AI洗白不会持续太久

今天的《AI每日简报》要聊的是，中国新发布的开源权重模型如何与企业AI思维的重大转变相关联。在此之前，先来看头条新闻：Palantir与迈向AI主权的进程。

Today on the AI Daily Brief, what a new Chinese open weight model release has to do with big shifts in enterprise AI thinking, and before that in the headlines, Palantir and the march to AI sovereignty.

《AI每日简报》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.

好了，朋友们。

All right, friends.

在正式开始之前，先快速说几个通知。

Quick announcements before we dive in.

首先，感谢今天的赞助商：KPMG、Airtable、Section和Blitzy。

First of all, thank you to today's sponsors, KPMG, Airtable, Section, and Blitzy.

想收听无广告版本的节目，请访问 patreon.com/aideallybrief，或者在Apple Podcasts上订阅。

To get an ad free version of the show, go to patreon.com/aideallybrief, or you can subscribe on Apple Podcasts.

如果想了解如何赞助本节目，请发送邮件至 sponsors@aideallybrief.ai。

And to learn more about sponsoring the show, send us a note at sponsors@aideallybrief.ai.

本周，科技公司财报继续发布，Palantir再次交出了一份惊人的季度成绩。

This week, tech earnings continue with Palantir registering another monster quarter.

季度营收达到19.4亿美元，比去年同期增长93%，超出预期。

Quarterly revenue came in at $1,940,000,000 up 93% from this time last year and beating expectations.

商业销售额同比增长149%，高于第一季度133%的增长率，这表明尽管有关于削减代币预算的夸张新闻，企业AI需求依然火爆。

Commercial sales were up 149% year over year, up from 133% growth rate in Q1, demonstrating that enterprise AI demand is still booming despite flashy headlines of token budget cuts.

确实，事实证明，Palantir还成功扩大了利润率，本季度净利润达到10亿美元，年增长率达225%。

Indeed, it turns out that caps Palantir also managed to expand profit margins, with net income reaching $1,000,000,000 for the quarter and growing at a 225% annual pace.

首席执行官亚历克斯·卡普称这个季度“超凡脱俗”，并利用这次财报机会，在他那高高在上的讲台上，越来越大声地宣扬他的观点。

CEO Alex Karp described the quarter as otherworldly, and used the earnings as a chance to proclaim the message that he has been getting increasingly loud about on his bully pulpit.

他基本上把Palantir的业绩描绘成对AI主权需求的体现。

He basically painted Palantir's results as an expression of the demand for AI sovereignty.

他说：Palantir是唯一一家证明能将代币转化为实际经济价值的公司。

He said: Palantir is the only company that has demonstrated it can transform tokens into actual economic value.

我们的客户信任我们，因为我们能让他们对运营、数据和决策拥有最大程度的控制权。

Our customers trust us to provide them with maximal control over their operations, data, and decisions.

Palantir上调了全年预测，导致股价在盘后交易中飙升10%，同时强调对运营、数据和决策拥有最大程度的控制权。

Palantir hiked annual forecasts, sending the stock surging by 10% in after hours trading, with maximal control over their operations, data, and decisions.

他在股东信里也传达了同样的信息。

This was the message that he echoed in his shareholder letter as well.

卡普写道：全球每个组织都开始意识到，把机构大门的钥匙交给语言模型的创造者，让这些模型在自己的地盘上随意运行，是件多么危险的事。

Karp wrote: Every organization in the world is awakening to the risks of handing the creators of language models the keys to their institutions, of letting these models loose within their homes.

我们合作伙伴的需求很明确：他们需要掌控数据、掌控模型接收的提示词，更根本的是，要掌控组织智能和商业情报——也就是他们的核心优势——而这些正是语言实验室不仅愿意、而且从结构上就设计好要从客户那里攫取的东西。

The demand from our partners is clear: it is for control over data, the prompts that the models ingest, and more fundamentally, the organizational and business intelligence, their alpha, that the language labs are not only ready and willing, but structurally designed to capture from their customers.

在信的后面，他继续说道：我们不是靠点击量、代币或聊天次数来赚钱的。

Later in the letter, he continued: We do not get paid for clicks or tokens or chats.

把现代经济史上最重要的进展搞成游戏化，在我们看来是搞错了方向。

The gamification of the most significant development in modern economic history seems to us misplaced.

平台的使用量或许能暗示其价值，但绝不是决定性因素。

The usage of a platform may hint at its value but is by no means dispositive.

现在很多人发现，光靠消费和使用，跟产生实际成果几乎没什么关系。

And many are now finding out that consumption and usage alone have little or nothing to do with the production of results.

为了给这一切再添把火，他接着说道：我们的业务带有马克思主义的色彩和潜台词。

Just to add a little fire to all of it, he then says: There are Marxist overtones and undertones to our business.

其他一些人，包括许多正在构建大型语言模型的人，有意或无意地，想要夺取他们所谓合作伙伴的生产资料。

Others, including many of those building large language models, intend, knowingly or otherwise, to capture the means of production of their purported partners.

代币工业综合体的局限性和缺陷，它曾威胁要接管并主导世界经济，如今已越来越多地被暴露出来。

The limitations and faults of the token industrial complex, which has threatened to overtake and dominate the world economy, have increasingly been exposed.

我们一直拒绝，并将继续拒绝与我们的合作伙伴建立寄生关系。

We have always declined and will continue to decline entering into a parasitic relationship with our partners.

在接受CNBC的后续采访时，卡普明确指出了他针对的是谁。

In a follow-up interview with CNBC, Karp made it clear exactly who he takes issue with.

他说：有人试图用毒品般的方式让我们上瘾，走向一个他们自认为掌控的未来。

He said: We have people trying to drug addict us to a future they believe they control.

