What to Use the Latest AI Tools For — 文稿

M10:00

这周发布了太多新的AI和模型产品，我不得不多做一期节目才能把所有事情都讲清楚。

This week had so many new AI and model product releases that I've had to do not one, but two episodes just to capture everything that's been going on.

有意思的是，第二波发布的产品中，很多都体现出了未来方向的鲜明趋势。

What's interesting about this second set is that many of them show some pretty distinct and important trends about where we're headed.

比如，很明显我们和电脑的互动会越来越多地靠语音来进行。

For example, it is very clear that we are going to be managing more and more of our interactions with computers via our voice.

这不会一蹴而就，但现在像ChatGPT Live这样的工具已经开放给开发者，能力更强了，能更好区分谁在说话、识别背景噪音。

It won't happen all at once, but now that we have things like ChatGPT Live available to developers to build around with a significantly increased capability set, better ability to distinguish who's talking, background noise.

所以你会看到越来越多跟语音相关的应用。

You're just going to see more and more applications that involve voice.

这些发布还体现了另一个趋势，就是越来越倾向于复杂的模型架构，人们可以优化成本和够用的能力，而不是一味追求最高性能。

Another example of a trend shown off in these announcements is the continued push towards complex model architectures where people can optimize for cost and good enough capability as opposed to just always seeking the highest capability.

重点是，有时候新品发布看的是你能用它做什么，有时候则在于它们预示了我们即将做什么。

The point is that sometimes new product releases are about what you can do with them, and sometimes they're about what they say about what we're all going to be doing soon.

《AI Daily Brief》是一档每日更新的播客和视频节目，聚焦AI领域最重要的新闻和讨论。

The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

好的朋友们，在我们开始之前，先快速说几个通知。

All right, friends, quick announcements before we dive in.

M11:26

欢迎回到《AI Daily Brief》新闻精华版，大约5分钟带你了解每天所有AI新闻。

Welcome back to the AI Daily Brief Headlines Edition, all the daily AI news you need in around 5 minutes.

今天我们先来看Anthropic的一份新报告，详细介绍了他们检测和应对AI滥用的努力。

We kick off today with a new report from Anthropic that details their efforts to detect and counter AI misuse.

这份报告里有很多有意思的细节，要不是这周有更大的AI安全讨论，它本来应该会成为更热门的话题。

There were a lot of juicy nuggets in this thing, and were it not for the larger AI safety conversation happening this week, it probably would've been a bigger topic of conversation.

我来挑几个重点：Anthropic说他们阻止了来自阿里巴巴、DeepSeek和小米的主要蒸馏攻击。

Going through some of the highlights, Anthropic said that they had disrupted major distillation attacks from Alibaba, DeepSeek, and Xiaomi.

每家公司都用虚假账户网络来提取Anthropic模型中的推理链条，作为训练数据。

Each company used networks of fraudulent accounts to extract reasoning traces from Anthropic models for use as training data.

不过，Anthropic还声称他们发现DeepSeek和Kimi K3（即Moonshot）把用户请求转发给Claude，然后把回复返回给自己的用户。

However, Anthropic also claimed that they had detected both DeepSeek and Kimi K3 Meeker Moonshot routing requests to Claude and serving the responses to their users.

Anthropic写道：在10天内，Moonshot将近30万次用户请求转发给了Anthropic，其中绝大多数都路由到了Opus。

Anthropic wrote, in one instance, over a 10-day period, Moonshot relayed almost 300,000 customer requests to Anthropic, the vast majority of which were routed to Opus.

根据报告来看，Moonshot似乎不是用这种方法来伪装他们自己的模型，假装他们有更强的能力。

Based on the report, it doesn't seem like Moonshot was using this method to spoof their models, using an Anthropic backend to make them appear more capable.

相反，这看起来像是为了收集真实用户查询，用于他们的蒸馏流程。

Instead, it seems like this was a way to gather realistic user queries for their distillation pipeline.

不过，Anthropic指出，这种方法多次暴露了中国政府和企业用户的敏感信息。

Still, Anthropic notes that this method exposed sensitive information from Chinese government and corporate users on multiple occasions.

在生物学章节中，Anthropic讨论了5月份发生的一起事件，当时他们的分类器阻止了一份科研经费申请。

In the biology section, Anthropic discussed an incident in May when their classifier blocked work on a grant application for scientific funding.

这份申请讨论的是对一种叫基孔肯雅热的蚊子传播病毒的增益功能研究。

The application discussed gain-of-function research on a mosquito-borne virus called chikungunya.

Anthropic写道：因为基孔肯雅热在自然界中本来就存在，所以作为生物武器故意释放的话，很难和自然爆发区分开。

Anthropic wrote, because chikungunya circulates naturally, a deliberate release as part of a bioweapon would be difficult to distinguish from a natural outbreak.

该资助计划试图识别基孔肯雅病毒的增强突变，将其改造为感染性克隆，并在活体中选择毒力。

The grant sought to identify enhancing mutations of the chikungunya virus, engineer them into infectious clones, and select for virulence in vivo.

也就是说，病毒在反复感染活体动物时会变得越来越有害，研究人员每轮都保留致病性最强的变种。

In other words, the virus would become progressively more harmful as it repeatedly infected live animals, with researchers keeping the most disease-causing variants in each round.

Anthropic指出，这项研究在疫苗开发方面可能有正当用途，但他们担心这可能带有恶意，因为研究计划是在一家军事研究机构进行的。

Anthropic noted that this research could have innocent applications in vaccine development, but they were concerned it might have been malicious because research was intended to be conducted at a military research institute.

报告没有指明是哪个国家资助的这项研究，但指出在研究人员被禁止使用后，AI提示仍然通过灰色市场转售商继续出现。

The report did not identify which nation was sponsoring the research, but did note the AI prompts continued through gray market resellers after the researchers were banned.

报告还包含了几十个其他AI滥用的例子，包括俄罗斯行为者利用Claude进行间谍活动和宣传，中国、俄罗斯和也门行为者试图编写常规武器软件，以及大量网络攻击和金融诈骗尝试。

The report also contained dozens of other examples of AI misuse, including Russian actors using Claude for espionage and propaganda, Chinese, Russian, and Yemeni actors attempting to write software for conventional weaponry, and a host of attempted cyberattacks and financial scams.

Anthropic指出，这份报告是精心准备的，他们写道：'我们在这里分享的案例不是典型的滥用，而是我们迄今为止发现的最值得注意和新颖威胁活动的例子。'

Anthropic noted the report was put together with intention, writing, 'The cases we share here aren't typical misuse, but rather examples of the most notable and novel threat activity we've identified to date.'

我们发布这份工作是因为我们认为有责任披露对我们服务的恶意滥用。

We're publishing this work because we believe we have a responsibility to disclose malicious misuse of our services.

随着模型能力越来越强，风险也会增加，除非AI开发者和社会的保卫者采取行动让它们更安全。

As models become increasingly capable, their risks will increase unless AI developers and society's defenders act to make them safer.

值得注意的是，这份报告完全聚焦于Haiku、Sonnet和Opus模型的滥用问题，只有一个案例涉及Fable和Mythos的蒸馏。

Notably, the report was entirely focused on misuse of Haiku, Sonnet, and Opus models, with only a single case of distillation related to Fable and Mythos.

