The Most Useful New AI Features and Tools to Try — 文稿

M10:00

这些里面哪个最让你兴奋？

Which of these sound most exciting to you?

Claude 新增的能通过浏览器窗口帮你完成任务的能力？

A new ability for Claude to use a browser window to be able to do tasks for you?

ChatGPT 能够根据需要灵活切换临时模式的能力？

The ability for ChatGPT to effectively switch in and out of temporary mode as suits you?

一个新的语音模型，不仅能准确转录你说的内容，还能清理语音并理解你的意图？

A new voice model that not only can faithfully transcribe what you actually say, but even clean it up and get at your intent?

Google 推出的新视频模型，可控性大幅提升，能生成真正可用的视频？

A new video model from Google with massively more controllability for actually generating videos that you can use for real things?

或者一个完全不同的视频模型，生成视频的时间比观看它还要短？

Or a totally different video model that takes you less time to generate the video than it does to watch it?

这些只是本周发布的新功能、新模型和新工具中的一部分。

These are just some of the new features, models, and tools that released this week.

《AI Daily Brief》是一个每日播客和视频节目，关注 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.

好了朋友们，在深入之前先快速宣布几件事。

Alright friends, quick announcements before we dive in.

M11:18

首先，确认 Hugging Face 正在被收购，而且没错，买家就是 NVIDIA。

Kicking off, confirmation that Hugging Face is being sold and that yes, it is to NVIDIA.

《The Information》在周三晚间的财报电话会议后爆出消息，据他们透露，Nvidia 已同意以 129 亿美元收购 Hugging Face。

The Information broke the story late on Wednesday night following the earnings call, with their sources saying that Nvidia has agreed to buy Hugging Face for $12.9 billion.

这笔交易的传闻其实上周末才开始出现。

Now, rumors of the deal only started emerging over the weekend.

当然，虽然交易可能早在外部知情之前就已经在推进了，但看起来谈判过程相当迅速。

And while, of course, this could have all been going on before anyone externally knew, it seems like it might have been a pretty speedy negotiation.

很多报道都聚焦于其高昂的估值倍数。

Much of the reporting focused on the strong valuation multiple.

Hugging Face 的年化营收只有 1.5 亿美元，也就是说 Nvidia 实际上支付了大约 80 倍营收的价格。

Hugging Face only has $150 million in annualized revenue, meaning that Nvidia is technically paying around 80 times revenue as a price.

作为参考，SpaceX 以 600 亿美元收购 Cursor 的估值大约是 15 倍营收。

As a point of reference, SpaceX's $60 billion acquisition of Cursor was around 15 times revenue.

不过，正如我前几天谈到这笔交易潜力时所说，Nvidia 收购 Hugging Face 绝不是为了它的营收。

Still, as I said the other day when I talked about the potential of this deal, Nvidia is very much not buying Hugging Face for their revenue.

相反，这笔交易似乎证实了 Nvidia 正在通过收购来构建一个全栈开源商业模式。

Instead, the deal seems to confirm that Nvidia is acquiring their way into a full-stack open business model.

Hugging Face 是开源模型的首要分发渠道。

Hugging Face is the premier distribution channel for open models.

Nvidia 上周与 Poolside 达成的 60 亿美元授权协议，使他们能够雇佣大约 100 名资深 AI 研究员来充实团队。

Nvidia's $6 billion licensing deal with Poolside announced last week allowed them to hire about 100 veteran AI researchers to staff up their team.

此外，其他投资，比如与 Perplexity 的战略合作，也为他们提供了坚实的工具链以及应用层面的布局。

And other investments like their strategic partnership with Perplexity give them solid harnesses as well as an app layer play.

更重要的是，在大约六个月的时间里，Nvidia 将其开源模型系列 Neutron 从一项技术演示变成了开源模型中的真正竞争者。

What's more, over the course of around 6 months, Nvidia has taken their open model series, Neutron, from a tech demo to a real contender among open models.

Akash Gupta 表示，这可能既是进攻性策略，也是防御性策略。

Akash Gupta suggests that as much as it might be an offensive play, it's also a defensive play.

他写道，Google 已经运行自己的 TPU 长达十年。

He writes, Google has run its own TPUs for a decade.

那些给 Nvidia 开出最大支票的实验室，正是最努力想摆脱它的。

The labs writing Nvidia the largest checks are the ones working hardest to stop.

开源模型就是那个平衡力量。

Open models are the counterweight.

当 Meta、Mistral 或 DeepSeek 发布权重时，数百万开发者下载、微调并部署这些模型，绝大多数都是通过 CUDA 在 Nvidia 的 GPU 上运行的。

When Meta or Mistral or DeepSeek releases weights, millions of developers download them, fine-tune them, and serve them, overwhelmingly on Nvidia GPUs through CUDA.

定制芯片对这个生态系统毫无办法。

Custom silicon has no answer for that ecosystem.

我认为这种防御性定位不仅关乎 Nvidia 的芯片业务，也关乎他们为那个我们一直在讨论的世界所做的布局——在那个世界里，公司会使用多种模型，许多日常核心流程不再依赖最顶尖的前沿模型，而是运行在更开放、更轻量、更灵活、更可定制的东西上。

I think the defensive positioning is not just relative to Nvidia's chip business, but also its positioning for the world that we keep talking about, where companies simply use a diversity of models, and many of their common and core day-to-day processes are run not on the state-of-the-art frontier models, but on something open, lighter, more nimble, more customizable.

不久前，Nvidia 发现了这个机遇，从此便全力朝这个方向冲刺。

A while ago, Nvidia spotted that opening and has been careening towards it ever since.

现在，从大家的反应来看，没有人不理解Nvidia为什么这么做。

Now, in terms of how folks are receiving it, no one has a hard time understanding why Nvidia made the move.

Nvidia显然不只是想卖铲子和镐头，他们想更多地拥有建立在它们之上的AI经济。

Nvidia clearly don't just want to sell the picks and shovels, they want to own more of the AI economy built on top of them.

Nvidia拥有GPU，CUDA拥有开发者，Hugging Face拥有开源AI生态系统，Jensen正在慢慢地把整个技术栈整合到一起。

Nvidia owns the GPUs, CUDA owns the developers, Hugging Face owns the open-source AI ecosystem, Jensen is slowly putting the whole stack together.

Nvidia正在变得远不止是一家芯片公司。

Nvidia is becoming much more than a chip company.

不过也有人担心，在新的所有权下，人们还会不会继续留在Hugging Face。

Others, though, are wondering if people will stick around Hugging Face given its new ownership.

Todd Saunders写道，Hugging Face之所以有价值，部分原因是它被认为是独立且大致中立的。

Todd Saunders writes, Hugging Face is valuable partly because it is perceived as independent and broadly neutral.

他们的用户包括研究人员、初创公司、企业、云服务提供商、硬件公司以及Nvidia的竞争对手。

Their users include researchers, startups, enterprises, cloud providers, hardware companies, and competitors to Nvidia.

如今，开发者可以使用这个平台，而不认为它旨在把他们推向某个特定供应商。

Today, developers can use the platform without believing that it is designed to push them towards one vendor.

这正是他们喜欢它的部分原因。

It's partially what they love about it.

Nvidia的收购可能会破坏这种信任，但这也有可能是一次大获全胜。

An Nvidia acquisition could threaten that trust, but this could be a home run.

Pratham认为这也是一个更大的趋势。

Pratham thinks that this is a bigger trend as well.

他们写道，真正的问题不是Nvidia为什么收购Hugging Face。

They write, The real question isn't why Nvidia bought Hugging Face.

而是当最大的参与者开始购买开源AI所在的平台时，开源AI会变成什么样。

It's what happens to open-source AI when its biggest players start buying the places where it lives.

Eric Hartford认为我们会看到一个机会。

Eric Hartford thinks that we will see an opportunity.