我和达里奥以及有效利他主义团队相处了很长时间。

Now, I've spent a lot of time with Dario and the Effective Altruism crew.

他们想告诉你，我们必须走向一个未来，在那里我们一无所有，你的企业不盈利，我们都没有工作，而我们的对手获胜。

They want to tell you we have to march into a future where we own nothing, where your businesses aren't profitable, where none of us have jobs, and where our adversaries win.

当然，这类评论并非每个人都会当真，它们引发了各种各样的看法。

Now, of course, these are not the type of comments that everyone is going to take at face value, and they generated a wide range of opinions.

总结一下最乐观的看法，Amit投资公司写道：人工智能时代的关键不在于估值，而在于赋能员工、增强自主性、推动GDP增长。

Summing up the most optimistic version of the take, Amit is investing writes: The age of AI is not about valuations, but about empowering workers, enabling agency, and growing GDP.

说到那些跟前沿实验室叫板的公司，苹果现在当然正跟OpenAI打官司，声称OpenAI通过一名从苹果跳槽到OpenAI的员工，窃取了苹果的商业机密。

Now speaking of companies that are taking issue with the frontier labs, Apple right now is, of course, embroiled in a lawsuit with OpenAI, claiming that via an employee who left Apple to join OpenAI, the LLM lab stole Apple trade secrets.

OpenAI正在反击。

OpenAI is biting back.

在一封新信中，他们写道：苹果搞错了。

In a new letter, they wrote Apple is getting this wrong.

他们写道，苹果是有史以来最伟大的公司之一，以对细节的极致追求而闻名。

Apple, they write, is one of the greatest companies of all time, and built a reputation for obsessing over the smallest details.

但这起粗心、激进又奇怪的诉讼，很遗憾，配不上这个名声。

This careless, aggressive, and oddly personal lawsuit sadly doesn't live up to that reputation.

最让人震惊的一句话是这句。

Now, the big jaw drop line was this one.

苹果声称他们在二月份联系过OpenAI，但我们没有回应。

Apple had claimed that they contacted OpenAI in February and that we didn't respond.

他们现在承认，他们的外部律师发邮件时搞错了人，把两个亚洲姓氏搞混了，直到我们指出这个问题他们才发现。

They now admit that their outside lawyers emailed the wrong person after confusing two Asian last names only after we brought this to their attention.

说实话，如果这是真的，那这起诉讼让很多旁观者都产生了不少疑问。

Which, honestly, if true, brought up some big questions with the lawsuit for many folks watching this from the outside.

现在这个阶段，这出心理剧比我们节目通常讨论的内容要戏剧化一些。

Now, at this stage, this is a little bit more psychodrama than we normally get into on this show.

如果这真的能影响OpenAI在硬件方面的计划，那还是值得关注的。

It'll be worth paying attention to if it actually shifts the plans that OpenAI can make with hardware.

目前来看，这个错误实在太离谱了，我不得不在这里分享给大家，让大家感受一下AI社区的氛围。

For now, that is just such a wild mistake to make that I had to capture the zeitgeist of the AI x community by sharing it with you here.

接下来，我们看看其他一些引起关注的评论。

Moving on to some other comments that are getting attention.

DeepMind的首席战略官将高昂的资本支出重新定义为对递归自我改进的预付款。

DeepMind's Chief Strategy Officer has reframed sky high CapEx as a down payment on recursive self improvement.

在加州大学伯克利分校的一个小组讨论中，Jasjit Sikhan表示，递归自我改进是投资理念的关键部分。

During a panel at UC Berkeley, Jasjit Sikhan said that RSI was a key part of the investment thesis.

你仔细听的话就会发现，这个财报季的核心话题就是AI的投资回报率如何与不断攀升的资本支出相匹配。

Now, you've been listening closely, this earnings season has been all about how the ROI of AI lines up with endlessly ramping CapEx.

特别是谷歌，因为其AI收入未能跟上短期支出，自由现金流转为负值，而受到了市场的惩罚。

Google, in particular, was punished for their AI income failing to live up with spending on the short term as free cash flow flipped negative.

这位谷歌高管承认，目前的AI收入“并不能支撑我们到目前为止的资本支出”，从而造成了“我们可能陷入AI空窗期的风险，即支出发生了，但收入却没有出现”。

This particular Google executive acknowledged that current AI revenues, quote, don't sustain the capital expenditures we're making so far, creating a, quote, danger we could hit an AI air pocket such that the expenditures happen but the revenues don't show up.

不过，他认为资本支出的扩张不是为了短期收入，而是人类文明有史以来最大的一次科学赌注。

However, he argues that the capex build out is not about near term revenue, but instead the biggest scientific bet civilization has ever made.

接下来是一个有趣的网络安全好消息：Claude帮助研究人员发现了一个在刑事起诉中使用的DNA证据数据库中存在数十年的漏洞。

Next up, an interesting good news story in cybersecurity as Claude helps researchers uncover a decades old vulnerability in DNA evidence databases used in criminal prosecution.

在全国各地的实验室里，法医科学家将DNA样本保存为数字文件，以创建可搜索的数据库。

At labs across the country, forensic scientists keep DNA samples as digital files to create a searchable database.

一组研究人员发现，他们可以使用Claude编写的代码来修改这些文件，整个过程大约需要45分钟。

A group of researchers have found that they could alter the files using code written by Claude in a process that takes around forty five minutes.

核心问题在于，这个数据库软件创建于1995年，没有任何现代软件那样的防篡改保护措施。

The core issue is that this database software was created in 1995 and includes none of the tamper evident protections of modern software.

其中一位研究员，法医学家、纽黑文大学教授劳拉·盖达什·科姆说：实际上，我们手头这些被合法称为法医学黄金标准的数据文件，却连一个纸袋所要求的防篡改标记都没有。

One of the researchers, forensic scientist, and New Haven University professor Laura Gaydash Kom said: Effectively, what we have are data files that are legitimately referred to as the gold standard of forensic science, that lack the same level of tamper evident markings that we require for a paper bag.

