OpenAI领导层重组、AI独角兽企业与白领工作

欢迎收听本播客。

Welcome to the podcast.

我是主持人杰登·谢弗。

I'm your host, Jayden Shafer.

今天节目中，我们将涵盖一系列不同的新闻故事。

Today on the show, we're covering a bunch of different news stories.

OpenAI正在重组其领导层。

OpenAI is reshuffling their leadership.

他们正努力夺回在企业市场失去的阵地，而这一点似乎长期以来都被Anthropic超越了。

They're trying to get back a bunch of lost ground on enterprise, which it feels like Anthropic has kind of been beating them at for a while now.

有一家关键的OpenAI基础设施合作伙伴刚刚达到独角兽估值，当然，还有一大批推理初创公司正展现出2026年AI资金真正将流向的领域。

There is a key OpenAI infrastructure partner that has just hit unicorn status, and of course, there's a whole wave of inference startups that are showing basically where I think a lot of the AI money is actually going to be flowing in 2026.

此外，我想分析一项新基准，它揭示了为什么白领工作实际上并未如过去预测的那样迅速消失。

Plus, I want to break down a new benchmark that shows why white collar jobs are actually not disappearing nearly as much as they were predicted in the past.

同样，一位前红杉资本合伙人认为，AI代理终于能够解决日程安排问题，如果你想了解一些有趣的AI应用场景，这值得一看。

And again, a former Sequoia partner is betting that AI agents can finally fix calendar scheduling if you wanna get into some interesting AI use cases.

这是AI聊天播客，一个每日更新、聚焦AI领域最重要新闻与对话的节目。

This is the AI Chat podcast, which is a daily podcast covering the most important news and conversations in AI.

那么我们开始本期节目吧。

So let's get into the episode.

据《The Information》报道，我认为——顺便说一句，如果你还没关注《The Information》，它真的是最好的科技内幕新闻媒体之一，免费给你们安利一下。

I think according to reporting from the information they have, basically if you by the way, if you don't have the information, it's I think one of the best insider tech news, so, you know, free plug to them.

但OpenAI发布了一份报告，称已任命布雷特·佐夫负责企业销售业务。

But OpenAI, they have a whole report on which says that OpenAI has appointed Brett Zof to lead their enterprise sales efforts.

佐夫在公司内部非常熟悉。

Zof is pretty familiar inside of the company.

他此前曾担任训练后推理的副总裁，2024年离职后与前OpenEye首席执行官或首席技术官米里亚姆·拉迪共同创立了Thinking Machine Labs。

He previously was the vice president of post training inference before he left in 2024 to go cofound Thinking Machine Labs with former OpenEye CEO or CTO, Miriam Radi.

仅这一件事本身，就引发了大量关注。

So this, I mean, just in and of itself, this got a lot of headlines.

他离开OpenAI去创办这家公司，随后融资十亿美元，不到一年、甚至还没推出任何产品，就又回到了OpenAI，担任这个新职务，这确实挺讽刺的。

The fact that he left OpenAI to go found this company, and then, you know, they raised like a billion dollars, and less than a year later, before any products were shipped, he came back to OpenAI, and what, you know, it's like, has this new role inside, which is kind of funny.

他上周回来了，但关于他离开Thinking Machine Labs的具体原因，我们目前并不清楚，情况依然不明朗。

He came back last week, and I think there's not like a lot of new, like, circumstances that we know about why he left Thinking Machine Labs, so that's pretty unclear.

这个时间点非常引人注目。

The timing is pretty notable.

目前，OpenAI在企业市场面临着巨大压力，我认为这个职位将使他成为其2026年增长战略的核心人物。

OpenAI right now has a ton of pressure in the enterprise market, and I think this role is gonna put him kind of at the center of their 2026 growth strategy.

尽管OpenAI比许多竞争对手更早推出了ChatGPT Enterprise，但其市场份额却一直在稳步下降。

I think despite launching ChatuchPut Enterprise earlier, than a lot of the other competitors, OpenAI's market share has actually steadily declined.

很多人可能不知道，但Menlo Ventures发布的一些数据显示，OpenAI的企业级LLM使用率从2023年下降了约50%，到2025年仅剩27%，这意味着你可以观察其他竞争对手，了解行业内的其他动态。

I think not a lot of people know this, but there's some data that came out of Menlo Ventures that says that OpenAI's enterprise LLM usage has fallen roughly 50% in 2023 to about, 27% by the '25, which means that you can look at some of their other competitors to see, you know, what else is happening in the industry.

例如，Anthropic目前的市场份额已达到约40%，而谷歌的Gemini也在持续取得进展。

Anthropic, for example, now leads has around 40%, and Google's Gemini has it's been making a bunch of gains.