他写道，我为Hugging Face成功退出感到高兴，但这将是开源AI的一大损失。

He writes, I'm happy for Hugging Face for their exit, but this is gonna be a loss for open-source AI.

Hugging Face不再与硬件无关。

HF is no longer hardware agnostic.

这是一个新的标杆崛起的机会。

This is an opportunity for a new standard bearer to arise.

不过，这种担忧是人们面对收购时总会有的。相比之下，我认为这次收购比通常情况要乐观得多，这说明Nvidia在AI社区建立信任方面做得有多好。

Still, that's the kind of concern that people always have around acquisitions, and relative to expecting that to be the case, I would suggest that there is a lot more optimism around this particular acquisition than you normally see, which shows what a good job Nvidia has done building trust in the AI community.

我想不出比Nvidia更适合收购Hugging Face的公司了，这很合理，因为它与Nvidia的整体战略是一致的。

I can't think of a better company to buy Hugging Face other than Nvidia, and it makes sense as it aligns with what they are doing as a whole.

我敢打赌，未来我们会看到更多令人惊叹的功能以及训练和使用模型的方法出现在Hugging Face上。

I bet we will be seeing loads of more amazing features and ways to train and use models on Hugging Face going forwards.

简直是天作之合，也是推动多元化AI的一大步。

Such a great fit and such a great push towards plural AI.

Nvidia卖的是算力。

Nvidia sells compute.

Hugging Face让人们很容易通过多种方式将算力转化为价值，降低了任何单一实验室、公司或国家垄断前沿AI的可能性。

Hugging Face makes it easy to access a plethora of ways to turn that into value, reduces the odds that any one lab, company, or nation will monopolize frontier AI.

多元化AI让世界更美好。

The world is better with plural AI.

Nvidia这周还发布了财报，可以说是大超预期。

Nvidia also reported earnings this week and kinda blew the doors off.

过去一年营收增长加速至106%，比第一季度末的增长率高出21个百分点。

Revenue growth accelerated to 106% over the past year, 21 points higher than their growth rate at the end of Q1.

本季度销售额达到962亿美元，距离单季度1000亿美元的里程碑仅一步之遥，目前只有其他五家公司达到过这个里程碑。

Quarterly sales for the quarter came in at $96.2 billion, tantalizingly close to the $100 billion a quarter milestone that only 5 other companies have achieved.

不过，尽管预期超出了预期，Nvidia仍表示第三季度增速将略微放缓至89.5%。

Still, while forecasts were ahead of expectations, Nvidia is still flagging that growth will slow slightly to 89.5% for Q3.

CFO Colette Kress表示，供应链问题将继续拖累增长，使Nvidia在2027财年的增速限制在70%。

CFO Colette Kress said that supply chain snarls would continue to weigh on growth, limiting Nvidia to 70% for fiscal year 2027.

Nvidia还透露，主要客户要求延长付款时间。

Nvidia also disclosed that major customers are asking for more time to pay.

自由现金流同比下降了50%，只有210亿美元。

Free cash flow fell by 50% compared to a year ago, coming in at $21 billion.

与此同时，应收账款增加了50%，两者形成了鲜明对比。

This fall was mirrored by a 50% rise in accounts receivable.

英伟达表示，原因是“与某些投资级客户签订的大型多季度协议延长了付款期限”。

Nvidia said the cause was, quote, extended payment terms on large multi-quarter agreements with certain investment-grade customers.

这显然对投资者来说是个危险信号，因为应收账款周转天数从45天增加到了60天。

Certainly, this was the big red flag for investors, with days sales outstanding rising from 45 days to 60 days.

除了财务数据，英伟达还宣布正式重返中国市场。

Alongside the numbers, Nvidia announced that they are officially back in the Chinese market.

他们报告称，根据新的出口管制规则，H200芯片的销量很小，但远低于特朗普政府允许的总额。

They reported a small volume of sales of H200 chips under new export control rules, but noted that the figure was well short of the total permitted by the Trump administration.

总的来说，虽然这份财报里确实能找到一些负面因素，但很难忽视那惊人的销售数字。

Ultimately, while there was room to find some negatives in the report, it's difficult to overlook the raw sales numbers.

去年夏天英伟达就已经在出货数量惊人的芯片，而自那以后，他们的销售额几乎翻了一番。

Nvidia was already shipping an incredible number of chips by last summer, and they've almost doubled sales since then.

为了证明这一点，亚马逊宣布，随着他们加速数据中心建设，将额外购买200万块英伟达芯片。

Reinforcing that point, Amazon announced that they would be buying an additional 2 million Nvidia chips as they accelerate their data center buildout.

英伟达没有透露每年总共卖出多少芯片，但据信不到1000万块，这意味着这是一笔巨大的额外订单，将在未来几年内持续支撑强劲的需求。

Nvidia doesn't disclose how many total chips they sell per year, but it's believed to be less than 10 million, meaning that this is a huge additional order that will keep demand solid for years into the future.

这份财报足以让投资者重新青睐英伟达，股价在盘后交易中上涨了4.8%。

The earnings were enough to drive investors back to Nvidia, sending the stock up by 4.8% in after-hours trading.

考虑到英伟达在人工智能生态系统中日益扮演着独特角色，有一个有趣的支线故事：该公司暂停了与NeoClouds的收入分成协议，因为他们正在重新考虑对行业兜底的做法。

One interesting substory, given the unique role that Nvidia is increasingly playing in the AI ecosystem: the company has paused revenue-sharing deal with NeoClouds as they rethink their approach to industry backstops.

不到两个月前，英伟达宣布将开始提供新的融资方案，客户可以以部分利润为交换条件获得信贷支持。

Less than 2 months ago, Nvidia announced that they would start offering new financing deals where customers could receive credit support in exchange for a cut of the profits.

当时，这被视为英伟达之前模式的改进，即他们保证数据中心需求，并希望这能转化为从私人信贷公司或银行获得的更便宜融资。

This was seen at the time as an improvement on Nvidia's previous model, where they would guarantee data center demand and hope that translated into cheaper financing from private credit firms or banks.

消息人士告诉《华尔街日报》，在一些员工表达担忧后，该项目已被搁置。

Sources told The Wall Street Journal that this program has been put on hold after some employees expressed concerns.

有人认为，这种安排可能会引起反垄断方面的关注，而且对于英伟达在客户业务中施加多少控制权也存在敏感问题。

There is a belief that the arrangement could attract scrutiny for antitrust, and there were sensitivities around how much control Nvidia would exert over customer businesses.

《华尔街日报》指出，暂停的具体原因尚不清楚，因此理论上，该计划未来经过一些调整后可能会重新启动。

The Journal noted that the precise reason for the pause couldn't be learned, and so theoretically the program could return in the future with some tweaks.

英伟达方面否认了这些报道，并表示：“我们在7月份推出的新商业模式，旨在为快速发展的AI生态系统提供计算访问权限，目前仍在运行，并因需求旺盛而持续发展。”

Nvidia, for their part, denied the reporting and stated, the new business model we introduced in July that opens up compute access to the fast-growing AI ecosystem is still in place and continues to evolve due to high demand.

现在，稍微聊一下华尔街。信不信由你，Salesforce——我们也会在正片里讨论他们——凭借一份强劲的财报引领了SaaS的复苏。

Now, staying on Wall Street for a minute, believe it or not, Salesforce, who we will talk about in the main episode as well, is leading a SaaS comeback with a strong earnings report.

六个月前，市场的一个主流叙事是软件行业已经完蛋了，公司很快就要用“vibe coding”来自己开发替代SaaS的产品了。

6 Months ago, one major market narrative was that software was done for, and companies would soon be vibe-coding their own SaaS replacements.

这就是当时热议的“SaaS末日”，如果你还记得这个词的话。

This was the much-discussed SaaS-pocalypse, if you remember that term.