现在，后果令人担忧：一个对DNA检测有基本了解、并能访问数据库的恶意行为者，就可能篡改文件，甚至修改证据来陷害无辜的人。

Now, the ramifications are concerning: a bad actor with rudimentary knowledge of how DNA testing works and access to the database could corrupt files or even modify evidence to frame an innocent person.

全国各地的实验室都没有报告过这类证据篡改事件，但他们也一直没能找到检测是否发生过篡改的方法。

None of the labs around the nation have reported any kind of this evidence tampering, but they also haven't been able to figure out a way to detect if tampering has occurred.

作为披露的一部分，研究人员指出，目前仍在使用的某些加密技术，依赖的是一个已经在互联网上流传多年的加密密钥。

As part of their disclosure, researchers noted some of the encryption still being used relies on an encryption key that's been available on the Internet for years.

该数据库软件的创建者表示，他们一直与美国网络安全和基础设施安全局密切合作，研究缓解措施，并已推送了一个软件更新，以实施数字签名。

The creator of the database software said that they have been working closely with the US Cybersecurity and Infrastructure Agency on mitigations, and have pushed a software update to implement digital signatures.

珀尔马特司法正义中心的政策与改革主任莎拉·邱表示，这项研究凸显了法医领域有多么落后。

Sarah Chew, the director of policy and reform at the Perlmutter Center for Legal Justice, said that the research highlights how badly behind the forensic sector is.

她说：从其他行业吸取的教训，并没有被认真地引入法医学领域。

She said: Lessons learned from other industries haven't been imported into forensic science in a serious way.

我们落后太久了。

We've been behind the ball for so long.

这些事最终都归结到了这起事件上。

That kind of all rolls downhill into this incident.

对于AI领域的人来说，这正好印证了许多人一直强调的观点：在尖端模型上设置防护栏是多么关键。

Now, for the AI folks, this highlights what many have been saying about how critical it is to be very conscientious about guardrails on frontier models.

这是一个例子，说明防御型研究人员能够在仍在高风险领域使用的老旧系统中发现严重漏洞。

This is an example where defense focused researchers were able to find serious vulnerabilities in ancient systems still being used in a very high stakes field.

毫无疑问，社会上还有成百上千个类似的老旧系统在运行关键基础设施。

It is absolutely the case that there are hundreds or even thousands of outdated systems like this running critical infrastructure across society.

成本是一个巨大的因素，可能还是最大的因素，导致这些系统没有被彻底改造，甚至没有进行过适当的漏洞测试。

Cost is a huge part, perhaps the biggest part, of why these systems haven't been overhauled or even properly tested for vulnerabilities.

虽然这显然不完美，但用有限的预算借助AI进行哪怕是基础的测试，对于这类软件来说可能是一个彻底的变革。

While it's obviously not perfect, the ability to do even rudimentary testing using AI on a modest budget could be an absolute game changer for this sort of software.

虽然网络犯罪分子拥有了新的强大工具确实令人担忧，但AI同样也是网络防御者的一次重大升级。

While, yes, it is scary that cybercriminals have a new suite of powerful tools, AI is and remains a massive upgrade for cyber defenders as well.

最后，今天还有一个话题，我们明天肯定会继续讨论：白宫正在召集一批AI公司，讨论政府针对尖端模型的自愿审查框架。

Lastly today, one that we will pick up on I m sure tomorrow, the White House is hosting set of AI companies to discuss the administration's voluntary review framework for frontier models.

我很期待看看会有什么结果，不过咱们还是等更可靠的报道出来再说吧。

I am very interested to see what comes out of that, but let's wait until we have some more solid reporting.

好了，今天的头条新闻就到这里。

For now, that's gonna do it for today's headlines.

接下来是本期的主节目。

Next up, the main episode.

现在人工智能领域最重要的问题之一，不是谁在用AI，而是谁用得好。

One of the most important AI questions right now isn't who's using AI, it's who's using it well.

毕马威和德克萨斯大学奥斯汀分校刚刚分析了140万次真实职场中的AI交互，发现了一个令人惊讶的现象：那些使用效果最好的人，并不是更擅长写提示词，而是把AI当作一个推理伙伴。

KPMG and the University of Texas Austin just analyzed 1,400,000 real workplace AI interactions and found something surprising: the highest impact users aren't better prompt engineers they treat AI like a reasoning partner.

他们会阐述问题、引导思考、反复迭代，并推动AI给出更好的答案。

They frame problems, guide thinking, iterate, and push for better answers.

好消息是什么？

And the good news?

这些行为是可以大规模培养的。

These behaviors are teachable at scale.

如果你想从使用AI转向真正掌握AI能力，KPMG关于高级AI协作的研究值得你花时间了解。

If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time.

了解更多请访问kpmg.com/us/sophisticated，网址是kpmg.com/us/sophisticated。

Learn more at kpmg.com/us/sophisticated That's kpmg.com/us/sophisticated.

本期《AI每日简报》由Hyperagent赞助播出，你可以在Hyperagent上运行多个智能体，让团队一起管理。

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

新用户可获得1000美元的推理额度。

New users get $1,000 in inference.

别再依赖本地智能体和等待你笔记本电脑提示的聊天工作流了。

Forget local agents and chat workflows waiting on your laptop to be prompted.

Hyperagent在云端部署始终在线的智能体，在你团队已经使用的工具中完成实际工作。

Hyperagent deploys always on agents in the cloud, doing real work across the tools your team already uses.

市场部的智能体能把竞争对手的动态转化为落地页。

Marketing's agent turns competitor moves into landing pages.