虽然进展较慢，但它确实在逐步抢占市场份额。

It's it's a lot slower, but it is making some it is gaining some ground there.

所以，我认为OpenAI的管理层对此并非无动于衷。

So I think this isn't, you know, lost on OpenAI's leadership.

萨姆·阿尔特曼显然已指出，OpenAI的增长速度过快是公司内部的一大担忧。

Sam Altman has apparently kind of flagged the fact that OpenAI is growing as a big concern inside of the company.

曾经出现过一个著名的警示时刻。

There was kind of the famous, you know, red flag moment.

首席财务官莎拉·弗莱尔也表示，企业增长是他们今年的首要任务。

And the CFO, Sarah Fryer, also said that enterprise growth is a top priority this year for them.

他们还扩展了与ServiceNow的合作关系。

They've also expanded their partnership with ServiceNow.

我认为这表明OpenAI正计划为大型企业客户争取得更加积极。

I think that kind of shows OpenAI is planning to fight a little bit harder for larger business customers.

这将会很有趣。

So this is gonna be interesting.

接下来我想聊一聊一家叫LiveKit的公司。

The next thing I wanna talk about is a company called LiveKit.

它是OpenAI最重要的基础设施合作伙伴之一，并且已正式成为独角兽企业。

It's one of OpenAI's really key infrastructure partners, and they have officially reached unicorn status.

他们是一家实时语音和视频基础设施公司。

So they are a real time voice and video infrastructure company.

他们为聊天GPT的语音模式提供支持。

They power chat GPT's voice mode.

他们刚刚以10亿美元的估值融资1亿美元。

They just raised a $100,000,000 at a 1,000,000,000 valuation.

本轮融资由一级风投Index Ventures领投。

The round was led by Index Ventures, a tier one VC.

他们还获得了Altimeter、Redpoint以及其他一些原有投资人的跟投。

They also had Alimeter and Redpoint and some other preexisting investors that all jumped in on that.

但LiveKit本质上是在疫情期间作为开源项目启动的。

But LiveSkip or LiveKit essentially started as an open source project during the pandemic.

创始人Russ Dassault和David Zhao专注于打造无中断的音视频工具，随后业务自然迅速发展。

The founder, Russ Dassault and David Zhao, they're focused on building interruption free audio and video tools, and then the business just sort of took off from there.

我觉得这很有趣。

I think it's interesting.

当许多公司开始要求他们提供这项服务并代为管理时，这项业务迅速发展起来，于是他们构建了一个基于云的语音AI解决方案。

It took off especially when a lot of companies started asking for them to basically do it but manage it, so they built a cloud based solution for the voice AI.

如今，LiveKit的客户包括OpenAI、XAI、Salesforce、Tesla，以及多家应急服务机构和一些心理健康服务机构。

So today, LiveKit's customers, they have OpenAI, XAI, Salesforce, Tesla, a bunch of emergency services, some mental health providers.

但对我来说，真正有趣的是，随着AI语音变得越来越主流，基础设施层正成为这一技术栈中极具价值的部分。

But I think the thing that's really interesting for me is that as AI voice is becoming a lot more mainstream, the infrastructure layer is turning into a really valuable part of this stack.

因此，他们能够筹集到1亿美元，而且还在为OpenAI的语音模式以及其他众多玩家提供支持，这充分说明了这家公司及其类似企业将具有多么重要的价值。

And so, I think the fact they were able to raise a $100,000,000, and of course they're fueling like OpenAI's voice mode and a lot of other, a lot of other players shows just how valuable this company and others like it will be.

好的。

Okay.

同样地，我想谈谈另一家公司，名叫Infraq。

So, in that same theme, I want to talk about another company, which is called Infraq.

这是一家将许多流行的开源项目商业化的企业，其中特别突出的是一个名为VLLM的项目。

This is a company that is commercializing a lot of popular open source projects, but one in particular is called VLLM.

他们刚刚完成了1.5亿美元的种子轮融资。

And they have just raised a $150,000,000 in seed funding.

他们的估值为8亿美元。

They have an $800,000,000 valuation.

本轮融资由a16z和Lightspeed共同领投。

The round was co led by a16z and LightSpeed.

VLLM以及一系列类似项目专注于推理，也就是在生产环境中高效运行模型的过程。

VLLM and a bunch of similar projects are focusing on inference, or essentially the process of running models really efficiently in production.

我认为，随着AI正从大规模训练突破转向更多实际部署，投资者正在大量投入资金，用于开发让模型更快、更便宜、更具可扩展性的工具。

I think right now, as we're seeing kind of a shift in AI from these sort of big like training breakthroughs to a lot more real world deployment, investors are putting a lot of money into the tools that make models faster and cheaper, cheaper and more scalable.