显然，那种程度的颠覆并没有发生。我们实际上看到的是，传统软件产品中集成AI工具正在带来收入增长。

That level of disruption obviously did not play out, and increasingly what we're actually seeing is adoption of AI tools within traditional software products is leading to a revenue boost.

周三晚上，Salesforce的财报就印证了这一点。他们报告称，Agentforce产品今年有望带来15亿美元的收入，高于第一季度的预测12亿美元。

That was certainly the story on Wednesday night when Salesforce reported that their Agentforce product is on track to deliver $1.5 billion in revenue this year, up from a forecast of $1.2 billion in the first quarter.

当季总销售额为115亿美元，同比增长11%，所以AI收入在整体中的占比仍然相对较小。

Overall sales were $11.5 billion for the quarter, growing at an 11% pace, so AI revenue is still a relatively modest part of the picture.

不过，这已经足以让市场将Agentforce视为一款成功且正在增长的产品。

This was still enough, though, for the market to identify Agentforce as a successful and growing product.

周四，该股飙升了22%，如今Salesforce距离完全填平“SaaS末日”期间的跌幅仅差5%。

The stock was up a staggering 22% on Thursday, with Salesforce now just 5% away from filling the entire drawdown from the SaaS-pocalypse period.

退一步看，我们一个接一个地看到，一些曾经被严重打压的软件股，凭借稳健的财报重新焕发生机。

Taking a step back, one by one we have seen some of the most beaten-down software stocks deliver solid financials and surge back to life.

ServiceNow、Figma、Atlassian、CrowdStrike、Okta在过去几个月里都实现了大幅反弹。

ServiceNow, Figma, Atlassian, CrowdStrike, Okta have all staged major comebacks over the past months.

KeyBanc的分析师Jackson Adder指出，这些公司甚至不需要发布超预期的财报就能带来巨大的回报。

Key bank analyst Jackson Adder noted that these companies haven't even needed to post blockbuster earnings to generate a huge return.

Adder说：“我们都在逐渐接受一个事实，那就是事情会更有持续性，并不会消失。”

Said Adder, we're all coming around to the idea that things are going to be more durable and are not going away.

所以你会看到Salesforce交出了一份不错但算不上惊艳的成绩单，但是市场反应却格外强烈。

That's why you see Salesforce come out with fine but not stellar results, but you see a really outsized reaction.

Benioff在财报电话会上当然得意了一番，说我真的想从开头讲起，聊聊这个SaaS末日论。

Benioff certainly took a victory lap on the earnings call, saying, I really want to start here at the beginning and talk about the SaaS-pocalypse.

怀疑论者说座位数会下降，但Agentforce、销售、服务以及Slack——所有这些板块的座位数都同比增长了。

Skeptics said that seats would decline, and agent force, sales, service, and Slack— all seats grew year over year.

怀疑论者说客户会流失，但客户流失率却降到了历史最低点附近。

Skeptics said customers would leave, but attrition was near its lowest level ever.

怀疑论者说定价能力会被削弱，但预订量环比翻了一倍多，而且所有业务板块——包括新业务和续约——的合同期限都变得更长了。

Skeptics said pricing power would erode, but bookings more than doubled quarter over quarter, and contract length terms improved across all segments, new business and renewals.

而且通过模型上下文协议调用和CLA调用，平台上的智能体使用量激增了六倍。

And agentic use of the platform surged sixfold via model context protocol calls and CLA calls.

应用程序并没有消亡。

Apps are not dying.

前人工智能主管David Sacks迅速抓住这个新闻，发帖说人工智能资本支出是泡沫以及SaaS已死的说法，今天早上被撕得粉碎。

Former AI czar David Sacks pounced on the news, posting, The AI CapEx is a bubble and the SaaS is dead narratives getting shredded this morning.

Box的Aaron Levy写道，本周的科技公司财报电话会是一个关键的提醒，让我们看到软件和人工智能之间的关系有多么宝贵。

Box's Aaron Levy writes, this week's tech earnings calls were a critical reminder of how valuable the relationship is between software and AI.

软件和智能体会协同工作，并最终随着时间的推移，大幅扩展信息技术总市场规模。

Software and agents are going to work together and ultimately grow the size of the IT TAM dramatically as a result over time.

现在，把这个营收增长的故事带到应用层。根据The Information提供的新数据，Cognition已经实现了年化收入九亿美元。

Now bringing the booming revenue story to the app layer, according to new numbers sourced by The Information, Cognition has hit $900 million in annualized revenue.

这比年初时，也就是编程智能体Devin的创造者宣布的数字，增长了超过三倍。

That's more than a 3x increase since the beginning of the year for the creators of coding agent Devin.

而且看来全行业情况都差不多。

And it appears to be a similar story across the board.

Higgsfield，一个视频和图像模型的界面封装器，今年也翻了三倍，年化收入达到了七亿美元。

Higgsfield, an interface wrapper for video and image models, has also tripled this year to reach $700 million in annualized revenue.

Perplexity的年化运行率翻了三倍，达到了七亿五千万美元。

Perplexity has tripled their run rate to $750 million.

Open Evidence，那个面向医生的聊天机器人，收入翻倍到了三亿美元。

Open Evidence, the chatbot for doctors, has doubled to $300 million.

甚至连Manus，在Meta收购案被搞砸之后，也声称他们的年化运行率达到了四亿美元，比去年翻了四倍。

And even Manus, after the messy unwind of the Meta acquisition, claims they have a run rate of $400 million, quadrupling from last year.

而且，尽管总的数字还远远落后于美国，但领先的中国人工智能公司也看到了类似的增长。

And if the total numbers still remain far behind the US, leading Chinese AI companies are seeing similar types of growth.

The Information还报道说，DeepSeek今年实现了七千万美元的收入，这比2025全年的数字增长了十倍。

The Information again reports that DeepSeek has generated $70 million in revenue this year, which is a 10x increase from their full-year numbers for 2025.

现在，在我们快要讲完的时候，快速看一下政策方面的情况。

Now, as we round the corner, a quick check-in on the policy side of the house.

一项关于人工智能监管的新行政命令，在最近几周草案被悄悄传阅之后，已经停滞不前了。

A new executive order on AI regulation has stalled out after drafts were quietly circulated in recent weeks.

据报道，特朗普政府最近提出了一项行政命令，要成立一个人工智能行业的自律监管机构。

Reports say that the Trump administration recently proposed an executive order that would create a self-regulatory body for the AI industry.

你可能还记得，这个想法是在DeepMind董事长Demis Hassabis在七月份的一篇博客文章中正式提出后，开始获得支持的。

You might remember this idea starting to gather support after DeepMind Chairman Demis Hassabis formally proposed it in a July blog post.

这个新机构将仿照FINRA，也就是金融服务行业的自律监管组织来建立。

The new body would be modeled on FINRA, the self-regulatory organization for the financial services industry.

基本上，这种做法要求人工智能实验室在安全问题上互通信息，对新的模型进行安全测试，并制定行业监管规则。

Essentially, this approach would require AI labs to share notes on security issues, conduct safety testing for new models, and create regulations that govern the industry.

财政部长Scott Bessant，不知为何在人工智能政策中扮演了关键角色，公开支持了这个想法。

Treasury Secretary Scott Bessant, who, for whatever reason, has a key role in AI policy, publicly supported the idea.

据报道，高级官员在过去一个月起草并传阅了一份行政命令草案，打算成立这个新机构。

Reportedly, senior officials drafted and circulated an executive order that would create this new body over the past month.

虽然还不完全清楚这项努力停滞不前的原因，但一些消息来源指出，前白宫人工智能主管David Sacks的反对可能是原因之一。

And while it's not totally clear exactly why the effort has stalled out, some sources flagged opposition from former White House AI czar David Sacks as a potential cause.

