你在大会上。

You're at the conference.

你收到了红色警报。

You got the you got the code red.

你拿到了新模型。

You got the new model.

我的意思是

I mean

是的。

Yes.

确实有点睡眠不足，但很享受。

Definitely a little a little sleep sleep deprived, but enjoying

这一切。

it all.

非常感谢你来到这里。

Well well, thank you so much for being here.

我们非常感谢。

We really appreciate it.

让我们谈谈这个新模型，它叫Garlic。

Let's talk about the new model, which is called garlic.

好吗？

Okay?

我们稍后再谈代号的事，但Garlic有什么不同之处呢？

And we'll get to maybe the code names later on, but what's going to be different about Garlic?

是的，OpenAI确实有一些有趣的代号。

Yeah, no, definitely some interesting code names from OpenAI.

但本质上，Garlic是一个新的预训练模型，OpenAI的高管们表示，它在编码和推理等领域的表现与谷歌的Gemini 3和Anthropic的Opus 4.5模型相当，而这对于所有这些公司来说都非常重要。

But essentially what Garlic is, it's a new pre trained model that OpenAI executives have been saying performs as well as Google's Gemini three and Anthropix's Opus 4.5 models in areas like coding and reasoning, which of course are very important to all these companies.

我认为这之所以如此有趣，是因为预训练一直是OpenAI历史上相对薄弱的领域。

And I think kind of the reason why this is so interesting is because pre training is an area that OpenAI has kind of historically struggled with.

为了提供一些背景，预训练基本上是模型训练的第一步，即向模型展示海量的网络数据、书籍和其他来源的数据，以教会它在这些数据之间建立联系。

And to give some context, pre training is basically the kind of first step of model training where you show a model just, you know, tons and tons of data from the web, from books, from other sources to teach it to make connections between all of them.

所以，尽管OpenAI过去在这方面表现不佳，但谷歌在预训练方面做得非常好。

So, you know, while OpenAI has kind of struggled in this in the past, Google has done really well on pre training.

这之所以如此重要，是因为许多研究人员相信，当你改进预训练时，也能提升模型在训练数据之外发现新事物的能力。

And the reason why that's so important is because a lot of researchers believe that when you improve pre training, you also improve a model's ability to come up with new discoveries beyond what it's been trained on.

因此，这正是所有AI公司的终极目标。

And so that's kind of the holy grail of all these AI companies.

你希望模型能够找到类似治愈癌症的方法，或者开展AI研究的新途径。

You want the models to become to come up with, you know, like cures for cancer or like new ways to do AI research.

因此，这就是所有AI实验室都在努力追求的终极目标。

And so that's kind of this holy grail that all the AI labs are working towards.

我们觉得这个模型会多快发布？

How quickly do we think this model is going to come out?

具体时间还不清楚，但看起来OpenAI正在迅速推进其下一波模型的发布，尤其是考虑到谷歌Gemini三号获得的积极反响。

So, don't know the exact timing, but it does seem like, you know, OpenAI is moving very quickly to release its next couple models, especially given, you know, the kind of positive reaction to Google's Gemini three.

所以，我认为我们肯定会在明年年初或未来几个月内看到基于Garlic的模型。

And so, think I we'll definitely see some model that's based on garlic, you know, by early next year, in the next couple months or so.

所以一定要留意这个。

So definitely keep a lookout for that one.

我得说，我们之前看到GPT-5发布了盛大的主题演讲和各种宣传。

I have to say, we saw, I think it was GPT-five that had the big sort of splashy keynote and and everything like that.

而显然，从那以后，情况在许多方面都发生了巨大变化。

And we obviously know the story has changed a lot since then in in in many ways.

我几乎在想，既然模型竞争如此激烈，未来的发布会和发布活动会不会都变得低调一些了。

I almost wonder if the keynotes and the releases everything's gonna be toned down a little bit now that the model wars are so competitive, honestly.

既然一个月后，别人可能就发布了一个产品，完全把你甩在后面，那何必为六、七、八、九号模型大张旗鼓呢？

Like, it's why make a huge deal out of six, seven, eight, nine when a month later, you know, somebody else could release something and totally blow you out of the water?

是的。

Yeah.

我认为OpenAI已经从GPC中吸取了教训。

I think OpenAI has definitely learned its lesson from GPC.

所以我认为我们不应该期待很快看到大规模的发布，比如GPT-6或GPT-7。

So I don't think we should expect to see huge releases or GPT-six, GPT-seven anytime soon.

它们可能是5.2、5.3、5.5。

They go maybe 5.2, 5.3, 5.5.

所以这更像是渐进式的更新，而不是一次盛大的发布。

So it's much more kind of like incremental updates versus a big splash g launch.

我想聊聊你本周在专栏中写到的另一个话题。

I wanna talk about another topic that you wrote about this week in your column.

你写到了一个叫做‘分裂脑问题’的东西。

You wrote about something called the split brain problem.

嗯。

Yeah.

这正是我想说的。

That's what I was gonna say.

那是之前的内容，我只是想确认我读对了。

That was a previous trying to make sure I read it right.

分裂脑问题。

The split brain problem.

那是什么？为什么OpenAI在它上面遇到困难？

What is that, and why is OpenAI struggling with it?

是的。

Yeah.

请耐心一下，因为接下来会有点技术性，但我尽量讲清楚。

So bear with me because it's going get a little technical, but I will try to break it down.

这本质上是一个问题，OpenAI的研究人员以及很可能其他实验室都注意到了：你向模型提问的方式，会导致答案出现极其不同的变化。

So this is essentially a problem that researchers at OpenAI and most likely the other labs have kind of noticed where basically just the way that you phrase a question to a model can lead to extremely different, you know, variations on the way that answers.