说到对AI风险态度的转变，Sam Altman已经告诉员工他愿意放缓AI的发展节奏。

Now, speaking of the changing winds around AI risk, Sam Altman has told staff he's open to an AI slowdown.

在本周的一次全体会议上，Altman告诉团队，OpenAI可以控制新AI模型的开发速度。

During an all-hands meeting this week, Altman told his team that OpenAI could pace the development of new AI models.

不过，他说这很可能会与其他AI实验室协调进行，同时他也担心并非所有竞争对手都会同意。

However, he said this would likely happen in coordination with other AI labs and expressed concerns that not all rivals would agree.

这些言论是在关于放缓AI开发步伐的讨论浪潮之后发表的，这股浪潮甚至在Jacob Coxon本周早些时候的病毒式推文之前就开始了。

These comments come after a wave of discussion about slowing the pace of AI development that started even before Jacob Coxon's viral tweet earlier this week.

许多OpenAI员工在七月份签署了《领跑前沿》公开信，首席科学家Jacob Pachocki最近警告说，递归自我改进强化了需要放缓的理由。

Many OpenAI staffers signed on to the Pacing the Frontier open letter in July, and chief scientist Jacob Pachocki recently warned that recursive self-improvement amplified the case for a slowdown.

在上周末发布的一篇博文中，Pachocki写道，他希望行业协调以及——引用他的原话——“自愿放缓”能够成为常态，直到建立起共同的安全标准。

In a blog post published last weekend, Pachocki wrote that he hopes for industry coordination and, quote, voluntary slowdowns to become commonplace until shared safety bars are established.

这似乎是Altman第一次表示愿意放缓脚步。

This appears to be the first time that Altman has indicated a willingness to slow down.

与他在Anthropic的同行Dario Amodei不同，他没有签署《领跑前沿》公开信。

Unlike his counterpart Dario Amodei at Anthropic, he did not sign the Pacing the Frontier open letter.

而且直到最近，他在公开声明中一直不置可否。

And had been noncommittal in public statements until very recently.

周二，在评论OpenAI解决了一个千禧年大奖难题时，Altman写道：“现在世界已经拥有了极其强大的模型。”

On Tuesday, remarking on OpenAI finding a solution to a Millennium Prize problem, Altman wrote, 'the world has extremely capable models now.'

我没预料到这么重大的结果会这么快发生。

I did not expect a result of this magnitude to happen so soon.

我们一直在讨论需要控制进度以确保安全。

We've been talking a lot about the need to pace progress to ensure safety.

对我而言，这是迄今为止最有力的证据，表明这件事迫在眉睫。

For me, this is the strongest evidence yet of the urgency.

再加上本周早些时候AI安全研究员Paul Christiano加入OpenAI董事会，Altman的言论可能表明行业协同放缓正变得越来越可能。

Alongside the addition of AI safety researcher Paul Christiano to the board of OpenAI earlier this week, Altman's comments could indicate a coordinated industry slowdown is becoming more likely.

另外，在OpenAI的新闻中，该公司似乎遇到了算力瓶颈，因此将暂停高端订阅服务。

Now, also in OpenAI news, the company appears to have hit a compute wall and will pause top-end subscriptions as a result.

我们之前在Anthropic遇到所有麻烦时就讨论过，这将会成为每个人都要面临的问题，而现在显然已经来了。

We talked about this back when Anthropic was having all of its trouble, that this was going to be a problem coming for everyone, and it is apparently now here.

在GPT-6 Astra发布后不久，OpenAI的产品负责人Thibault Sotiao发帖说：“Astra的需求真的是前所未有的。”

Now, shortly after the release of GPT-6 Astra, OpenAI product leader Thibault Sotiao posted, 'demand for Astra is really unprecedented.'

我们正想尽一切办法来应对需求，但我以前从没见过这种情况，而我们之前也经历过非常快速的增长。

We're pulling all the levers possible to sustain the demand, but I've not seen anything like it until now, and we went through very steep growth before.

优先事项始终是为现有用户保持优质服务，但如果这种情况持续下去，我们可能不得不暂时暂停新的Pro订阅。

Priority will always be to keep excellent service for existing users, but we might have to pause new Pro subscriptions for a bit if this continues.

Sam Altman转发了这条消息并说：“这会很糟糕，但我们会优先为客户提供优质服务，直到我们重新掌控局面。”

Sam Altman reposted that saying, 'this would suck, but we will prioritize great service for customers until we can get back on top of things.'

嗯，看来他们现在已经无计可施了，Thibault周四又发了帖子。

Well, it seems like they have now run out of levers to pull with Thibault posting on Thursday.

为了确保我们现有用户能获得极佳的体验并持续使用Astra，我们将暂停200美元Pro计划的订阅。

To make sure our current users have an incredible experience and continued access to Astra, we are going to pause subscriptions to our $200 Pro plan.

这个计划对我们的系统压力最大，我们想采取最小的影响步骤，同时继续提供尽可能广泛的访问权限。

These put the most strain on our systems, and we wanted to take the smallest step that allows us to continue giving the broadest access possible.

所有其他计划和API仍然可用。

All other plans and the API remain available.

现有账户不受影响，我们正在尽快增加更多容量。

There is no impact to existing accounts, and we are working on adding more capacity as fast as we can.

Matt Schumer直截了当地写道：“补贴代币的时代正在结束。”

Stating it bluntly, Matt Schumer wrote, 'the era of subsidized tokens is ending.'

做好准备吧。

Prepare accordingly.

而Alex Barish认为200美元的计划不会恢复了。

And Alex Barish thinks the $200 plan will not be returning.

他们会发一篇关于规模适配的帖子，然后告诉我们，200美元计划的用户之间存在巨大差异：大多数用户其实更适合100美元的计划，只有一小部分用户一直在不断刷上限。

We will get a post talking about right-size fit, and they will tell us there was a large split between people on the $200 plan where most would have been better off on the $100 plan and a small percentage of users that were constantly maxing out limits.

说到算力，微软在经历了最近几年的算力短缺之后，计划把数据中心容量扩大到现在的三倍。

Now, speaking of compute, Microsoft plans to triple their data center capacity after battling their own compute shortages over recent years.

彭博社援引消息人士的话说，微软在新计划下全面加速了AI基础设施建设。

Sources told Bloomberg that Microsoft has turbocharged their AI buildout strategy under a new plan.

他们现在的目标是到2032年实现38吉瓦的全球容量，相比目前的12吉瓦大幅提升。

They now intend to have 38 gigawatts of global capacity by 2032, up from their current 12 gigawatts.

目前他们的算力中只有2吉瓦专门用于AI计算，但未来他们打算把这一比例提高到总容量的三分之一。

Only 2 gigawatts of their current fleet is dedicated to AI compute, but they will now aim to grow the share to a third of total capacity.

此外，该计划还会增加CPU来支持代理型工作流。

In addition, the plan will see them add CPUs to power agentic workflows.

重要的是，在整个AI建设过程中，微软是所有超大规模云服务商中最保守的一家。

Importantly, throughout the AI buildout, Microsoft has been the most conservative of the hyperscalers.