销售部的智能体能丰富潜在客户信息、起草邮件并更新CRM系统。

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

运营代理负责处理文书和跟踪预算。

Ops agent chases the paperwork and tracks the budget.

每个代理都能访问共享上下文，并遵循你关于范围和审批的规则。

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

是时候用上那些像队友一样的代理了。

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

去Hyperagent招一个吧，它是由Airtable团队打造的。

Hire yours at Hyperagent built by the team at Airtable.

在hyperagent.com/aidailybrief领取你的1000美元推理额度。

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

说个残酷的事实：你的公司可能花了成千上百万美元买AI工具，但大部分都没怎么用起来。

Here's a harsh truth: Your company is probably spending thousands or millions of dollars on AI tools that are being massively underutilized.

一半的公司有AI工具，但只有12%真正用它们创造商业价值。

Half of companies have AI tools, but only 12% use them for business value.

大多数员工还在用AI来总结会议记录。

Most employees are still using AI to summarize meeting notes.

如果你在公司负责AI的推广落地，那你需要Section。

If you're the one responsible for AI adoption at your company, you need Section.

Section是一个平台，能帮你管理整个组织的AI转型。

Section is a platform that helps you manage AI transformation across your entire organization.

它会指导员工实际应用场景，追踪谁在用AI创造业务价值，并清晰展示AI在哪些地方产生了价值、哪些地方没有。

It coaches employees on real use cases, tracks who's using AI for business impact, and shows you exactly where AI is and isn't creating value.

结果呢？

The result?

你就能从单纯部署工具，转向推动可衡量的AI价值。

You go from rolling out tools to driving measurable AI value.

你的员工也会从总结会议纪要，升级到解决真正的业务问题，而且你能证明投资回报率。

Your employees move from meeting summaries to solving actual business problems, and you can prove the ROI.

别再猜测你的AI投入是否有效了。

Stop guessing if your AI investment is working.

去sectionai.com看看Section吧。

Check out Section at sectionai.com.

网址就是 sectionai.com。

That's sectionai.com.

Blitzy 在写代码之前，会先深入理解你的整个代码库。

Blitzy deeply understands your codebase before it writes code.

第一个体现价值的地方就是：人工智能时代的安全问题。

Here's the first place that pays off: security in the age of AI.

漏洞不是孤立存在的。

Vulnerabilities don't live in isolation.

它们藏在数百万行相互关联的代码深处，你修好一个，可能悄悄就搞坏了另外三个。

They live buried inside millions of lines of interconnected code, where patching one thing quietly breaks three others.

所以表面扫描根本没用。

That's why surface level scans fail.

Blitzy 从整个应用的知识图谱出发，识别并呈现所有资产中的 CVE 漏洞，主动推荐补丁，还能直接执行 PR。

Blitzy starts from its knowledge graph of your entire application, identifies and surfaces CVEs across the full estate, proactively recommends patches, and can execute the PR.

每个修复都基于系统之间的连接和验证方式，确保不会引发新的问题。

Each fix is grounded in how your systems connect and validate so nothing new breaks.

而且知识图谱会动态更新，让你始终领先于不断加速的威胁环境。

And the knowledge graph dynamically updates, keeping you ahead of an ever accelerating threat landscape.

Blitzy 的一个客户在四天内解决了六个核心微服务中的 21 个活跃 CDE。

One Blitzy customer resolved 21 active CDEs across six core microservices in four days.

零编译错误，每次验证扫描都干净，原本计划数月的工作在不到一周内完成。

Zero compile errors, every validation scan clean, months of planned work fixed in less than a week.

基于真实架构上下文的安全修复，速度堪比计算速度。

Security remediation grounded in real architectural context at the speed of compute.

在 blitzy.com 上加固你的代码库。

Harden your codebase at blitzy.com.

网址是 blitzy.com。

That's blitzy.com.

欢迎回到《AI 每日简报》。

Welcome back to the AI Daily Brief.

今天我们来聊聊最新的模型发布——Quen 3.8 Max。

Today, we are talking about the latest model release, which is Quen 3.8 Max.

不过我们这次聊的语境跟平时有点不一样。

But the context we're putting it in is a little bit different than normal.

你知道，我刚从毕马威上周举办的科技与创新研讨会回来，最大的感触就是，过去一年里大家的讨论方向发生了多么根本性的变化。

As you know, I am just back from KPMG's Tech and Innovation Symposium last week, and my biggest takeaway was about just how radically the conversation had shifted over the last year.

去年我去参加他们2025年的那届时，大家还在认真讨论什么“有多少组织有一、两个或三个AI用例”这类话题。

When I was there for their 2025 edition, we were still genuinely talking about things like the percentage of organizations that had one, two, or three AI use cases.

而今年，我们聊的是企业如何应对人人都拥有强大编码工具后带来的复杂治理问题，以及人们如何处理成本配置问题、在组织不同部门使用不同模型，还有是否要考虑微调开源权重模型、制定开源权重模型政策等等。

This year, we were talking about how organizations are handling the complex governance issues that arise from everyone having powerful coding tools, and how people are dealing with cost provisioning issues and different models across different parts of their organization, and whether they be looking into fine tuning open weights models and have an open weights models policy.

说白了，就是企业对AI的理解深度，更重要的是，他们提出的问题质量，已经大幅提升了。

Point being that the level of sophistication around AI, and more importantly, the quality of the questions that enterprises are asking, has increased dramatically.

这意味着，我猜，第一次有相当数量的企业采购者——不只是早期采用者那种开发者类型，而是真正的企业IT人员——会在像新的Quen 3.8 Max模型发布时，主动关注起来。

Meaning that my guess is, for the first time, some meaningful number of enterprise buyers not just early adopter developer types, but actual enterprise IT type folks are actively paying attention when things like the new Quen 3.8 Max model comes out.

所以，我们先来聊聊这个模型本身。

So first let's talk about the model.