而这基本上都是因为模型的训练方式造成的。

And this is basically all because of the way that models are trained.

因此，研究人员会给模型提供数据，以帮助改善模型在特定领域的行为，比如推理、数学或科学。

And so essentially, researchers give models data to help improve the model's behavior in very specific areas like reasoning or math or, you know, science.

例如，研究人员可能会给模型提供正式的数学证明，以教会它更准确地回答数学问题。

And so, for instance, like a researcher might give a model kind of formal math proofs to teach it how to better answer math questions more accurately.

然后，另外，研究人员可能给模型提供更通用的问答对，以教会它以礼貌、友好、符合人类喜好的方式回应。

And then separately, maybe the researcher gives the model just more general kind of question answer pairs to teach it to respond politely and friendly and in ways that humans like.

例如，使用大量表情符号和项目符号来分解很长的答案。

So using lots of emojis and bullet points, for instance, to break down a really long answer.

因此，在未来，每当模型运行像HATGBT这样的系统时，它可能会根据它认为自己所处的情境做出不同的回应。

And so in the future, whenever the model is powering to HATGBT, for instance, it might respond differently based on which scenario it believes that it's in.

这种第一种情境，比如它在回答非常正式的数学证明，或者第二种情境，它在回答更一般性的问题，此时它的重点更放在友好、语气和回答格式上。

This kind of first scenario where maybe it's answering very formal math proofs or the second scenario where it's answering more general questions and where its focus is more on being friendly and the tone and the formatting of its answer.

所以，如果有人以正式证明的风格提出数学问题，模型可能会给出正确的答案。

So, if somebody, you know, phrases a math question like in the style of a formal proof, maybe the model will answer correctly.

但如果你以更随意的方式提问，模型可能会误以为自己处于那种因回答友好、格式美观而获得奖励的情境中。

But if you phrase it more like casually, the model might accidentally think it's in the scenario that it gets rewarded for like having a really friendly, nice formatted answer.

因此，它可能会给你一个错误的数学答案，但也许它的回答格式非常精美，充满了表情符号、项目符号等，这显然不是你想要的。

And so, it might give you like an incorrect answer to the math problem, but maybe it's like formatted super nicely and it has a bunch of emojis and bullet points and things like that, which is obviously not what you want.

所以，我的意思是，归根结底，这其实又回到了一个问题：你输入模型的内容很大程度上决定了你能得到多好的回应。

So so I mean, at its core, this is sort of it's it's coming back to the issue that it depends a lot what you put into the model in terms of how how good a response you get out.

我想知道，你所识别出的这个问题的细致程度，是否反映了模型已经取得了多大的进展？

I wonder if if the granularity of this problem that you've identified, does that reflect how far the models have come?

或者反过来，这是否反映了它们还有多大的提升空间？

Or does it, on the flip side, reflect how much still they have to improve?

这个问题在这一系列问题中处于什么位置？

Where where does this problem sit in that stack?

我认为是第二种情况。

I think it's more the second.

我的意思是，这非常有趣，因为一方面，你可以说人类也会被问题的措辞所迷惑。

I mean, it's really interesting because on one hand you could argue that humans get tripped up on the way that people word questions to you.

比如，我给你一个措辞巧妙的问题，你可能会答错。

Like maybe I give you a tricky wording to a question and you might get it wrong.

但你期望这些 supposedly 即将变得超人的模型，能够找到治愈癌症的方法。

But you kind of expect these models which are supposedly going to be superhuman soon, they're going to like find a cure for cancer.

你自然会期望它们不会出现同样的问题。

You would kind of expect them to not have those same issues.

是的，它们应该能够理解你说的话。

Yeah, they should be able to interpret what you're saying.

你知道我的意思吧，老兄。

You know, catch my drift, man.

但即使是人类，你也可以想象，比如一个数学教授，如果我用正式的数学证明方式问他一道数学题，他不会答错；但如果是我在对话中随意地问，他可能会出错。

But even like, I mean, even for humans, like you can imagine, like a math professor would not answer a math question incorrectly if I gave it to them in a formal math proof or something versus if I asked them very casually in a conversation.

所以这真的很有趣，但也挺奇怪的，模型居然还在犯这些有点基础的问题。

So it's just very kind of interesting, but like very weird that models are still having these somewhat, you know, basic problems.

对。

Right.

我最后问你一个问题。

Last question for you.

你本周参加了一个会议，你发表的《Code Red》文章引发了大量热议。

You are at a conference this week and there has been a lot of buzz about the Code Red story you published.

我想知道现场的讨论情况如何，不只是关于OpenAI发布的《Code Red》，而是更广泛地看这个活动上的整体反响。

I wonder what the chatter is on the ground, not just about the Code Red that OpenAI has issued, but also on the ground more broadly at this event.

是的。

Yeah.

这周我正在圣地亚哥参加NeurIPS会议。

So this week I'm at NeurIPS in San Diego.

非常有趣。

It's been super fun.

很高兴见到这里的每一位研究人员。

Loved meeting all the researchers here.

这个会议有上百万人参加。

There's like a million people at this conference.

关于Code Red，确实有一些对话提到了它。

I think for Code Red, yeah, definitely has been coming up in some conversations.

我觉得一方面，所有研究人员之间都会互相交流。

I do think that on one hand, like all the researchers talk to each other.

所以我认为，早在Gemini Three发布前几周，很多人就已经知道它会是一个出色的模型，OpenAI很可能会对此做出某种回应。

So I think a lot of them knew even, you know, weeks before Gemini three was released that it was going to be a good model, that OpenAI was probably going to have some sort of reaction to it.

因此，一方面，这对他们来说并不特别令人惊讶。

So I think on one hand, it wasn't like incredibly surprising to them.