众所周知，他们在2025年初缩减了计划，取消了一些大规模租赁，这导致了市场的重大回调。

They famously scaled back plans in early 2025, canceling some large-scale leases, which contributed to a major market pullback.

但在过去一年里，微软多次表示算力限制正在制约增长，导致他们不得不拒绝客户。

But over the past year, Microsoft has repeatedly said that compute constraints are limiting growth and causing them to turn away customers.

微软的新计划表明，AI基础设施远未过剩，推动这些建设的公司预计还需要至少五年高强度的资本支出才能满足需求。

Microsoft's new plan reinforces that AI infrastructure is nowhere near overbuilt, and the companies driving the buildout expect at least 5 more years of elevated CapEx spend before they can meet demand.

说到持续加速的建设，英伟达CEO黄仁勋在高盛的一次会议上坚持预测增长率将达到70%。

Speaking of the buildout continuing apace, Nvidia CEO Jensen Huang has doubled down on forecasts of 70% growth at a Goldman Sachs conference.

黄仁勋在周四的发言中说，增长只受限于供应链，他评论说，尽管我们的需求远大于70%，但我们的供应能力让我们有信心实现70%的增长。

Speaking on Thursday, Jensen said that growth is limited only by the supply chain, commenting, even though our demand is much greater than 70%, our supply allows us to confidently deliver 70%.

黄仁勋向来不吝于自我宣传，但这是他过去一年来首次谈论前瞻性预测，说明他信心更足了。

Huang is never shy about self-promotion, but this is the first time he's discussed forward estimates over the past year, suggesting an added layer of confidence.

面对一群投资者，黄仁勋解释说，他们中的很多人都没跟上英伟达产品近年来的演变。

Before the crowd of investors, Huang explained that many of them haven't kept up with how Nvidia's product has evolved over recent years.

他说，大多数人都以为英伟达只是造芯片，但实际上，你需要用飞机才能运走我们造的东西。

He said, most people think Nvidia builds a chip, I mean, you need airplanes to ship what we build.

现在一块GPU不是399美元。

One GPU now is not $399.

而是850万美元。

It's $8.5 million.

那是一块GPU，通过NVLink全部连接起来，有两百万个零件，对吧？

That's one GPU, all connected with NVLink, 2 million parts, right?

功耗25万千瓦。

250,000 Kilowatts.

这就是一块GPU，而且我们卖出去好几千块。

That's a GPU and we ship thousands of them.

黄仁勋还补充说，他们机架级系统的订单目前以每月27%的惊人速度增长。

Huang added that orders of their rack-scale systems are currently growing an astonishing 27% month over month.

总的来说，黄仁勋试图传达的核心思想是，尽管每隔一周就有关于新竞争对手的头条新闻，但英伟达仍然处于AI建设的中心。

And overall, the core idea that Huang tried to communicate was that Nvidia is still at the center of the AI buildout despite headlines about new competition every other week.

不过，他们还得应对一个新情况，那就是美国司法部正在调查他们与Groq之间价值200亿美元的交易，以确定这是否是为了规避反垄断法而设计的结构。

Still, one new thing they have to contend with is the fact that the DOJ is investigating their $20 billion Groq deal to determine whether it was structured to circumvent antitrust laws.

这笔交易在去年12月宣布时，芯片初创公司Groq将其描述为非排他性的许可协议，允许英伟达使用他们的技术集成到未来产品中。

When the deal was announced in December, chipmaking startup Groq framed it as a non-exclusive licensing agreement that gave Nvidia access to their technology for integration into future products.

作为交易的一部分，CEO John Ross和COO Sunny Madra也加入了英伟达。

As part of the deal, CEO John Ross and COO Sunny Madra also joined Nvidia.

《纽约时报》报道说，司法部在这笔交易宣布后不久就开始调查了。

The New York Times reports that the DOJ has been investigating the deal since shortly after it was announced.

这笔交易是近年来AI行业众多“非收购”交易之一，包括谷歌与Character AI和Windsurf的交易，以及Meta与Scale AI的交易。

The deal was one of many not acquisitions in the AI industry over recent years, including Google's deals with Character AI and Windsurf, as well as Meta's deal with Scale AI.

通过把这些交易设计成非排他性的许可协议，它们不会自动触发美国联邦贸易委员会的审查，也不会像传统并购那样被阻止。

By structuring the deals as non-exclusive licensing agreements, they don't trigger automatic review by the FTC and can't be blocked in the same way as traditional mergers and acquisitions.

不过，这种结构在二月份受到了批评，多位参议员呼吁美国联邦贸易委员会和司法部调查这个具体问题。

Still, that structure came under fire in February, with a number of senators calling on the FTC and the DOJ to investigate this particular issue.

参议员们在信中写道，这些交易“实际上起到了并购的作用”。

In a letter, the senators wrote that the deals, quote, function as de facto mergers.

它们让公司能够整合人才、信息和资源，同时却显然试图绕过通常适用于并购的审查。

Allowing the companies to consolidate talent, information, and resources, all while apparently attempting to bypass the scrutiny typically applied to mergers and acquisitions.

现阶段这还只是调查，不清楚会不会有结果，但自从这些交易开始发生，这种调查迟早是逃不掉的。

Now, at this stage, it is just an investigation, and it's unclear whether anything will come of it, but when these deals started happening, it was always inevitable that this sort of investigation was going to happen.

最后今天，Meta 的新个人助手 Muse 似乎挺受消费者欢迎的，而且不仅如此，它在华尔街也绝对是个大热门。

Lastly today, Meta's new personal agent Muse seems to be finding some decent traction with consumers, and beyond that, it is certainly a big hit on Wall Street.

Sensor Tower 的新数据显示，这个应用在美国 iOS App Store 上线当天被下载了 83000 次。

New data from Sensor Tower shows the agent was downloaded 83,000 times on launch day in the iOS App Store for the US.

这个数字对 Meta 的新产品来说其实不算特别亮眼。

Now, that's not particularly impressive for a new Meta product.

Threads 上线当天有 430 万次下载，而 Meta AI 应用则有 10.8 万次下载。

Threads achieved 4.3 million downloads on launch day, while the Meta AI app reached 108,000 downloads.

不过，这个首发表现也足以让 Muse 冲到 App Store 第二名。

Still, the debut was good enough to push Muse to second place in the App Store.

但在华尔街，Muse 被视为 Meta 在 AI 领域依然有竞争力的证据。

However, on Wall Street, Muse is being viewed as evidence that Meta can still compete in AI.

发布当天股价上涨了 6%，而且基本上保住了涨幅。

The stock jumped 6% on release day and is largely holding on to the gains.

周四，摩根大通的分析师将这只股票评级上调为买入，认为 Muse 预示着更多进展。

On Thursday, JPMorgan analysts upgraded the stock to a buy, viewing Muse as a sign of more to come.

他们在一份研究报告中写道，Meta 已经做好充分准备，能为其大约 40 亿用户提供面向消费者的 AI 产品，而规模化分发是一个重要的竞争优势。

In a research note, they wrote that Meta is well-positioned to deliver consumer-driven AI products to its base of around 4 billion users, and that scaled distribution is a significant competitive advantage.