在上周末的All In播客节目中，Sacks称针对人工智能的FINRA是一个，引用原话，糟糕的主意，是一个企图封杀开源模型的特洛伊木马。

During last weekend's episode of the All In podcast, Sacks called FINRA for AI a, quote, horrible idea and a Trojan horse for an open source model ban.

他认为，一旦这个机构成立，前沿实验室就会推动制定对所有模型都适用的标准，也就是说，开放模型的开发者会发现这些标准根本无法满足。

He believes that once the body is set up, the frontier labs will push for standards that apply equally to all models, i.e., standards that open model developers will find impossible to comply with.

Sacks还指出，FINRA在促进竞争和创新方面名声并不好。

Sacks also pointed out that FINRA doesn't have a good reputation for promoting competition and innovation.

许多想进入金融行业的初创公司认为FINRA主要关心的是用监管来保护现有企业。

Many startups looking to enter the financial industry view FINRA as largely concerned with using regulation to protect incumbents.

不管怎样，目前还没有针对AI的FINRA，但这并不意味着行业自律的努力没有在进行。

Whatever the case, for now, no FINRA for AI, but that doesn't mean that industry self-regulatory efforts are not happening.

值得注意的是，OpenAI、Anthropic和其他100多家公司呼吁采取紧急的新方法来应对网络防御。

Notably, OpenAI and Anthropic and more than 100 other companies have called for an urgent new approach to cyber defense.

这封公开信由科技、金融和工业领域的公司以及Google、Microsoft和AWS共同签署。

The call came in the form of an open letter, signed by companies across the tech, finance, and industrial sectors alongside Google, Microsoft, and AWS.

名单还包括Visa、通用汽车和普华永道等公司。

The list also includes companies like Visa, General Motors, and PwC.

信中警告说，我们在未来几个月内加强网络防御的时间窗口有限。

The letter warned, we have a limited window to strengthen cyber defenses in the coming months.

随着全球模型能力越来越强，AI驱动的网络攻击将变得更加广泛和复杂。

AI-enabled cyberattacks will become far more widespread and sophisticated as models around the world become increasingly capable.

我们社区依赖的公司和公共服务，从医院到水处理厂再到支撑互联网的基础设施，都面临风险。

The companies and public services our communities depend on, from hospitals to water treatment plants to the infrastructure that powers the internet, are at risk.

当今AI的进步已经为防御者提供了修复多年积累弱点的新方法。

Today's AI advances are already giving defenders new ways to fix weaknesses that have accumulated for years.

如果我们果断行动，就可以利用防御者的窗口让我们的数字世界变得更加安全。

If we act decisively, we can use the defender's window to make our digital world much more secure.

这封信提出了一个集体应对方案。

The letter proposes a collective response.

信中声称，现状行不通了，所有计算机系统都存在长期存在的错误、未修补的漏洞和技术债务。

It asserted that the status quo isn't going to cut it, with longstanding bugs, unpatched vulnerabilities, and technical debt across all computerized systems.

信函呼吁为网络防御者提供更好的AI工具，并加强跨行业的协调。

The letter called for cyber defenders to be provided with access to better AI tools as well as better coordination across industries.

它认为需要全球响应，建立新的伙伴关系以提高安全标准，并为新兴的网络威胁找到新的解决方案。

A global response is necessary, it argues, requiring new partnerships to raise security standards and find new solutions for emerging cyber threats.

Sam Altman在评论这封公开信时，将安全置于竞争之上，他写道：“这是AI网络防御的关键时刻。”

Remarking on the open letter, Sam Altman positioned security as more important than competition, posting, this is a critically important moment for cyber defense with AI.

没有太多时间可以行动了。

There is not much time to act.

我们很高兴你们想与我们或我们的任何竞争对手或合作伙伴合作，但请认真对待这个时刻。

We are happy if you want to work with us or any of our competitors or partners, but please take this moment seriously.

只有紧急而强烈的集体响应才能奏效。

Only an urgent and intense collective response will work.

作为强调，Ethan Mollick教授写道：“在OpenWeight的Mythos级模型和工具可用之前，你的组织在加强网络安全方面的投入还不够。”

Putting an exclamation point on it, Professor Ethan Mollick writes, your organization is not spending enough of its effort bolstering cybersecurity during the window before OpenWeight's Mythos-class models and harnesses become available.

Hugging Face事件表明，即使没有故意的恶意行为者，你也可能暴露在AI黑客攻击之下。

The Hugging Face incident shows us you don't even need intentional bad actors to be exposed to AI hacking.

就像我说的，周末有很多重磅消息，但现在，让我们进入正题，谈谈你现在就可以使用和尝试的一大批新功能和工具。

Like I said, quite a meaty headlines to end the week, but for now, let's move into the main episode and talk about the slew of new features and tools that you could be using and trying right now.

F114:44

毕马威和德克萨斯大学奥斯汀分校的一项新研究发现，当人们与AI合作时，相似的技能并不能保证相似的结果。

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.

研究人员研究了500多名早期职业专业人士，发现表现最好的人通过指导、评估和优化AI的输出，持续放大了AI的价值。

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.

这些被称为AI放大器的最优秀者，定义他们的不仅仅是他们所知道的东西，而是他们如何与AI合作。

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

M117:47

欢迎回到AI每日简报！

Welcome back to the AI Daily Brief!

就像几乎不可能跟上AI新闻的冲击一样，任何一个人也很难跟踪每周发布的所有新功能、模型和工具。

Just like it is nearly impossible to keep up with the onslaught of AI news, so too is it extremely difficult for any one person to keep track of all of the new features and models and tools that come out in any given week.

我们的注意力主要集中在大模型发布上，这些发布可能会更根本地改变我们与AI的互动方式。

Mostly our attention just goes to when there are big model announcements that might more fundamentally shift how we interact with AI.

尽管在2026年，我们对工具更新的熟练程度也提高了不少。

Although over the course of 2026, we've also gotten a lot more adept at harness updates as well.

不过，有时值得梳理一下模型和工具公司发布的所有功能更新、生活质量改进和新能力，看看是否有任何与我们的工作方式相关的东西。

Still, sometimes it's worth going through and just taking an inventory of all of the feature updates and quality of life upgrades and new capabilities that the model and harness companies have released to see if any of them can be relevant for the way that we work.

带着这个想法，今天我们要聊的是，大概过去一周多时间里我看到的十几个更新，这些东西你在近期内可能会觉得特别有用、特别实用。

And with that in mind, today we are going through, I don't know, a dozen updates that I have seen over the last 7 days or so that are things that you might find extremely useful and practical in the immediate term.

首先，Claude 有个更新。

First up, we have an update for Claude.

Claude Cowork 现在内置了一个浏览器，让基于网页的任务更方便了。

Claude Cowork now has a built-in browser to make web-based tasks more convenient.

如果你的任务涉及网站上的操作，比如填写表单、和网页应用互动，那就会有一个浏览器窗口在你的会话旁边打开，供 Claude 使用。

If your task includes something that happens on a website, like filling out a form, interacting with a web app, a browser window will open alongside your session for Claude to use.

这个功能也可以用于智能浏览，类似于之前一些 AI 浏览器吸引大家注意的早期功能，比如 Perplexity 的 Comet 或者 The Browser Company 的 Dia。

The feature can also be used for agentic browsing, similar to some of the early features that caught people's attention from AI browsers like Perplexity's Comet or Dia from The Browser Company.

Claude 指出，不需要额外安装任何东西。

Claude points out that there is nothing new to install.

这个功能直接内置在桌面应用里，和你自己的浏览器以及登录信息是分开的。

That it's built directly into the desktop app and stays separate from your own browsers and logins.

换句话说，这是它自己的浏览器，不会接管你的浏览器。

In other words, this is its own browser, it's not taking over yours.

不过，对于那些想在自家浏览器里操作的人，也可以通过 Chrome 里的 Claude 来实现。

Although for those who want to stay in their own browser, that feature exists via Claude in Chrome.