和KimiK3一样，这也是一个大模型（这个模型有2.4万亿个参数）。

Like KimiK3, it is a large model (this one coming in at 2,400,000,000,000 parameters).

确实，Latent Space 写道，如果不是最近发布的 KimiK3，这个模型本可以成为全球最顶尖的开源模型。

Indeed, Latent Space writes that the model would have been the top open model in the world, but for the recent KimiK3 release.

阿里巴巴通义千问的官方账号称其为编程和协作的新标杆，并列举了它在自主编程方面的成果：从空文件夹到生产环境，十几天内自我进化式开发，全程无需人工干预，还在 GitHub 上分享了完整的项目轨迹。他们还称赞了它在数百个专业领域内达到生产质量级别的交付成果、具备闭环自适应学习的系统级自主规划能力，以及原生的多模态智能。

The official Alibaba Quen account calls it a new bar for coding and co work, and points to examples of its work on autonomous coding: ten plus days of self evolving development from empty folder to production without handholding sharing a complete project trace on their GitHub They also laud production quality deliverables across hundreds of professions, systems level autonomous planning with closed loop adaptive learning, and native multimodal intelligence.

他们写道，视觉不仅仅是输入，更是规划、执行和自我修正的持续反馈循环。

Vision they write isn't just input, it's a continuous feedback loop for planning, execution, and self correction.

现在，至少从他们自己报告的基准测试来看，确实有一些相当亮眼的数据。

Now, at least on the self reported benchmarks, there are some pretty impressive numbers.

阿里巴巴报告了一个 Terminal Bench 分数，介于 Fable 5 和 GPT-5.6-SOL 之间；一个 CoWork Bench 分数，介于 SOL 和 Fable 5 之间；视觉推理和研究复现能力都达到了顶尖水平，等等。

Alibaba reports a Terminal Bench score that puts them between Fable five and GPT-five-six SOL, a CoWork Bench score between SOL and Fable five, visual reasoning and research reproduction that are totally state of the art, and so on and so forth.

他们还报告了在 OS World Verified 上达到了顶尖水平，这指的是智能体式的计算机使用能力——如果这个结果准确的话，这将是它作为智能体工作工具最显著的应用之一。

They also report state of the art on OS World verified, which is agentic computer use, which if that result is accurate, is one of the more significant when it comes to its applicability as an agentic work tool.

有意思的是，网上很多讨论并不是关于模型本身，而是关于他们发布模型时附带的那段视频广告。

Now interestingly, a lot of the discourse online was not about the model itself, but about the video ad that they released it with.

这个广告的概念极其简单，但执行得非常出色：画面中是一台笔记本电脑，正在执行各种不同类型的工作。

The ad, which is an incredibly simple concept very well executed, is a laptop in the foreground, doing all sorts of different types of work.

有编程工作、科研工作，还有各种类型的知识型工作。

There's coding work, science work, different types of knowledge work.

而在背景里，你能看到那些本该和笔记本电脑配对的人，正在做更多让他们享受生活的事情。

And in the background, you see the humans presumably paired with that laptop off doing more of the things that make them, well, enjoy their lives.

画面里有人钓鱼，有人打网球，有女士在攀岩，还有另一位女士在读书。

There's a guy fishing at one point, another man playing tennis, a woman rock climbing, and another woman reading.

换句话说，几乎没用什么文字，Quen 就在传达：这些极其强大、先进的模型，对个人生活的价值，就是让你有更多时间去做自己真正想做的事。

In other words, without using basically any words, Quinn is saying that the value to your life as an individual of these incredibly powerful and advanced models is to give you more time to do the things that you actually wanna do.

All In 播客的 Jason Kalakanis 发推说：中国不仅在制造越来越有竞争力的模型，还在做世界级的正面 AI 营销。

Jason Kalakanis from All In tweeted: Not only is China making increasingly competitive models, they're also making world positive AI marketing.

电脑会替我们干活，我们全都去海滩享受生活。

The computers are going to do our jobs for us, and we're all going to the beach.

Nansen 的 Alex Fanavic 写道：这是什么营销啊？

Nansen's Alex Fanavic writes: What is this marketing?

没有墓地吗？

No graveyards?

初级岗位不会消失吗？

No entry level jobs disappearing?

维托里奥还指出，他们似乎有点在调侃我们，因为其中一个例子是在法布尔严格限制下验证蛋白质来源。

Vittorio also points out that it seems like perhaps they are trolling us a little bit, with using one of the examples being verifying protein sources given these strict guardrails on Fable.

然而，这次发布的一个重要意义在于，它标志着阿里巴巴和千问系列回归开源模型。

And yet, one of the big deals about this announcement is that it represents a return for Alibaba and the Quen series back to open weights models.

去年，千问团队经历了很多人员变动，他们最大的Max系列模型从开源转向了闭源。

Last year, there was a lot of personnel shifting around Quen, and for their largest Max series models, they moved from open to closed.

事实上，这引发了一个更广泛的问题：随着中国模型越来越先进、越来越接近前沿，这种做法是否会成为常态。

In fact, it brought up the question more broadly of whether this was something that was just going to be more commonplace for Chinese models as they got more advanced and got closer to the frontier.

现在他们回来了，正如余晨金指出的，这将是千问首次开源Max级别模型的权重。

Well, now they are back, and as Yu Chenjin points out, this marks the first time Quen will open source the weights of a Quen Max class model.

完整权重将在下周发布。

The full weights will be released next week.

草莓实验室的希拉·加萨波塔写道：这对开源模型来说意义太重大了。

Strawberry Lab's Shira Gasarpota writes: This is so huge for open models.

中国实验室显然在利用Anthropic最近的公关危机，全力押注开源权重。

Chinese labs are clearly taking advantage of Anthropic's recent PR mess and are going all in on open weights.