Meta 目前还处于发布前沿模型和除广告之外 AI 驱动产品的早期阶段，所以仍有很大的上涨潜力。

There's still meaningful upside potential as Meta is in the early stages of releasing frontier models and AI-driven products beyond advertising.

对一个重要的新产品类别来说，这是不错的进展，不过，今天的头条新闻就到这里。

Good progress for what is an important new product category, but for now, that is going to do it for the headlines.

接下来，是正片环节。

Next up, the main episode.

M113:02

到了这个阶段，问题已经不是企业是否在积极使用 AI 了。

At this point, it's no longer a question of whether companies are actively using AI.

但能不能用好它，那又是另一回事了。

Using it well, on the other hand, is a whole different story.

不过，Robots and Pencils 这家公司是我可以拿出来举例的，它真的是为这个时代而生的。

Robots and Pencils, though, is a company that I can point to that is actually built for this time.

他们是一家应用型 AI 工程公司，直接与客户合作，解决对业务真正重要的问题，而不是只停留在幻灯片里的实验。

They're an applied AI engineering firm working directly with clients on problems that matter to the business, not experiments that live in a slide deck.

每个项目都是从客户实际需要的结果出发，反过来倒推步骤。

Every engagement starts by working backwards from the outcome a client actually needs.

如果你想在这个领域里分辨出真正的 AI 工程和噪音，这才是关键区别。

If you're trying to tell real AI engineering apart from noise in this space, that's the difference maker.

M114:13

欢迎回到《AI 每日简报》。

Welcome back to the AI Daily Brief.

今天我们又回到实用性话题，因为即便最新一轮关于 AI 风险的讨论已经引爆主流，但那些现在就能用得上、而且基本没人说它们有实际危险的新产品，还在不断推出，创造新的机会。

Today we are back to the practical because even as the latest round of the AI risk conversation has exploded into the mainstream, new products that are useful for you right now and which basically no one is arguing are actively dangerous have continued to come out and create new opportunities.

所以今天我们来讲讲，最新的 AI 工具可以用来做什么。

And so today we're going to talk through what to use the latest AI tools for.

我会介绍过去几天里大概八九款新产品发布。

I'll be going through 8 or 9 product launches from the last couple of days.

快速说一下这些产品是什么、谁该考虑使用、以及应该怎么用它们来做什么。

With quick notes on what the launch is, who should be considering using it, and what they should be thinking about using them to do.

首先，OpenAI 在 API 中发布了 GPT-4o Live。

First up, OpenAI has released GPT-4o Live in the API.

这就是 OpenAI 在七月份推出的那个语音模型，还记得那个让人印象深刻的奶奶视频吗？

This is the voice model that OpenAI released back in July with that memorable grandma's video.

这是一个实时语音模型，也就是说它可以同时听和说，不需要轮流来。

This is a live voice model, meaning it can talk and listen at the same time without needing to take turns.

它在嘈杂环境和应对打断方面表现更好，而且其架构可以把任务交给后台的推理模型处理。

It boasts better handling of noisy environments and interruptions, as well as architecture that hands off tasks to a backend reasoning model.

这样一来，模型可以在后台完成任务，同时继续对话。

This allows the model to complete tasks in the background while still carrying on the conversation.

API 发布同时还附带了一个很酷的演示视频，展示了这个模型对开发者来说为什么有用。

The API release came with another nifty launch video demonstrating why this model could be useful for developers.

OpenAI 展示了这个模型的能力，即使在用户几英尺外有人运球，它也能检测到语音。

OpenAI showed off the model's ability to detect voice even with a basketball being dribbled a few feet away from the user.

它还展示了该模型可以在进行对话的同时控制机器人和显示器，展示了同时使用多种工具的能力。

It also shows the model can control a robot and a display while carrying on the conversation, demonstrating multiple simultaneous tool use.

该模型现已可用，音频模型每分钟收费 0.05 美元，后端模型则额外按标准 API 定价收费。

The model is now available for $0.05 per minute for the audio model and standard API pricing for the backend model on top.

你现在可以在我们已向 ChatGPT 上十亿用户推出的同一语音系统之上进行开发。

You can now build on top of the same voice system that we shipped to a billion users in ChatGPT.

全双工系统加上可靠的工具调用，有太多有趣的应用了。

So many fun applications of a full duplex system with solid tool calling.

在这之后回到纯文本体验，感觉比我想象的更难。

Going back to text-only experiences after this feels harder than I had expected.

这实际上是我期待已久的发布。

This is actually the release I've been waiting for for a long time.

黑客的梦想成真了。

A hacker's dream come true.

你现在可以在 Jony Ive 和 OpenAI 的设备发布之前就造出它来。

You can now build the Jony Ive OpenAI device before they even release it.

我打算在 Amazon 上买一个小型开发屏幕和麦克风。

Going to be buying a small dev screen/mic on Amazon.

把它插到我的电脑上。

Plug it into my computer.

让 Astra 破解它，把 LiveOne API 装上去。

Tell Astra to hack it to put LiveOne API on it.

现在无论你去哪里，你都可以直接跟设备说话，让它为你干活。

Now anywhere you go, you can just talk to the device and have it do work for you.

基本上就是你走到哪儿，ChatGPT 的语音就跟到哪儿。

Basically ChatGPT's voice everywhere you go.

给它下达指令。

Give it commands.

它就开始在你家里的电脑上干活。

It starts doing work on your home computer.

你现在基本上可以打造自己的个人助理了。

You can basically now build your own personal assistant.

在 Amazon 上有很多便宜的方法，用小型开发设备就能实现。

Tons of cheap ways to do this on Amazon with small dev devices.

直接搜就行了。

Just search for them.

像这样的事情让 AI 变得有趣，我觉得我们对乐趣的关注还不够。

Things like this just make AI fun, and I don't think we focus enough on having fun.

那么说到谁应该使用这个产品，在很多方面，它都将是许多现有语音体验的严格升级。

So when it comes to who should be using this product, in a lot of ways this is just going to be a strict upgrade to many existing voice-based experiences.

显然，这对客服主管、联络中心以及那些因为忙于工作而错过很多电话的小型服务企业来说将非常有价值。

Obviously, this is going to be extremely valuable for customer support leaders, contact centers, small service businesses that have a lot of missed calls because they can't be talking on the phone while they're doing their work.

但我也认为，人们应该更多地尝试将语音引入新的体验类型。

But I also think that people should be experimenting a lot more with bringing voice into new types of experiences.

我总的感觉是，随着语音技术的发展，人们会越来越多地从打字与设备交互转向直接与设备对话。

My general sense is that as voice technology gets better, people are going to shift more and more from typing to their devices to talking to their devices.

这种转变已经在发生了。

This is a shift that's already been happening.

这是代际性的变化。

It's generational.

老实说，在某种程度上，这种转变没有更普遍发生的原因仅仅是因为像 Siri 这样的原生语音识别技术历来表现太差。

And honestly, to some extent, the only reason it hasn't happened more is that the native voice recognition technology in places like Siri have been so bad historically.

随着越来越多不同类型的用户对语音感到习惯，很多可能性就打开了。

As more different types of users get comfortable with voice, it opens up a lot of possibilities.