对很多人来说，这个功能真的是千呼万唤始出来。

For many, this one was a big old finally.

终于做到了，各位！

They did it, folks!

早该有了。

It was about time.

非常好的功能。

Very good feature.

之所以说这个功能意义重大，是因为虽然它看起来只是个小更新，但当你把它和 ChatGPT Work 以及 GrokBot 这些新玩意儿结合起来，就会发现这其实是一个相当重要的转折点。

Gets at why, although this seems like a small feature update, When you combine it with ChatGPT Work and other new things like GrokBot, it all adds up to a fairly meaningful inflection.

他写道，AI 在同一个月里从写文字进化到了替你浏览网页。

AI, he writes, went from writing text to browsing the web on your behalf in the same month.

或者像 Enoms 说的那样，感觉 AI 助手和 AI 操作员之间的界限一下子变得模糊多了。

Or as Enoms put it, feels like the line between AI assistant and AI operator just got a lot thinner.

这感觉可比单纯的 Claude 有了浏览器要大得多。

This feels like a much bigger update than just Claude got a browser.

有趣的地方在于，Claude 现在真的能帮你做网页上的工作了，而不只是告诉你该怎么做。

The interesting part is that Claude can now actually do the web work instead of just telling you how to do it.

需要查资料、浏览网站、填写表单或者完成某个任务？

Need to research something, navigate a website, fill out forms, or complete a task?

浏览器会在 Cowork 里打开，然后 Claude 和你一起处理整个过程。

The browser opens inside Cowork, and Claude handles the process alongside you.

我们正在慢慢告别那种只会给出答案的 AI。

We're slowly moving away from AI that just gives answers.

下一阶段是 AI 直接打开标签页把事情搞定。

The next phase is AI that opens the tab and gets the job done.

不过，也有人指出，Anthropic 这次来得至少有点晚了。

Now, as some pointed out, Anthropic is at least a little late to the party here.

Token Gremlin 说 OpenAI 在 ChatGPT Work 里基本已经实现这个功能好几个星期了。

Token Gremlin says OpenAI has already had basically this in ChatGPT work for weeks.

Julia 的 CEO Rahul 提到了他们七月发布的浏览器代理。

Julia's CEO Rahul points to the browser agent that they released back in July.

基本上就是说，欢迎来迟。

And basically said, welcome to the party.

所以对很多人来说，这个功能的明显灵感来源是 GrokBot。

So for many, the obvious inspiration for this was GrokBot.

GrokBot 发布时，它能在自己的浏览器里运行这个能力是大家觉得特别强大的一个点，坦白讲也是最可能被其他实验室模仿的一个点。

When GrokBot was released, the ability for it to work in its own browser was one of the things that people latched onto as really powerful, and frankly most likely to be copied by all the other labs.

我觉得 ChatGPT Work 更像高级的 RPA。

I find ChatGPT Work more like fancy RPA.

它运行得不错，但更偏向工作流导向，而 GrokBot 则更像一个真正的虚拟同事。

It works well, but it is more workflow-oriented, where GrokBot really is a virtual coworker.

给它一个独立的浏览器，还能通过Slack聊天。

Give it its own browser, chat over Slack.

它很能适应你的风格和偏好。

It does a great job adapting to your style and preferences.

现在，关于ChatGPT Work版本的这项功能，官方ChatGPT账号写道，你可以让它为新公寓设置水电煤气，预约车管所或护照办理，通过保险检查报销费用，根据你的空闲时间找并预约网络内的医生，查看你的汽车注册什么时候到期并准备续期文件，比较你的租房保险政策和你要给房东发邮件反映的问题。

Now, in terms of ChatGPT Work's version of this feature, the official ChatGPT account, they write, you can ask it to set up utilities for a new apartment, book a DMV or passport appointment, check reimbursement costs through your insurance, find and book an in-network doctor around your availability, check when your car registration expires and prepare the renewal paperwork, compare your rental insurance policy with an issue you're emailing your landlord about.

找到并保存符合你条件的房源列表，上传一张照片就能补充库存，等等等等。

Find and save apartment listings that match your criteria, restock something just by uploading a photo, and so on and so forth.

还有大概十多个其他点子。

There are about a dozen other ideas.

鉴于我们一直在讨论AI用于生活和AI用于工作之间的平衡，也就是消费者AI和企业AI，我觉得有趣的是，他们把所有这些更偏向生活场景的任务作为ChatGPT Work的新功能来推。

Given our ongoing conversations about the balance between AI for life and AI for work, i.e., consumer AI versus B2B AI, it's interesting to me that they're pitching all of these much more life-oriented tasks as a new capability for ChatGPT Work.

OpenAI的开发者体验负责人Romain Huet写道，ChatGPT Work现在有了自己的云端电脑。

OpenAI's head of developer experience, Romain Huet, writes, ChatGPT Work now has its own computer in the cloud.

当它能用网络并替你行动时，可能性几乎是无限的。

When it can use the web and take action for you, the possibilities start to feel almost limitless.

另外，ChatGPT还有一个小提升生活质量的地方：他们最近意识到，需要让人们更方便地在临时对话和永久对话之间切换。

Now, one other small quality-of-life upgrade from ChatGPT: the company realized recently that they need a better way for people to switch back and forth between temporary and permanence.

在X上，他们发帖说，我们收到了你的反馈，有时候临时对话变成了值得保存的内容。

On X, they posted, we heard your feedback that sometimes a temporary chat becomes something worth saving.

现在你可以选择用你已有的记忆、自定义指令和插件来个性化一个临时对话，如果想保留的话就把它保存到历史记录里。

Now you can choose to personalize a temporary chat with your existing memories, custom instructions, and plugins, and save it to your history if you want to keep it.

临时对话不会创建新的记忆，也不出现在侧边栏里，除非你选择保存它们。

Temporary chats don't create new memories and stay out of your sidebar unless you choose to save them.

换句话说，对于那些对隐私更关注、可能希望默认使用临时对话功能而不是默认保存所有操作的人来说，这个功能现在让临时对话也能轻松带入相关的ChatGPT上下文，以及在合适的时候升级为非临时对话。

In other words, for people who have a little bit more concern around privacy and who would want to be perhaps default using the temporary chat feature as opposed to saving everything that they do by default, this feature now makes it easier to pull in relevant ChatGPT context even into temporary chats, as well as to upgrade them to non-temporary when it makes sense to.

另一个很大的生活质量提升——坦率地说，这个功能在新闻价值上会远超其体量——我是在开始录音前注意到的。

Another big quality of life upgrade One which will frankly punch way above its weight class in terms of its newsworthiness, I noticed just before I started recording.

OpenAI的Jason写道，插件现在支持多个Gmail账户了。

Jason from OpenAI writes, multiple Gmail accounts available in plugins now.

也就是说，你不再只能关联一个Gmail，而是可以关联多个。

In other words, instead of connecting only one Gmail, you can connect multiple.

这对于很多重度用户来说一直是他们的心愿单上的内容，因为像他们这样的人，日常工作中通常有——我也不知道——五六个不同邮箱需要认真处理。

This has long been on many power users' wishlists, given that the normal course of business for people like this, folks, is to have, I don't know, a half dozen different emails that you interact with in some meaningful way.

IAI的Claire Vo最近就一直在夸，这实际上是她最喜欢的新Grokbot的功能，就是不限制只连接一个Gmail或一个Slack。

How IAI's Claire Vo was recently raving about how this was in fact her favorite feature of the new Grokbot, that it didn't constrain connectors to a single Gmail or a single Slack.

而关于这一点，我不在乎是谁先做的，只要他们都能做到就行。

And boy howdy, is this one where I don't care who did it first as long as they all do it.

说到这个，很多人会很乐意使用大公司给我们的任何工具，不管是OpenAI、Anthropic、SpaceX AI还是谷歌，或者其他公司。

Speaking of that, many of you will always be quite happy to use whatever tools the big companies give us access to, whether it's OpenAI or Anthropic or SpaceX AI or Google or whomever.