所以，InfraX的故事与另一个最近的衍生公司RadiaX Arc非常相似，后者也商业化了SG Lang。

So, InfraX story, I think is very close to another recent spin out, which is called RadiaX Arc, which also commercialized SG Lang.

这两个项目都源自加州大学伯克利分校的AI研究生态，我认为这展现了一个行业大趋势：训练固然能吸引头条，但真正创造商业价值的地方在于推理。

Both of those projects originated from UC Berkeley's AI research ecosystem, and I think they show like a pretty big trend in the industry, which is that training is gonna definitely grab headlines, but inference is where a lot of the real business value is actually being created.

他们能够筹集8亿美元，这一点足以证明这一点。

And I mean, the fact that they were able to raise $800,000,000, I think goes to prove this.

现在，我们来看当前AI领域最重要的研究进展之一，我认为这个进展能穿透我们常看到的诸多炒作：Merkur推出了一项新基准，名为Apex Agents。

So, now for one of the most important research stories in AI right now, and I think this is one that cuts through a lot of the hype that we see, there's a new benchmark from Merkur, it's called the Apex Agents.

这本质上是他们用真实白领工作来测试主流AI模型，涵盖了一系列来自咨询、法律、投资银行等领域的任务，我认为结果非常令人印象深刻，或者至少非常有趣。

And this is essentially they tested leading AI models on real white collar work, and they have a whole bunch of tasks that are drawn from, you know, like consulting, law, investment banking, and I think the results were really impressive, or very interesting in any case.

即使是表现最好的模型，也只能正确回答约25%的问题。

So even the best performing models struggled to get more than about 25% of the questions right.

Gemini 3 Flash 在这里表现最佳。

Gemini three flash was actually the leader here.

它取得了24%的得分。

It did 24%.

紧随其后的是GPT 5.2，得分为23%。

Right behind it was GPT 5.2 at 23%.

我认为其他大多数模型的得分都接近18%。

I think a lot of the others were closer to 18%.

最大的失败点不仅在于推理和孤立任务，更在于跨多个领域协同操作的能力。

The biggest failure point was not just kind of like reasoning and isolation, but it was actually operating across multiple domains.

这需要具备真实的知识，我认为所有这些都要求从邮件、文档、内部政策和协作工具中提取信息，比如Slack和Google Drive。

So having real knowledge, and I think all of this requires pulling information from emails or documents or internal policies and collaboration tools, right, like Slack and Google Drive.

考虑到将所有这些复杂地整合在一起，我认为当前的代理模型在跨这些环境保持连贯性方面仍然面临巨大挑战。

And with all like, the complexity of tying all that together, I think today's agent models are still really struggling to maintain coherence across all of those environments.

我认为这背后的含义非常重大。

And I think the implication to that is actually really big.

基本上，人工智能正在快速进步，但尚未准备好真正取代公司内部许多高价值的专业岗位。

Basically, AI is moving and improving a lot faster, but it's not yet ready to actually replace a lot of the high value professional roles inside of companies.

所以至少目前来看，这些模型看起来更像需要严密监督的实习生，而不是自主的工作者。

So I think for now anyways, the models look a lot like a lot less like autonomous workers and probably more like interns who need heavy supervision.

这正是我目前发现它们所处的位置。

This is where I've found them.

你当然可以让这些AI模型出色地完成某一项特定任务，从而为你节省时间，但它还远未达到能完全接管一个完整岗位的程度。

You definitely can get these AI models to do one particular task very well and save you time on it, but it's not like it's gonna go take over a full role yet.

还不到时候。

Yet.

关键词是‘还不到时候’。

Keyword yet.

好的。

Okay.

今天终于有一位前红杉资本合伙人押注AI代理能够最终解决工作中最棘手的问题之一——日程安排。

Finally today, a former Sequoia partner is betting that AI agents can finally fix one of work's most heated problems, which is scheduling.

我觉得这个故事挺有趣的。

This one I thought was kind of a funny story.

基本上，卡齐·基米吉在红杉资本工作了大约六年，最近他创办了一家AI日历初创公司，名叫Block Kit。

Essentially, Kazi Kimiji, who spent about six years at Sequoia Capital, he just launched an AI calendar startup, it's called Block Kit.

他们刚刚完成了500万美元的种子轮融资，由红杉资本领投。

They just raised $5,000,000 seed round, was led by Sequoia.