例如，对于 B2B 销售和集客营销团队来说，很多体验都是从强迫人们预约演示开始的，但也许你可以让感兴趣的潜在客户直接说出他们的需求，并在进入演示阶段之前就获得相关的产品解释。

For example, for B2B sales and inbound marketing teams, so many of those experiences start with forcing people to schedule a demo, but maybe instead you could just let interested prospects talk through their needs and receive a relevant product explanation before they even get to the demo stage.

显然，这里对于语言学习和更广泛的辅导来说有大量的机会。

It obviously feels like there's a ton of opportunity here for language learning and for tutoring more broadly.

这在构建逼真实践和模拟场景方面创造了一些非常有意思的机会，让人们可以练习工作中会遇到的特定互动类型。

It creates some really interesting opportunities for realistic practice and simulation type of scenarios to the extent you're building experiences where people are practicing a particular type of interaction that's going to come up in their jobs.

而这些只是冰山一角。

And these are just the tip of the iceberg.

有一家已经基于GPT Live构建产品的公司是Cognition，他们刚刚发布了Devin Voice，推广语非常简单。

One company that has already built on GPT Live is Cognition, who have just announced Devin Voice with a very simple pitch.

你最爱的AI软件工程师现在有座机了。

Your favorite AI software engineer just got a landline.

你一说，Devin就搞定。

You say it, Devin ships it.

这个功能在语音层面由OpenAI的GPT Live驱动，后台由Cognition的新模型Sui 2支持。

The feature is powered by OpenAI's GPT Live on the voice layer and Cognition's new Sui 2 model on the backend.

你可能也猜到了，我们马上就会聊到Sui 2。

And as you might have guessed, we will get to Sui 2 in just a minute.

这类语音编码的更新，跟回合制语音模式相比，是巨大的生活质量提升。

Now, these sort of updates while using voice for coding, are huge quality of life upgrades over turn-based voice modes.

可以打断、唠叨、纠正自己，这让整个体验自然多了，也更本土化。

The ability to interrupt, ramble, and correct yourself just makes the experience a lot more natural and native.

Cognition团队的Nader Davit说：“我90%的日常现在都用语音完成了。”

Nader Davit from the Cognition team says, 90% of my day-to-day work is now done with my voice.

所以把它变成Devin的头等公民功能，完全合情合理。

It only makes sense to make it a first-class citizen in Devin.

OpenAI的Kova写道：“语音是与AI互动最神奇的方式。”

Writes OpenAI's Kova, voice is the most magical way to interact with AI.

未来，把一个想法变成现实，从说出来开始——这真的很美。

There's something beautiful about a future where bringing an idea to life starts with saying it out loud.

所以这个案例里，谁该用、该用来做什么，其实更简单明了。

So in this case, the who should be using this and what they should be using it for is a little bit more simple.

该用这个的人当然是Devin用户，他们该做的就是把行为从“打字来提示和互动”转成“说话来提示和互动”。

The people who should be using this are, of course, Devin users, and what they should be doing with it is shifting their behavior from typing to prompt and interact to talking to prompt and interact.

即使你不是Devin用户，我也想借这个机会再次恳请你：如果还没试过语音，就赶紧开始用吧。

And even if you are not using Devin, I'll use this chance once again to implore you to start using voice if you haven't yet.

如果你是ChatGPT用户，一个非常好用的方式就是原生ChatGPT语音功能。

A really great way to do that is the native ChatGPT voice feature if you are a ChatGPT user.

但就算你用别的工具和LLM，我也强烈建议在电脑上装个像Whisper Flow这样的东西，然后开始有意识地试着把行为往语音转移。

But even if you're using other tools and other LLMs, I highly recommend setting up something like Whisper Flow on your computer and starting to very intentionally try to shift your behavior over to voice.

一旦你开始用，我猜你不仅会发现速度提高了，还会发现你可以在做其他事——比如散步——的同时，完成更多工作。

Once you do, my guess is that you'll not only find your speed improved, but you'll also find you can do more of your work while doing other things like taking a walk.

看吧，从最初这几个发布就能看出，整个行业已经在这条路上推你了。

Look, it's pretty clear from these first couple announcements that the industry is pushing you there anyways.

不如就顺势抓住这个新机会。

You might as well go surf this new opportunity.

接下来要讨论的新发布，就是之前提到的Sui 2。

Next new launch to discuss is the previously mentioned Sui 2.

这是Cognition最新的内部编码模型，也是他们模型的第一个重大迭代，替代了Sui 1.7。

It is Cognition's latest in-house coding model and their first major iteration of their model, which replaces Sui 1.7.

这个版本是基于KIMI-K3的后训练版本，专门为编码优化，不做别的。

This version is a post-trained version of KIMI-K3 optimized for coding and nothing else.

过去一年，我们从Cognition和Cursor那看到了好几代这类模型。

We've seen a few iterations of these types of models from Cognition and Cursor over the past year.

通常，它们的定位是在便宜的价格点上提供“足够好”的模型，高效处理日常任务。

Generally, they've been aimed at delivering a good enough model at a cheap price point to efficiently handle everyday tasks.

实际上，Cursor和Cognition都有点超前地意识到：我们在能力上已经达到了某个转折点，同时AI承担的任务量也大大增加，这时不仅要优化纯能力，还要优化效率，这开始变得真的很重要。

In fact, Cursor and Cognition were both a little ahead of the curve in understanding that we had reached a certain inflection point in capability as well as a magnitude of tasks that are being undertaken by AI where optimizing not just for pure capability but also for efficiency was going to start to really matter.

有意思的是，从Sui 2来看，最近几个月“足够好”的定义已经大大提高了。

What's interesting is that with Sui 2, it's clear that the definition of good enough has dramatically increased in recent months.

鉴于这个模型基于Kimi K3，它接近前沿性能也就不奇怪了。

Given that the model is built on Kimi K3, it unsurprisingly offers near frontier performance.

举个例子，模型在Frontier Code 1.1上得了50%，略高于GPT-5.6 SOL，略低于Fable 5.1，但成本降低了64%。

For example, the model scored 50% on Frontier Code 1.1, putting it slightly ahead of GPT-5.6 SOL and just behind Fable 5.1, but at a 64% reduction in cost.

这个模型甚至比基于更小的Kimi K 2.7构建的Sui 1.7还要便宜。

The model is even cheaper than Sui 1.7, which was built on the much smaller Kimi K 2.7.

Cognition还增加了努力级别，这样你可以降低算力消耗来进一步省钱。

Cognition has also added effort levels so you can turn down the juice for even cheaper usage.

现在，这个时间点有趣的一点是，尽管Cursor越来越深入整合到SpaceX AI中，导致了一些后果，比如OpenAI切断了Cursor内部对其模型的访问，但Cognition表示他们想保持独立，刚刚以480亿美元的估值融资了20亿美元。

Now, one of the interesting things about the timing for this is that even as Cursor gets further integrated into SpaceX AI, leading to consequences like OpenAI cutting off access to its models inside Cursor, Cognition has signaled that they want to stay independent, just raising $2 billion at a $48 billion valuation.

我猜这意味着那些看重独立性、希望不受限制地选择模型的开发者，可能会去尝试Cognition的产品，比如Devin。

My guess is that that means some number of developers who prize independence and the ability to select the models they want without constraint might go check out Cognition's products like Devin.