这个信号非常明确。

The message is loud.

他们希望中国的开源模型能在全球占据主导地位。

They want Chinese open models to dominate globally.

不过，也许更大的看点在于价格。

Yet maybe the even bigger deal is the price.

我最近在节目里经常提到，很多人，尤其是政策领域的人，仍然认为所有中国模型的价格都只是美国顶尖竞争对手的零头。

One of the things that I've talked about on the show a lot recently is that many folks, particularly around the policy sector, still have this idea that every Chinese model is cents on the dollar relative to their state of the art American competitors.

但这种情况其实早就不是这样了。

That hasn't been the case for a while.

举个例子，虽然Kimi K3比Opus便宜，但也就便宜了大约40%（每百万输出token 15美元对25美元）。

In fact, to give you one example, while Kimi K3 is cheaper than Opus, it's only cheaper by about forty percent (fifteen dollars per million output tokens versus $25 per million output tokens).

而Quen则把价格大幅拉低了。

Quen, however, takes it down significantly.

千问的定价是每百万输入token两美元，每百万输出token六美元，这比Kimi便宜了三分之一左右，比Opus便宜了五分之四。

Quen is being priced at $2 per million input and $6 per million output tokens, making it a little more than a third of the price of Kimi and a fifth of the price of Opus.

虽然这还不是有些人以为的‘白菜价’，但相比Kimi K3来说，降价幅度已经大得多了。

Now that's still not the pennies on the dollar that some people assume, but it is a much more significant decrease than, for example, Kimi K3.

考虑到这些情况，很多人的第一反应都很兴奋，这也不奇怪。

Now given all of this, it's not surprising that a lot of the first impressions were excited.

来自《周四AI》播客的Alex Volkov写道：阿里巴巴要回归了吗？

Alex Volkov from the Thursday AI podcast writes: Alibaba is back?

千问3.8 Max即将开源，拥有2.4万亿参数。

QUEN 3.8 MAX is about to get open weighted with 2,400,000,000,000 parameters.

这个模型在运行了十六天的项目上，表现非常接近K3。

This model comes very close to K3 running projects for sixteen days.

不过，在这个早期阶段，还是应该保持一点谨慎。

But it is worth being at least a little cautious at this early stage.

在独立基准测试方面，Artificial Analysis确实发布了千问3.8 Max的53分，比Kimi K3低了4分，甚至比Grok 4.5还低了1分。

When it comes to independent benchmarks, artificial analysis did publish Quen3.8 Max's score of 53, which puts them four points behind Kimi K3 and even a point behind Grok 4.5.

更糟的是，人工分析那边没给任何解释就迅速撤下了那些评分，所以我们得看看究竟是怎么回事。

Now adding some injury to the whole thing, artificial analysis quickly took down those scores without any explanation, so we'll have to see what was going on with that.

在伊桑·莫利克的一些测试中，他写道：经过一系列实验，初步印象是，这是个不错的模型，但就我目前的体验来看，还达不到Kimi K3的水平。

On some of Ethan Mollick's tests he writes: Basic impression after a bunch of experiments is that it is a solid model, but not Kimi K3 level in my experience so far.

Datum写道：Quen 3.8 Max根本没法用，说他们在编码、设计、规划、智能体编排以及多个测试框架上进行了测试，并与Kimi K3、Grok 4.5、GLM 5.2、GPT Luna和Opus 5进行了对比，结果Quen每次都垫底。

Datum writes: Quen 3.8 Max is unusable, saying that they tested it across coding design, planning, agent orchestration, and multiple harnesses, and running it against Kimi K3, Grok 4.5, GLM 5.2, GPT Luna, and Opus five, with Quen coming in last every time.

他说这个模型极其缓慢、不稳定，而且消耗量惊人。

He says it's extremely slow, unstable, and burns through usage like crazy.

它经常第一次尝试就失败，第二次也失败，有时甚至第三次还失败。

It often fails on the first attempt, then the second, and sometimes the third.

帕维尔·赫伦也在他的Bug Bench测试中进行了测试，该测试包含两个真实代码库，共105个隐藏漏洞，结果发现它只找出了19个。

Pavel Huron also tested it on his Bug Bench test, which has two real repos with 105 hidden bugs, and found that it found 19 out of 105 bugs.

Kimi K3和Opus 5都领先，修复了21个。

Kimi K3 and Opus five were both ahead, fixing 21.

GPT56 SOL以42个位居第一。

GPT56 SOL led at 42.

他还提到，虽然Qwen确实修复了一个其他10个测试模型都没发现的漏洞，但光是让它跑起来就试了五次。

And while he noted that Quen did fix one bug that none of the other 10 models he tested found, that, quote, just getting it to run took five attempts.

有个编程代理半小时就用完了五小时的订阅额度。

One coding agent used up the five hour subscription quota in half an hour.

按量付费的密钥先用完了免费额度，然后一直不扣费，直到我换了个控制台才正常。

The pay as you go key exhausted its free tier, then refused to bill until I opened a different console.

总的来说，成本大约31美元，这让Palantir得出结论：我帮你省点钱和时间吧。

All in all, the cost was about $31 leading Palantir to conclude, Let me save you some money and time.

GPT 5.6 Luna跑同样的基准测试只花了1.8美元，修复了33个漏洞。

GPT 5.6 Luna ran the same benchmark for a buck 80 in fixed 33.

Grok 4.5在25分钟内修复了16个。

Grok 4.5 judged 16 in twenty five minutes.

所以初步证据表明，有充分理由对公布的基准测试数据持怀疑态度。

So first evidence suggests some good reason to be skeptical about the published benchmark numbers.

但归根结底，这次发布之所以让人兴奋，是因为它采用了开放权重，这意味着人们能以完全不同的方式拿到手。

But ultimately, why it's still an exciting release to people is the fact that it's coming in open weights, which means they're going to be able to get their hands on it in a totally different type of way.