所以，我想他们肯定希望他去创办这家公司，并且认为这是个很棒的点子。

So, you know, I'm sure I'm sure they wanted to to get him to go start that, and thought it was a great idea.

但与Calendly之类的众多工具不同，它并不依赖链接或人工协调。

But unlike a bunch of tools like Calendly, basically what it is doing, it's not relying on links or kind of manual coordination.

相反，它们使用AI代理直接相互协商以安排会议。

Instead, they have AI agents that are negotiating directly with each other to schedule meetings.

这有点像我最近在X上看到的一个梗或帖子，有人说：‘我会让我的Claude联系你的Claude，然后给你回复。’

It's kind of like this meme I recently, or this post I recently saw on X, which someone said, I'll have my Claude contact your Claude and get back to you.

我觉得这挺搞笑的，因为今天这确实正在发生。

And I just thought that was kind of funny because today, like that was really what's happening.

这些AI代理拥有日历代理，会查看你的日历，找出合适的时间，而对方的代理也会做同样的事，它们来回协商，最终选定一个时间，而你们俩根本不需要亲自沟通。

These AI agents who just have like a calendar agent, and it's gonna go look at your calendar, and it's gonna find a good time, and theirs is also gonna do that, and they negotiate back and forth, and they pick a time, and neither of you actually have to talk about it.

说实话，谁喜欢发邮件说‘我周三到周四这段时间有空’，哪怕只是发个链接，对吧？

Which, to be fair, who likes having to send an email that's like, I'm available from this time to this time on Wednesday to Thursday, and even dropping a link, right?

这纯粹是在浪费精力和时间。

It's just energy and time.

所以，如果人们想约时间，让代理去处理，我觉得这很棒。

So, I think this is really cool if people want to, you know, book a time with each other, have the agents do it, that sounds great.

我们来开个会吧，他们说‘好的，听起来不错’，然后代理们自动搞定一切，这对我来说简直太棒了。

Let's get on a call, and they say, yep, that sounds good, and all of sudden the agents do everything else, that sounds phenomenal to me.

有趣的是，它们在做这一切时，还会考虑各种偏好、优先级、语气和灵活性，综合这些个性化因素，为你安排日程并搞定所有细节。

What's interesting is they do all of this, but they take into account a bunch of preferences, priorities, tone, flexibility, and they, you know, take all these custom customized stuff and pick times on the calendar and get it all figured out for you.

有趣的是，BlockKit 的智能体可以直接通过邮件或 Slack 调用，并且它们还经过了针对用户特定规则的训练，比如哪些会议可以调整、哪些不可协商，甚至如何根据邮件结尾判断其紧急程度。

And what's interesting is BlockKit's agents can be invoked directly on email or Slack, and they're also trained on user specific rules, like which meetings are movable, which ones are non negotiable, and even how to interrupt the urgency of an email based on its closing.

对我来说，这一切都非常有趣。

So all of that to me is very interesting.

目前已有超过 200 家公司正在使用它，包括 Brex、Gethr AI 以及多家顶级风投公司。

They're already being used by more than 200 companies, Brex, Gethr AI, and a bunch of other top VC firms are using it.

我认为这很好地展示了人工智能如何真正为我们节省时间，并重塑许多日常的工作流程。

I think this is a really good example of how AIs are actually gonna save us time, reshape a lot of daily workflows.

尽管它们今天还不会完全取代整个职位或角色，但我认为它们正在加速并自动化大量正在进行的工作。

And while they may not replace entire jobs today, entire roles completely today, I think they are speeding up and automating a lot of stuff that's happening.

今天我就分享到这里。

That is all that I have for you today.

如果你觉得这期内容有用，请确保在你收听播客的平台订阅我们，如果你还没做过的话，也恳请你为本节目留下评分和评论。

If you found this useful, make sure that you are subscribed wherever you get your podcasts, and I would really appreciate it if you could leave a rating and review on the show if you haven't already.

一如既往，请务必访问我自己的创业项目：aibox.ai。

As always, make sure to go check out my own startup, which is aibox.ai.

如果你想用无代码方式构建AI工具，喜欢通过对话来设计工具，只需告诉我你想构建什么，它就能帮你串联起来。

If you'd like to build AI tools, with no code, if you like to vibe code tools, can explain what you like to build and it will link together.

AI Box 会将不同的AI模型串联起来，自动填充提示词，并为你自动化整个流程。

AI Box will link together different AI models, fill in the prompts, and automate your processes for you.

你可以前往 aibox.ai 了解更多信息。

You can go check it out at aibox.ai.

我会在描述中留下链接。

I'll leave a link in the description.

非常感谢你的收听，我们下一期再见。

Thank you so much for tuning in, and I will catch you in the next episode.