而这次新模型的发布，当然也为大家提供了一个好时机，去看看Devin桌面版和命令行中这个更便宜的原生模型能做什么。

And this new model release, of course, creates a good context to go see what's possible with this cheaper native model that lives inside the Devin desktop and CLI.

说到底，谁该试试这个、用来做什么呢？任何正在尝试构建完整模型栈、希望更好地匹配能力与需求以控制成本的人，可能都会对这个感兴趣。

Basically, when it comes to who should try this and what they should use it for, anyone who is already in the midst of trying to build a complete model stack that can better align capability with need to keep costs down might be interested in checking this one out.

说到控制成本的模型，我们接下来的发布来自DeepSeek，他们发布了新的Flash模型版本——V4.1 Flash。

And speaking of models to keep costs down, our next release comes from DeepSeek, who have dropped a new version of their Flash model with the release of v4.1 Flash.

这个模型有5520亿参数，专为高延迟任务设计。

The model has 552 billion parameters and is designed for high-latency work.

基准测试成绩看起来很不错，在DeepSWE上得了74.2分，与GPT-5 6SOL和Opus 5基本持平。

The benchmarks look pretty solid with a score of 74.2 on DeepSWE, which puts it right in line with GPT-5 6SOL and Opus 5.

在TerminalBench 4.0上（这个基准目前还不太容易被刷分），V4.1 Flash得分31.2%，而SOL是39.9%，Opus是74%。

On TerminalBench 4.0, which is so far much less benchmaxable, V4.1 Flash scored 31.2%, compared to 39.9% for SOL and 74% for Opus.

不过，这个模型定价在市场的低端，每百万输入tokens收费0.30美元，每百万输出tokens收费1.20美元。

Still, the model is priced at the bottom of the market, charging $0.30 per million input tokens and $1.20 per million output tokens.

有趣的是，Artificial Analysis发现，Flash模型在智能指数上的表现实际上超过了DeepSeek全尺寸的Pro模型，而成本只有后者的四分之一。

Interestingly, Artificial Analysis found that the Flash model actually outperforms DeepSeek's full-size Pro model on the Intelligence Index at a quarter of the cost.

这是Artificial Analysis新版本4.3的指数，其中最高分是Fable 5.1和GPT-6 Astra的53分，而DeepSeek 4.1的总体得分是40，离前沿还有一段距离。

Now, this is Artificial Analysis's new version 4.3 index, where the top score is Fable 5.1 and GPT-6 Astra at 53, and DeepSeek 4.1 is still a way off the frontier with an overall score of 40.

这大致与GPT-5.6 Luna和Gemini 3.8 Flash相当，再次说明这可能是一个适合那些想构建更全面架构的团队的模型。

That puts it roughly in line with GPT-5.6 Luna and Gemini 3.8 Flash, and again means that this is likely a model for teams that are looking to build a more holistic architecture.

特别是对于有基础但高频次、计算密集型任务的团队来说，这个模型很可能值得一试。

Especially for teams that have basic but highly recurring compute-intensive types of tasks, this is one that's likely going to be worth a test.

接下来，我们再次回到OpenAI的世界，看看他们推出的“小企业插件集合”。

Next, we move once again back over into OpenAI world, with the introduction of the Small Business Plugin Collection.

这个功能并不复杂。

Now, this is not particularly complex.

OpenAI把他们最有用的一些插件打包成了一个小企业套装。

OpenAI has collected some of their most useful plugins as a pack for small business.

里面包括Dropbox、HubSpot、Canva、Figma、Shopify、DocuSign、PayPal、QuickBooks、Stripe、Gusto、Slack、Wix、Mercury等等。

It includes things like Dropbox, HubSpot, Canva, Figma, Shopify, DocuSign, PayPal, QuickBooks, Stripe, Gusto, Slack, Wix, Mercury, et cetera.

这其实就是想让使用ChatGPT的小企业主更方便，把所有东西都放在一个容易找到的地方。

Now, this is really just trying to make things a little bit simpler for small business owners who are using ChatGPT by putting everything all in one easily accessible place.

不过有意思的是，一些人开始意识到，在智能体时代，插件的整个概念有点不一样了。

What's interesting though is that some people are recognizing that the entire idea of plugins gets a little bit different when it comes to the agentic era.

针对OpenAI总裁Greg Brockman关于插件集合的帖子，Abdul Wasi写道：“v1插件之所以失败，是因为它们只是聊天框后面的API包装器。”

Responding to OpenAI President Greg Brockman posting about the plugin collection, Abdul Wasi writes, plugins v1 died because they were API wrappers behind a chatbox.

这次，智能体可以运行整个工作流程。

This time the agent can run the whole workflow.

不过Abdul也指出，未解决的部分依然是分发问题。

However, Abdul also points out the unsolved part is still distribution.

小企业主不会去浏览插件目录。

SMB owners don't browse plugin directories.

但既然现在这些插件被打包好了，这正好是这类用户应该去看看的东西。

Although given that they are now collected, that is exactly the type of user who should check this out.

如果你是小企业主，或者团队非常小，而且已经在使用这些应用，那么看看ChatGPT中的原生集成能做什么，可能会显著提升你的使用体验。

If you are a small business owner or have a very small team and you are using these apps already, seeing what the sort of native integrations in ChatGPT can do could be a significant upgrade to your experience.

OpenAI还推出了ChatGPT for Finance，对GPT-6 Astra进行了重大更新。

OpenAI has also released ChatGPT for Finance with a big update for GPT-6 Astra.

这个产品的新版本仍然是一套为金融专业人士构建的数据连接器和技能包，但已经针对OpenAI的新产品线进行了更新。

The new version of this product is still a bundle of data connectors and skills built for financial professionals, but it's been updated for OpenAI's new product lineup.

该产品现在位于ChatGPT Work中，并内置了对Daloopa、PitchBook、LSEG News和Crunchbase等优质数据源的访问权限。

The product now lives within ChatGPT Work and includes built-in access to premium data feeds including Daloopa, PitchBook, LSEG News, and Crunchbase.

这意味着用户不需要搞明白MCP连接，稍微配置一下就能直接上手。

This means that users don't need to figure out MCP connections and can get started right away with minimal configuration.

这个产品还整合了ChatGPT的企业级安全与治理控制，方便组织层面管理数据访问权限。

The product is also designed for ChatGPT's enterprise security and governance controls to allow for organizational-level control of data access.

OpenAI与摩根士丹利和Evercore合作进行重新设计，力求使其更贴合金融专业人士的需求。

OpenAI partnered with Morgan Stanley and Evercore for the redesign, trying to closely align it to the needs of financial professionals.

它专为构建财务模型、开展研究以及制作客户材料等任务而设计。

It's designed for tasks including building financial models, developing research, and building client materials.

在一次新闻发布会上，产品副总裁Nick Turley表示，我们实际上是在教ChatGPT像分析师一样进行研究，并像分析师一样为其结论提供支持。

And during a press briefing, VP of Product Nick Turley said, we're effectively teaching ChatGPT to research like an analyst and back up its conclusions like an analyst as well.