但对于那些想要更多控制权、可能已经喜欢上像OpenCLAW这样灵活性的用户来说，Hermes agent值得关注。

But for those who want more control, and who maybe got hooked on the flexibility that something like OpenCLAW enabled, it's worth keeping an eye on the Hermes agent.

Hermes的团队显然在密切关注所有框架和模型公司的动向，因为只要那些公司宣布什么有用的东西，几乎马上就会出现在Hermes上。

The team over at Hermes is clearly watching everything that the harness and model companies are doing, because as soon as anything useful gets announced from those companies, inevitably it finds its way over to Hermes extremely quickly.

比如，这周四他们宣布，Hermes agent现在可以无缝地替你浏览网页了。

So for example, on Thursday this week, they announced that the Hermes agent can now seamlessly browse as you.

他们写道，打开真实配置文件浏览功能，你的agent就会用你现有的Chrome配置文件的托管副本，配合你的登录信息来操作。

They write, turn on real profile browsing and your agent acts with your logins from a managed copy of your existing Chrome profile.

这是在大约十天前，也就是GrokBot发布之后不久，公司为Hermes Desktop推出了Bot Mode之后的事。

This follows from about 10 days ago, right after the GrokBot announcement, when the company introduced Bot Mode for Hermes Desktop.

在Bot Mode下，他们写道，你的agent配置文件会变成一系列命名的bot。

With Bot Mode, they write, your agent profiles become a series of named bots.

每个bot都有自己的角色、模型、记忆、技能和头像。

Each bot has its own role, model, memory, skills, and profile picture.

Bot可以使用任何模型，甚至可以互相通信。

Bots can use any model and even communicate with each other.

创建一次专家bot，就能永久使用。

Build a specialist bot once to use it forever.

为了说明为什么有些人可能对此感兴趣，制造Hermes的Nous Research账号写道，Hermes Agent也有自己的远程电脑，只是你不一定非要用它，而且每月只要几美元，而不是200美元。

Articulating why this might be of interest to some folks, the Nous Research account who makes Hermes writes, Hermes Agent has its own remote computer too, you just aren't locked into using it, and it costs a few dollars per month instead of $200.

其他一些比较大的区别：可以连接任何提供商、订阅或模型，完全免费在你自己的机器上使用，或者通过 Nous Portal 上的几个免费模型之一来使用，可以在机器人模式和经典会话模式之间切换，还有能够自我改进的智能体，它们可以学会更便宜、更有效地执行任务，而且是开源的，所以你可以根据自己的需求修改智能体的任何部分。

Some other big differences: connect any provider, sub, or model, use it completely free on your own machine or with one of several free models on Nous Portal, switch between bot mode and classic sessions mode, self-improving agents that learn to execute tasks cheaper and more effectively, open source so you can modify anything you want about the agent itself to suit your needs.

再说一遍，这对每个人来说都不是必需的，但我认为从消费者选择的角度来看，这相当酷。

Again, this is not going to be necessary for everyone, but I think it is pretty rad from a consumer choice perspective.

随着这些新的功能升级和能力上线，Hermes 基本上跟上了所有主要的框架。

That Hermes is very much keeping pace with all of the major harnesses as these new feature upgrades and capabilities come online.

顺便说一下，Hermes 不仅仅是在复制别人的东西，他们自己也在做一些非常酷的实验。

And by the way, Hermes isn't just copying things, they're also doing some really cool experimentation on their own as well.

他们有一个 Hermes HUD 模式，基本上是一个无缝叠加层，你可以一边做其他事情（比如玩游戏），一边跟踪工作生活中的事情。

They have a Hermes HUD mode, which is basically a seamless overlay where you can keep track of things going on in your work life as you're doing other stuff like playing a game.

官方新闻账号写道，我现在整天都在玩游戏，工作就在那个小文本框里完成了。

The official news account writes, I just play games all day and get my work done in that little text box now.

如果说 Hermes 是一端，那 Salesforce 和 Claude 的新合作就一定在另一端了。

Now, if Hermes is on one end of the spectrum, the new collaboration between Salesforce and Claude has to be on the other.

Salesforce 和 Anthropic 正在合作，他们称之为——听好了——Claude Force。

Salesforce and Anthropic are teaming up for a partnership that they are calling— wait for it— Claude Force.

它的核心是一套新的插件和技能，允许 Salesforce 用户通过 Claude 智能体来访问他们的客户关系管理系统（CRM）。

At its core, this is a new set of plugins and skills to allow Salesforce users to access their CRM using Claude agents.

他们可以使用 Salesforce 的数据作为上下文，而且依托 Salesforce 的 AI 框架（他们称之为 AI Force），可以在尊重现有治理规则的前提下执行智能体操作。

They can use Salesforce data as context, and thanks to Salesforce's AI harness, which they call AI Force, Agentic actions can be taken while respecting existing governance guardrails.

Salesforce 在他们的新闻稿中写道，几十年来，企业软件要求用户手动浏览静态界面来完成工作。

In their press release, Salesforce writes, for decades, enterprise software required users to manually navigate static UI to get work done.

现在，智能体可以直接访问数据、工作流和规则，软件成为一个驱动所有界面的系统。

Now that agents can access the data, workflows, and rules directly, software is a system that powers every interface.

这一举措突显出我们与像 CRM 这样的东西交互的方式正在发生不可避免的转变，使得智能体能够更自然地完成交互。

The move underscores that there is an inevitable shift in how we interact with things like CRM, making it easier for agents to more natively do so.

Salesforce 的 CEO Marc Benioff 说，这里的界面就是 AI，可以让你动态地构建自定义应用，并回答任何企业问题。

Salesforce CEO Marc Benioff said, here the UI is the AI, allowing you to build custom apps dynamically and answer any enterprise question.

仅靠概率智能无法运营一家公司，而确定性系统则无法推理。

Probabilistic intelligence alone doesn't run a company, and deterministic systems don't reason.

顺便说一句，这差不多是你能想到的最花哨的说法，来表达 AI 和软件各有各的角色。

Which, by the way, is just about the fanciest way to say that AI and software both have roles to play that you could have possibly imagined.

Benioff 继续说道，通过将 Claude 非凡的推理能力与每个企业所依赖的可信数据、工作流和治理相结合，我们正在交付一个能够思考、推理和行动的动态界面。

Benioff continues, by fusing Claude's extraordinary reasoning with the trusted data, workflows, and governance every enterprise runs on, We're delivering a dynamic interface that thinks, reasons, and acts.

这就是未来每个企业运行的方式。

This is how every business will run.

是的，甚至还为此拍了一个新的 Matthew McConaughey 广告。

And yes, it even got a new Matthew McConaughey ad.

对一些人来说，尤其是当 Dario 和 Benioff 一起出现在 CNBC 上时，这是一个可以开玩笑的时刻。

Now for some, especially as Dario appeared next to Benioff on CNBC, this was a moment for cracking jokes.

企业家 Varun Ram Ganesh 写道，Anthropic 在融资的时候说：我们的总目标市场是美国经济。

Entrepreneur Varun Ram Ganesh writes, Anthropic while raising VC money: Our TAM is the US economy.

我们是终极智能。

We are the ultimate intelligence.

我们要消灭整个应用层。

We are killing the entire app layer.

我们要创造50%的失业率。

We are going to create 50% unemployment.

并且让整个人类劳动力变得多余。

And make the entire human workforce redundant.

现实中，Anthropic 兴奋地宣布与 Salesforce 合作推出 Claude Force。

Anthropic, in reality, excited to announce Claude Force in partnership with Salesforce.

但这种玩笑极大地低估了像 Salesforce 这样的东西在多大程度上支撑着企业世界。

But this sort of joking massively undersells just how much of the enterprise world actually runs on things like Salesforce.