这就回到了我节目一开始的观点：尽管这种模型发布在过去只有开发者和早期用户才会真正关注。

Which gets us to my contention from the beginning of the show that despite this being the type of model release that in the past would only be something that was really paid attention to by developers and early adopters.

但越来越多的大型主流公司和企业内部的人也开始关注这件事了。

Increasingly, there are going to be folks inside big mainstream corporations and enterprises who are paying attention to this as well.

另外，昨天《纽约时报》发表了一篇评论文章，作者是Lululemon的前首席信息官朱莉·阿弗雷尔。

Now, also yesterday, The New York Times published an opinion piece from the former chief information officer for Lululemon, Julie Averell.

那篇文章的标题——注意，标题不是作者自己起的，而是《纽约时报》的编辑起的——叫做《我帮运营过的公司别再自欺欺人地看待AI了》。

The piece which, remember, is titled not by the author of the piece but by the headline writers at The New York Times was I Helped Run Companies Need to Stop Kidding Themselves About AI.

某种程度上，这可以说是当下关于企业AI的典型文章模板。

Now, to some extent, this is the platonic archetype right now of the average essay about enterprise AI.

第一部分讲的是，很多公司想用AI来提升公关形象，而不是真正把它融入战略。

The first part is all about how common it is for corporations to want to use AI for PR value rather than actually integrating it deeply into strategy.

朱莉创造了一个词叫“AI许愿”，指的是公司领导认为AI是魔法，只要对着难题挥一挥魔杖，就能跳过解决问题的过程。

Julie coins a term AI wishing, which she defines as the belief by company leaders that AI is magic, that you can wave its wand towards a hard problem and skip the work of solving it.

但别误会，这不是在批评AI本身。

But don't mistake this as a critique of AI.

并不是。

It is not.

它批评的是那些影响AI整合的真实人为过程。

It's a critique instead of the very real human processes that impact how AI gets integrated.

她写道：别把这些误解为对AI潜力的怀疑。

She writes: Don't mistake any of this for doubt about AI's potential.

这是我见过的最强大的技术，整个行业已经在被它重塑。

It's the most powerful technology I've seen, and entire industries are already being remade by it.

制药公司正在用它重建药物发现流程。

Pharmaceutical companies are rebuilding drug discovery around it.

银行用它进行欺诈和风险检测。

Banks run fraud and risk detection on it.

各国政府将其背后的芯片视为国家安全问题。

Governments are treating the chips behind it as a matter of national security.

然后，在最重要的一句话中，她总结道：这种工作不可能在一个季度内完成，认为可以做到就是一种陷阱。

And then in the most important line, she concludes, this kind of work doesn't happen in a quarter, and believing that it can is the trap.

所以在这种情况下，我们不仅看到了AI愿景，还看到了AI洗白。

And so in this context, not only did we get AI wishing, we got AI washing.

正如她所说，这是AI愿景的阴险表亲——当一家公司迫于压力要立即展示成果时，就会声称自己在AI方面做得比实际更多。

As she puts it, the insidious cousin of AI wishing that is when a company, under pressure to immediately show results, claims to be doing more with AI than it actually is.

其中最负面、最成问题的情况就是AI裁员。

Now the most negative and problematic version of this is the AI layoff.

引用她的原话：一家公司宣称，因为AI让运营更高效，所以不需要那么多人了。

Quote, a company proclaims it needs fewer people because AI made its operations more efficient.

但往往这种效率根本还不存在。

Often the efficiency doesn't exist yet.

裁员其实是为了腾出现金，有时甚至是为了在AI上投入更多。

The cut is really about freeing up cash, sometimes to spend more on AI.

她接着说，今年5月，美国雇主宣布裁员9.7万人，而一家研究机构的数据显示，其中40%的裁员被归咎于AI。

In May, she continues, US employers announced 97,000 job cuts, and according to one research firm, companies blamed 40% of them on AI.

另一项调查发现，大约三分之一因AI而裁员的招聘经理，后来又重新招聘了相同或类似的职位。

A separate survey found that around one third of hiring managers who had cut a role because of AI had already rehired for the same or a similar one.

那当然啦。

Of course they did.

他们还没重新设计工作流程就先裁了人，结果活儿全压在了留下的员工身上。

They eliminated positions before redesigning the work, so the work shifted onto the people who remained.

然后公司又悄悄把那些号称被AI取代的岗位给招了回来。

Then companies quietly hired back some of the capability they claimed AI had replaced.

这么折腾一圈，钱烧了，人才和经验流失了，更别提那些被要求留下的员工对公司的信任也没了。

That cycle burns money and loses the talent and experience that was pushed out of the door, not to mention the trust of the people asked to stay.

说到这儿，我觉得我的观点已经老生常谈了。

Now, I think at this point, my perspective on this is very well trodden territory.

我认为，那些把AI纯粹看作降本增效工具、而不是机遇型技术的公司，短期内或许能靠几个头条新闻风光一下，但最终绝对会被那些明白这是一次重新设计、能带来巨大新机遇的公司彻底碾压——这可不是为了让股东在第三季度对削减成本感到兴奋。

I think the organizations that view AI strictly as an efficiency technology rather than as an opportunity technology might be able to eke out a few headline wins in the short term, but are ultimately going to be absolutely pummeled by the companies who understand that this is a redesign moment that opens up massive new opportunities, not a chance to make shareholders excited about cost cuts for Q3.

简单来说，想把AI用好，是需要下功夫的。

In short, AI is going to take work to do well.

这涉及到组织架构、岗位职责、工作流程的重新设计，想走捷径注定会失败。

It is going to involve organizational redesign, job role redesign, process redesign, and shortcuts are bound to fail.