它原生使用GPT-6 Astra，OpenAI计划随着新模型的推出而持续更新。

It uses GPT-6 Astra natively, and OpenAI intends to keep it updated for new models as they become available.

OpenAI的Ryan Brewer写道，通过ChatGPT金融服务版，我们完成了对所需数据进行索引的艰巨工作，并使其易于被模型用于财务分析。

OpenAI's Ryan Brewer wrote, with ChatGPT for Financial Services, we've done the hard work of indexing the data that you need and making it easily accessible to the model for financial analysis.

我们创建了自定义图表体验、更好的引用功能、SEC文件查看器等更多功能。

We've created a custom charting experience, better citations, an SEC filing viewer, and much more.

Sundeep Srivastava写道，它不是一个披着金融外衣的通用聊天机器人。

Sundeep Srivastava writes, it's not a general chatbot with a finance skin.

它直接针对初级投行分析师一周的工作内容：公司研究和股权分析、杠杆收购建模和买家筛选、按照公司自有模板格式化的宣传册和客户演示文稿。

It's aimed squarely at what junior investment bankers spend their weeks doing: company research and equity analysis, LBO modeling and buyer screening, pitchbooks and client decks formatted to the firm's own templates.

现在，正如你可能预料到的，这引发了一些关于这些初级分析师是否会被取代的问题。

Now, as you might expect, this is opening up some questions about whether those junior associates are going to be replaced.

但我认为需要注意的是，认为高级银行家会仅仅因为AI代理在核心工作上变得更强，就自己动手做模型和幻灯片，这是不太现实的。

But I think it's important to note that it's not particularly realistic to think that high-level bankers are going to be making their own models and slide decks just because the agents have gotten much better at that core work.

仍然存在责任归属、迭代等问题，而这些事情更高级别的银行家并不想去做。

There's still questions of accountability, iteration, and all these things which more senior-level bankers are not going to want to do.

相反，我们在金融行业看到的是，初级银行家开始被要求展示AI熟练度。

Instead, what we're seeing in the finance industry is that junior bankers are starting to be required to demonstrate AI proficiency.

例如，瑞银最近开始要求未来初级银行家展示AI熟练度，作为招聘要求。

UBS, for example, recently started demanding that prospective junior bankers show proficiency in AI as a hiring requirement.

换句话说，我不认为这个产品会取代初级银行家。

In other words, I don't think this product is a replacement for junior bankers.

我认为它是初级银行家们的一个强力工具，能让他们在工作中更高效、更出色。

I think it is a power tool for junior bankers, which is going to allow them to be much more efficient and much better at their jobs.

OpenAI的最后一个产品是面向ChatGPT Work的新数据代理。

The last one from OpenAI is their new data agent for ChatGPT Work.

该代理专为处理组织内部的专有数据而设计，可与Amazon Redshift、Datadog、Google BigQuery、ClickHouse、Databricks、MongoDB和Snowflake等数据提供商连接。

The agent is designed for handling proprietary data within an organization with connections to data providers like Amazon Redshift, Datadog, Google BigQuery, ClickHouse, Databricks, MongoDB, and Snowflake.

OpenAI表示，这些功能可用于导入销售数据，并围绕销售转化率和留存率等核心指标生成洞察。

OpenAI says the features can be used to ingest sales data and generate insights around core metrics like sales conversions and retention.

目标是实现能够与数据对话并进行实际分析，而无需使用其他工具的承诺。

The goal is to realize the promise of being able to talk to your data and perform real analysis without needing to touch additional tools.

Tableau的首席产品官Souther Jones表示，将Tableau与ChatGPT Work连接，将可信的业务语义带入员工已有的工作环境，从而使得他们得到的答案基于团队所依赖的同一数据模型。

Souther Jones, the Chief Product Officer at Tableau, said, connecting Tableau with ChatGPT Work brings trusted business semantics into a place employees already work, so the answers they get are grounded in the same data model their teams rely on.

用户可以通过提问、探索证据，并利用内置设计和分析最佳实践，直接将新视图发布到Tableau，从而轻松将洞察转化为行动。

Users can easily transform insights into action by asking questions, exploring evidence, and publishing new views directly to Tableau using built-in design and analytics best practices.

退一步来看，如果你观察OpenAI这些功能更新的整体，它们都专注于针对特定类型的商业用户，更好地连接他们工作所需的工具和上下文，并通过更好的工具访问和上下文连接，使得将更多工作流程迁移到ChatGPT变得极其容易，而且说实话也更具吸引力。

Taking a step back then, if you look at the aggregate of these feature updates from OpenAI, they are all about honing in on specific types of business users, better connecting both the tools and context they need to do their work, and through that better tool access and better context connections, make it vastly easier and frankly more inviting to move more of their workflows into ChatGPT.

这并非一个新趋势，但我认为我们肯定会看到更多这样的发展。

This is not a new trend, but it's certainly one that I think we're gonna see a lot more of.

换句话说，就是垂直化，真正理解特定类型人群的工作方式。

In other words, verticalization that really understands how very specific types of people work.

另一个小例子是Grokbot，他们本周刚刚宣布，由于与Salesforce、HubSpot、Gong、Quake、Granola以及其他市场推广工具的原生集成，他们现在对销售团队更加强大了。

Another small example of that trend is Grokbot, who just announced this week that they're now more powerful for sales teams because of native integrations with Salesforce, HubSpot, Gong, Quake, Granola, and other go-to-market tools.

在此我们看到，垂直化、工具访问和上下文管理的趋势并非OpenAI独有，而是整个行业共同的发展方向。

And in this, we see the verticalization tool access and context management trends are not unique to OpenAI, but are of course shared across the industry.

在结束之前还有两个快速的要点。

2 More quick ones before we get out of here.

第一个是项目和光标。

The first is projects and cursor.

这个功能挺有意思的，虽然我们都已经习惯了在 ChatGPT 和 Claude 里用项目功能，但这个项目功能会带你到一些新的地方。

Now, this is a little interesting because while we're all used to using projects in ChatGPT and Claude, this feature takes it to slightly new places.

它不只是用来给某个项目积累上下文的地方，还被设计成可以容纳定时任务、长期运行的人工智能代理，以及任务编排。

Rather than just being a place to accumulate context around a particular project, this feature is also designed to house scheduled tasks, long-running agents, and orchestration.

换句话说，项目不仅仅是文件夹，它们实际上是一个真正的项目架构。

In other words, projects aren't just a folder, they are an actual project architecture.

Cursor 的项目标签页采用了类似于 Grokbot 的方法，让用户可以创建一个持久的代理，在需要时执行工作。

The projects tab in Cursor uses a similar approach to Grokbot, allowing users to spin up a persistent agent that performs work as required.

在它们的发布视频里，Cursor 说，它已经超越了回合制的对话模式。

In their launch video, Cursor said, it's graduated from turn-based chat.

你不需要事无巨细地管理每个代理，而是在和一个更自主的同事一起工作。

Instead of you micromanaging every agent, you're working with a much more autonomous colleague.

比如，如果你在开一个 PR 之后有一个标准工作流，那么项目就能自动搞清楚怎么做，并在你下次触发工作流时自动执行任务。

For example, if you have a standard workflow after opening a PR, the project can figure out how to do and carry out the tasks automatically the next time you trigger the workflow.