Claude Force 足够引人注目，让我重新思考我之前的看法。

Claude Force is notable enough to make me rethink my priors.

竞争信号正在发生变化。

The competitive signal is shifting.

除了标题之外，我觉得最值得注意的是，Anthropic 的首席商务官说，CRM 集成只是第一步，后面还会有更多。

Beyond the headline, what I find most notable is that the Anthropic CCO saying that CRM integration is the first of many.

实际上，软件供应商一直应该是前沿实验室最大的客户。

Realistically, software vendors always should have been the Frontier Labs' largest customers.

但到目前为止，似乎 Anthropic 和 OpenAI 都决定直接面向客户，而不是通过他们。

Up to now, however, it seemed like Anthropic and OpenAI had decided to go direct instead.

他们的策略正在改变，在一个动态的市场里，当事实发生变化时，你的看法也必须改变。

Their strategy is changing, and in a dynamic market, when the facts change, your view has to change.

Anthropic 显然意识到，在企业内部，当代理的护栏、治理和基础设施还不存在，而且未来一到两年内也不会存在时，直接取代软件是很困难的。

Anthropic is clearly recognizing that it's hard to displace software outright when the agentic guardrails, governance, and infra doesn't exist in the enterprise and won't for another 1 to 2 years.

现在的问题是，客户是否仍然想把数据从记录系统中移出来，自己构建原生方案，从而避免支付软件平台的 API 费用？

The question now is, do customers still want to move data out of systems of record to build natively and avoid paying software platform API fees?

还是他们会接受在现有软件基础设施上启用 AI 所带来的便利？

Or will they accept the utility of enabling AI on an existing software infrastructure?

这种叙事正在转变，直觉上，这些原生集成应该能以更少的投入，给客户带来他们想要的 AI 价值。

The narrative is shifting, and intuitively, these native integrations should deliver customers the value they want from AI without as much of a lift.

鉴于这种叙事发展的规模，现在很难对企业软件持悲观态度。

Hard to be bearish on enterprise software at present given the magnitude of this narrative development.

数据、运营嵌入的工作流、专业知识、在位优势仍然是软件的护城河，而 Anthropic 正在发出一个市场信号，承认这些护城河的持久性。

Data, operationally embedded workflows, expertise, incumbency are still the moats for software, and Anthropic is sending a market signal acknowledging the persistence thereof.

而且你看，这显然只是一些零星的例子。

And look, obviously it's just a handful of anecdotal examples.

但我和一些了解这个 Claude Force 交易的人聊过，他们平时是和 Salesforce 这类系统打交道的，这些人对此非常兴奋。

But the folks that I spoke to about this Claude Force deal that actually interact with systems like Salesforce were extremely excited about it.

而且至少有一个原本打算自己构建一些 Salesforce 替代方案的人，或者说至少是在探索这条路的人，基本上叫停了那个计划，想先看看这个方案能不能以合理的成本直接满足他们的需求。

And at least one who had been planning on potentially building some alternatives to Salesforce, or at least exploring down that path, basically threw the brakes on that so they could figure out if this just did what they wanted in a way that it justified the cost.

所以，虽然 X 上的那些发帖人可能会觉得好笑，但在真正的企业世界里，我想你们有些人坐在这里会觉得，这是我到目前为止说的所有事情里，最重要的一笔交易。

So while the X posters may get their yucks, over in the real enterprise world, some of you I imagine are sitting here thinking that this is the biggest deal of anything I've said so far.

现在我们已经介绍了大量功能，但还有几个模型值得一提，这周它们主要集中在音频和视频领域。

Now already we've gone through a ton of features, but there are a couple models that are worth noting, and this week they were clustered around audio and video.

Gemini 这周又上了新闻，他们发布了新版本的语音转文字模型，Gemini 3.5 Transcribe。

Gemini was back in the news this week with the release of a new version of their speech-to-text model, Gemini 3.5 Transcribe.

这个模型支持超过 85 种语言，谷歌说他们在可用性方面取得了巨大进步。

The model supports more than 85 languages, and Google says they've made some huge strides in usability.

如果你需要直接转录，模型可以工作在逐字模式；或者你也可以使用智能翻译模式，它会去掉填充词、压缩啰嗦的表述、处理自我修正，输出一个干净的用户意图版本。

The model can function in verbatim mode if you need a direct transcription, or you can use it in smart translate mode, which strips away filler words, condenses rambling thoughts, and handles self-corrections to deliver a clean version of user intent.

你甚至可以创建自定义词汇，来增强模型处理专业术语和俚语的能力。

You can even create custom vocabulary to enhance the model's ability to deal with jargon and slang.

谷歌表示，这个模型在嘈杂环境中表现良好，并且能区分多达三个不同的说话人。

Google says the model performs well in noisy environments and can distinguish between up to 3 different speakers.

这个模型可以在谷歌的 AI 套件中使用，如果你喜欢用语音模式和 AI 交互的话；它也可以作为 API 供第三方应用使用。

The model can be used across Google's AI suite if Voice Mode is your preferred way of interacting with AI, and is also available as an API to power third-party apps.

这实际上和 Whisper Flow 这类产品处于同一个领域，不过它不只是一个终端应用，更是其他人可以在此基础上构建的关键基础设施。

This is then playing very much in the space of something like Whisper Flow, although built not just to be an end application in itself, but also a key piece of infrastructure for other people who can build on top of it.

AI 开发者 Jan Kronberg 写道，这种发布会有很多人错过。

AI developer Jan Kronberg wrote, this is the kind of release a lot of people are going to miss.

在此之前，语音输入要求你基本上像写作一样说话——要有标点、结构良好的句子、过程中少做修正。

Until now, voice input required you to basically speak the way you write— punctuation, well-structured sentences, fewer corrections along the way.

Gemini 3.5 Transcribe 颠覆了这一点。

Gemini 3.5 Transcribe flips that around.

你按照自己实际思考的方式说话，系统会生成你大概想写的文字，一边说一边帮你清理和整理。

You speak the way you actually think, and the system produces the text you probably wanted to write, cleaning it up and organizing it as you speak.

顺便说一句，把这种清理做好真的很难。

Now, as an aside, it is really hard to do that cleanup well.

比如，我经常在 Whisper Flow 的轻度清理版本和无清理版本之间来回切换。

For example, I am constantly switching back and forth on Whisper Flow from the version where it does light cleanup to the version where it does no cleanup.

无清理版本确实看起来像是一段没有良好标点的啰嗦话，有时还会不清楚。

No cleanup really does look like rambles without good punctuation and sometimes gets unclear.

但即使是最轻度的清理设置，Whisper Flow 也会把我用作连接词而非填充词的词去掉，假设它们是填充词。

But even on the lightest cleanup setting, Whisper Flow strips away words that I am using as connectors rather than filler, assuming that they are filler.

不管怎样，就因为这个原因，我很期待看到谷歌在这里能提供什么不同的东西。

In any case, for that reason alone, I'm excited to see if Google offers something different here.

SpaceX AI 也推出了 Grok Voice Think Fast 2.0，并指出它现在在 Artificial Analysis Speech-to-Speech Index 上排名第一。

SpaceX AI also dropped Grok Voice Think Fast 2.0, noting that it's now number 1 on the Artificial Analysis Speech-to-Speech Index.

他们写道，这个指数衡量的是语音智能体能否对听到的语音进行推理、解决真实客户的问题，以及使用智能体工具正确完成任务。

They write, this index measures whether voice agents can reason over the speech it hears, resolve real customer issues, and correctly complete tasks using agent tools.

他们接着说，在 Starlink，我们每天使用 Grok Voice 处理超过一万五千通客户支持和销售来电。

At Starlink, they continue, we're using Grok Voice to resolve over 15,000 inbound customer support and sales calls a day.

Grok 可以诊断硬件问题、寄出替换品，并且每周通过语音电话和聊天完成超过三千个订单。

Grok diagnoses hardware issues, ships replacements, and fulfills over 3,000 orders a week via voice calls and chat.