Cursor 说，测试人员使用项目后，合并的 PR 数量是之前的 6 倍，合并率提升了 30%。

Cursor said that testers have merged 6 times as many PRs using projects and merge rates have increased by 30%.

这个功能最明显的用户群体是开发者，但非开发者的知识工作者也值得关注其中的模式。

Now, this is one where the obvious users initially are going to be developers, but the patterns are worth paying attention to for non-developer knowledge workers as well.

Prasenjit Sarkar 写道，从按任务聊天到持久协调员的转变，在我看来，是 Cursor 项目里比单纯增加子代理数量更重要的转变。

Prasenjit Sarkar writes, the move from chat per task to persistent coordinator is, in my read, the more important shift in Cursor projects than the raw sub-agent count.

我之前看过的每个编码代理，工作方式都一样。

Every coding agent I've watched before this worked the same way.

打开一个会话，描述一个任务，代理执行，会话结束。

Open a session, describe a task, agent executes, session ends.

Cursor 项目把这个模式颠倒过来了。

Cursor projects inverts the model.

一个协调线程在项目的整个生命周期中保持打开状态。

One coordinator thread stays open for the life of a project.

这个协调线程不写代码，它负责规划、把任务委派给子代理，然后把结果带回来检查。

The coordinator doesn't write code, it plans, delegates to subagents, and brings results back to check.

正如 Pressing G 指出的，这并不是一个独一无二的架构。

Now, as Pressing G points out, this isn't a unique architecture.

你有 Cognition 的 Devin、OpenAI Codex 以及其他工具，但他认为持久线程和主动触发系统是不同的。

You've got Devin by Cognition, OpenAI Codex, and others, but argues that the persistent thread and the proactive trigger system are different.

持久线程模式，我觉得很有意思。

Now, the persistent thread pattern is one that I think is interesting.

我们很多构建项目都必须在不同的实例之间传递上下文，比如 Claude Code，因为一个实例的上下文窗口满了，你就得重新开始。

A lot of our build projects involved having to hand off context between different instances of something like Claude Code because the context window in one was filled and you needed to start fresh.

到了 2026 年，情况已经大不一样了。

That has changed a lot in 2026.

Kansar 描述了变化的内容。

Kansar describes what changed.

他写道，在 2025 年第四季度，我带头大力推动教模块工程师高级上下文工程，以便充分利用 Claude Code 和同类工具。

He wrote, in Q4 of 2025, I led a big push to teach block engineers advanced context engineering to get the most out of Claude Code and Co.

到了三月和四月，我们开始大规模转向 Codex。

In March and April, we started switching to Codex in large numbers.

有了 5.4 版本，尤其是 5.5 版本，它能解决其他工具解决不了的问题，构建其他工具构建不了的东西。

With 5.4 and especially 5.5, it could solve problems and build things that other tools could not.

但我最喜欢的是它的压缩能力有多强。

But my favorite was how good it was at compaction.

不需要那些花哨的子代理技巧和有意的压缩操作，你只需要一直往下进行就可以。

Instead of all the fancy subagent tricks and intentional compaction, you could just keep going and going.

我有一条线程，构建了一整套同步协议、服务器和客户端，跨越多个仓库和语言，还把它们部署了。

I had one thread that built a whole sync protocol, server, and client across multiple repos and languages and also deployed them.

从那以后，压缩功能变得越来越好。

Since then, compaction continues to get better and better.

单线程现在是一个常见话题，但我还是很惊讶，一条线程里能做这么多事，而且一点问题都没有。

Monothreads are a common topic now, but I'm still surprised by how much you can do in one thread without any issues.

我发现自己无论在 Claude 还是 Codex 里，都越来越多地使用这种模式了。

This is a pattern I found myself using a lot more in both Claude and Codex.

比如说，我拥有的AI Daily Brief网站相关的所有内容，基本上都在一个很长的线程里，这个线程做了大量工作，已经运行了很长时间，而且从来没有真正遇到过任何上下文问题。

I have, for example, everything related to the AI Daily Brief website pretty much in one long thread that has done a ton of work and has been going for a very long time and hasn't really ever run into any context issues.

所以提醒一下，有时候新产品或功能的推出，重点不在于你是否会使用那个特定功能，而在于它反映了整体趋势的走向。

So the reminder is that sometimes a new product or feature launch is less about whether you are going to use that particular feature and more about what it says about where things are headed in general.

总之，各位，以上就是对过去几天发布的产品和功能的快速梳理。

Anyways, folks, that is a quick rip through the products and features that have launched in the last couple of days.

我认为它们适合哪些人，可以用来做什么，希望你能在周末之前获得一些新灵感。

Who I think they're for, what I think they're useful to do, and hopefully you have some new ideas heading into the weekend.

那么，今天的AI Daily Brief就先到这里了。

For now, that's going to do it for today's AI Daily Brief.

一如既往感谢你的收听或收看，下次见，Peace。

Appreciate you listening or watching as always, and until next time, peace.

已剔除 3 处广告（点击展开查看）

M11:09广告 · 已剔除

First of all, thank you to today's sponsors, KPMG, Blitzy, Robots and Pencils, and HyperAgent.

For an ad-free version of the show, go to patreon.com/AIDailyBrief, or you can subscribe on Apple Podcasts.

And if you want to learn more about sponsoring the show, send us a note at [email protected].

M111:42广告 · 已剔除

A new study from KPMG and the University of Texas at Austin found that when people work with AI, similar skills don't guarantee similar outcomes.

Researchers studied more than 500 early-career professionals and found that the best performers consistently amplified the value of AI by guiding, evaluating, and refining its outputs.

These top performers, called AI Amplifiers, weren't defined by what they knew alone, but by how they worked with AI.

Learn more about what separates AI Amplifiers from everyone else at kpmg.com/us/aiamplifiers.

Blitzy's understanding of massive codebases unlocks autonomous security fixes, modernization, and new features.

So what happens when there's no legacy code at all?

Greenfield is supposed to be the easy part, clean slate, no technical debt.

But even greenfield moves at human speed, one sprint at a time.

Blitzy changes the unit of work from the developer to the project, autonomously planning, building, testing, and validating entire applications from scratch.

Hundreds of thousands of lines of production-ready code.

One Blitzy customer stood up a brand new application, 534,000 lines of code, compressing a 65-week roadmap into 2 weeks.

Another shipped an entire application with no frontend engineer.

Legacy or greenfield, the answer is the same: software at the speed of compute.

Build what's next at blitzy.com.

That's blitzy.com.

M113:29广告 · 已剔除

Head to robotsandpencils.com.

This episode of the AI Daily Brief is brought to you by HyperAgent, where you run fleets of agents your team can manage together.

Forget local agents and chat workflows waiting on your laptop to be prompted— HyperAgent deploys always-on agents in the cloud, doing real work across the tools your team already uses.

Marketing agents turn competitor moves into landing pages.

Sales agents enrich leads, draft emails, and update the CRM.

Ops agents chase the paperwork and track the budget.

Every agent has access to shared context and follows your rules about scope and approvals.

It's time you had agents that feel like teammates.

Hire yours at HyperAgent.

Get $100 in credits at hyperagent.com/AIDailyBrief.