很多人展示了 demo，探索速度和延迟的升级究竟在多大程度上改变了你在智能体场景中可以用语音做的事情。

A bunch of people showed demos exploring just how much the speed and latency upgrades really change what you can do with voice in the agentic context.

不过，我们也从 Google 那里得到了一个新的视频模型。

We also, however, got a new video model from Google as well.

这个模型叫做 Gemini Omni 1.1 Flash，他们写道，通过这次更新，你可以扩展场景、指定一个镜头的起始和结束帧、添加视频输入参考、把你喜欢的片段提升到 4K，并且可以快速在 360p 下测试想法。

The model is called Gemini Omni 1.1 Flash, and they write that with this update, you can extend your scenes, specify starting and ending frames of a shot, add video input references, upscale your favorite takes to 4K, and test ideas quickly in 360p.

换句话说，这百分百是一次不仅仅关乎所谓的“更好的视频”的更新。

In other words, this is 100% an update that is not just about quote unquote better video.

它关乎的是视频生成整体可用性的大幅提升。

It's about massively more usability of video generation in general.

现在，我们又得到了一些新的基准测试结果。

Now, once again, we got some new benchmarks.

Arena 表示，Omni 1.1 Flash 在文本转视频的竞技场中排名第一，在图像转视频的竞技场中排名第二，不过说到新的视频模型，好与不好还得看实际表现。

Arena says that Omni 1.1 Flash hit number 1 in the text-to-video arena and number 2 in the image-to-video arena, although when it comes to a new video model, the proof is in the pudding.

看起来人们的第一印象是，这确实是比之前的 Omni 模型有所升级。

It seems like people's first impression is that this is indeed an upgrade over the previous Omni model.

Chablan 写道，最大的升级是语音，听起来明显更好、更自然，而且它生成的视频也比 Omni Flash 更逼真。

Chablan writes, the biggest upgrade is the voice and it sounds noticeably better, more natural, and it also generates more realistic videos than Omni Flash.

不过有些人发现，当把它和 Seed Dance 2.5 进行比较时，它并不算是明显的改进，甚至可能根本算不上改进。

Some found, however, that when they compared it to Seed Dance 2.5, it wasn't as clearly an improvement if it was an improvement at all.

在视频方面，Fal 也推出了 H3 Max，他们宣称它有更好的提示理解能力，而且速度要快得多得多。

On the video front, Fal also introduced H3 Max, which they pitched as having better prompt understanding and being just way, way faster.

他们声称平均延迟为 3.49 秒，而 Seed Dance 2.5 是 68 秒，Gemini Omni Flash 是 26.7 秒。

They claimed an average latency of 3.49 seconds compared to Seed Dance 2.5's 68 seconds and Gemini Omni Flash's 26.7 seconds.

来自 Andreessen Horowitz 的 Justine Moore 写道，这是 Fal 推出的一个疯狂模型。

Justine Moore from Andreessen Horowitz writes, this is an insane model from Fal.

我无法形容只需 15 秒就能生成一个长片段的感觉有多疯狂。

I can't describe how wild it feels to generate a long clip in just 15 seconds.

这对整个领域来说将是变革性的。

This will be transformative for the space.

细节方面我不太确定，但这个生成得很快，1.3 秒，只要 0.20 美元。

Very fuzzy on the detail here, but this generated fast in 1.3 seconds for $0.20.

F132:51

这也太离谱了。

That's ridiculous.

M132:52

H3 Max 可是件大事。

H3 Max is a big deal.

H3 Max 的第一印象。

H3 Max first impressions.

15 秒的生成时间。

15-Second generation time.

音频很棒，声音听起来很好，音效也很好。

Audio is amazing, voices sound good and sound effects too.

对提示词的遵循度令人难以置信。

Incredibly prompt adherent.

你说什么它就做什么。

It does exactly what you tell it to.

还需要更多测试，但很难有别的模型能比得上 H3 Max 的平衡性和价格。

Needs more testing, but it's hard to beat the balance and price of H3 Max.

AI 视频跨越了一条界限。

A line in AI video was crossed.

在我仅使用网页界面的实验中，H3 Max 现在生成一段质量尚可的 AI 视频所需的时间，比你观看这段视频还要短。

In my experiments with just the web interface, H3 Max can now create reasonably high-quality AI video in less time than it takes you to watch it.

是的，这些影响非常有趣。

Yes, the ramifications are super interesting.

它开创了实时交互的视觉环境，作为一个完全生成的世界。

It opens up live interactive visual environments as a fully generated world.

人们可以拥有大规模多人在线、完全响应且独一无二的用户体验，没有明显的切换或等待。

One can have massively multiplayer and completely responsive and unique user experiences with no visible cutting or waiting.

M133:50

那么，有了这些，从小的生活品质提升到全新的视频生成模型，有很多新东西可以去探索和尝试。

So with all of that, from little quality of life upgrades to entirely new video generation models, There is a lot of new stuff to go out there and play with.

希望本期节目给你带来了一些想法，让你想马上投入去尝试。

Hopefully this episode gave you a few ideas for something you want to jump in and try.

那么今天的人工智能每日简报就到这里。

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

一如既往地感谢你的收听或收看，下次再见，平安。

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

已剔除 9 处广告（点击展开查看）

M10:46广告 · 已剔除

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

To get an ad-free version of the show, go to patreon.com/AIDailyBrief, or you can subscribe on Apple Podcasts. To learn more about sponsoring the show, send us a note at [email protected].

And lastly, if you listen to today's episode and think to yourself, man, I want to use all those features, but I feel behind, I need to catch up in my AI skills, the next cohort of super intelligent training programs, which are incubated over here at AIDB, are coming up in September.

You can find a link to them on the AI Daily Brief website, or you can just go to training.besuper.ai.

F115:10广告 · 已剔除

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

M115:19广告 · 已剔除

If you listen to this show, you likely have a thesis.

F115:22广告 · 已剔除

Maybe it's enterprise adoption, maybe it's compute, maybe it's a specific lab.

Harbor Capital's AI Lab Ecosystem ETFs let you express it via 5 actively managed ETFs, each seeking exposure to the ecosystem around one major lab.

Anthropic, OpenAI, DeepMind, Meta, or xAI.

Your view of the AI race in ETF form.

Harbor Capital Advisors' AI Lab Ecosystem ETF suite gives investors a way to invest in the AI ecosystem they believe is best positioned for success.

Search Harbor AI Lab Ecosystem ETFs wherever you invest or follow @HarborCapital on X to learn more.

Visit harborcapital.com for a prospectus containing investment objectives, risks, fees, expenses, and other important information.

Read and consider it carefully before investing.

Risks include principal loss and artificial intelligence-related risks.

Harbor ETFs are distributed by Foresight Fund Services, LLC.

M116:05广告 · 已剔除

Harbor is not affiliated with AI Daily Brief, and the funds are not affiliated with, sponsored by, or endorsed by any AI lab.

F116:10广告 · 已剔除

This is a paid advertisement and not personalized investment advice.

Investing involves risk, including possible loss of principal.

M116:17广告 · 已剔除

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

F116:22广告 · 已剔除

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 B-L-I-T-Z-Y dot 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.

New users get $1,000 in inference.

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's agent turns competitor moves into landing pages.

Sales' agent enriches leads, drafts emails, and updates the CRM.

Ops' agent chases the paperwork and tracks the budget.

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

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

Hire yours at HyperAgent, built by the team at Airtable.

Claim your $1,000 in inference at hyperagent.com/AIDailyBrief.

M133:36广告 · 已剔除

Now, for those of you who want to try H3 Max out, you can do it over on Venice.ai.

Venice is a privacy-focused AI alternative that recently crossed the threshold of $100 million in ARR, suggesting that more privacy-conscious AI has a market.