那么，这两个故事是如何交织在一起的呢？

And so how do these two stories intersect?

一边是Quinn 3.8 Max，另一边是关于企业AI一厢情愿和盲目跟风的警示故事。

Quinn 3.8 Max on the one hand, and this cautionary tale about AI wishing and washing the Enterprise on the other.

关键在于，我提到的对话变化——就像KPMG那场活动所体现的——以及其中所蕴含的成熟度，表明越来越多的企业AI领导者正在提出正确的问题。

Well, what it comes down to is that the change in the conversation that I was mentioning as exemplified by the KPMG event, and the sophistication that it implicates shows that more and more enterprise AI leaders are asking the right type of questions.

例如，现在越来越多的人在讨论，开源权重模型能否成为完整AI系统的一部分，这个系统需要针对不同任务使用不同类型的模型。

For example, there is increasingly a discourse about whether open weights models can be part of a complete AI system that involves different types of models for different types of tasks.

尽管头条新闻只关注成本上限这类粗暴手段，但实际上，我接触到的每一位企业AI领导者都在深入思考：一方面希望增加AI使用量，另一方面又需要控制成本，他们到底应该采取什么样的策略。

Despite headlines which suggest only blunt tools like cost caps, the reality in practice is that basically every enterprise AI leader that I interact with is digging deep to actually understand what their approach to, on the one hand, wanting more AI consumption, and on the other, needing to manage costs, should actually be.

其中一部分讨论就围绕开源权重模型展开。

One part of that is discourse about open weight models.

一年前，几乎没有公司对这些模型有任何明确的政策，除了“哈哈，我们当然不会用中国模型”。

A year ago, almost no company had any sort of actually expressed policy when it came to these models, other than, LOL, of course, we're not gonna use Chinese models.

现在，这个话题已经彻底变了。

Now that conversation has really changed.

就在昨天，我看到一篇四位作者联合撰写的专栏文章，讲的是为什么领导者应该了解开放权重模型。

Just yesterday, I saw a four guest op ed about why leaders should learn about open weight models.

当然，常听我们节目的朋友都知道，这不仅仅是专栏文章那么简单。

And, of course, regular listeners will know that it's not just op eds.

现在越来越多的公司不仅提供更便宜的模型，还允许你根据自己组织的独特需求和数据进行定制和微调。

Increasingly, we have companies that are offering access not only to cheaper models, but to the ability to customize and fine tune them based on your organization's unique needs and data.

前OpenAI首席技术官米拉·莫拉迪创立的Thinking Machines Lab，去年年底就推出了一款名为Tinker的产品来实现这个功能。

Thinking Machines Lab, the spin off from former OpenAI CTO, Mira Moradi, introduced a product Tinker to do that late last year.

而微软，虽然显然没有使用开放模型，但正在他们成本更低的MAI模型之上构建一个前沿调优服务，明确要解决这个AI复杂度日益提升的时代的问题。

And Microsoft, while obviously not using open models, is building a frontier tuning service on top of their lower cost MAI models that is explicitly trying to solve for this era of increasing AI complexity.

当然，还有那些路由公司。

Then, of course, there are the router companies.

几乎每天都有新公司推出最新的路由产品，或者有消息显示这些公司变得多么有价值，比如有报道说Stripe打算以100亿美元收购OpenRouter。你可能会觉得路由只是取代了上一个热门的企业词汇，但根据我的经验，这绝对不是事实。

Every day, it seems, we have some new company introducing their latest router, or news that indicates how valuable these companies have come, like reports that Stripe is about to buy OpenRouter for $10,000,000,000 And while you might think that routers are just replacing the last thing as the buzzy corporate word, in my experience, that is just absolutely not the case.

我之前提到的那些企业AI领导者，是的，他们确实在探索路由解决方案，但并不是那种随便买一个Gartner推荐的最佳供应商就完事的态度。

Those same enterprise AI leaders that I was talking about before are, yes, exploring routing solutions, but not in a way where they're just hoping to buy whatever Gartner says is the best vendor, and call it a day.

正如digitalapplied.com上最近一篇文章所说，AI成本优化现在已经成了一门学科，而不是什么小技巧。

As a recent post on digitalapplied.com put it, AI cost optimization is now a discipline, not a hack.

企业们正在探索，第三方方案和内部方案怎么组合才是正确的处理方式，以及路由器到底是不是合适的解决方案。

Companies are exploring what combination of third party and internal solutions is going to be the right way to handle this, and whether or not a router is even the right solution.

但话说回来，他们能提出这类问题，本身就让我觉得非常值得乐观。

But again, the fact that those are the types of questions they're asking, I think, is worthy of a lot of optimism.

我知道，很多听众在自己公司里就是AI负责人，他们看到我描述的那种企业AI领导者的情况，可能会摇头，希望自己公司里也能是那种局面。

Now I know many listeners who are the AI leaders in their organizations will be shaking their heads, wishing that the situation I described among enterprise AI leaders were the situation that they were dealing with inside their company.

我当然不想低估那些努力推动AI正确落地的倡导者们面临的挑战有多大。

And I certainly don't wanna minimize how steep the hills to climb are for many opportunity AI advocates who are trying to get their organizations to handle this the right way.

但基本上，自从ChatGPT发布以来，我观察到，企业界对AI的主流看法第一次开始朝着更正确的方向走了。

But for basically the first time since ChatGPT launched, my observation is that the enterprise conventional wisdom around AI is getting more directionally correct.

我觉得，我们对AI那些浮夸宣传和表面文章的态度正在发生根本性的变化，以前那种靠AI概念获取公关价值或董事会认可的做法会逐渐失效，这有望切断那种错误使用AI的激励循环。

I think our attitudes around the types of things that go into AI wishing and washing are changing radically, and the PR value or board plot its that people got before are going to stop, which hopefully will cut off the incentive loop to do AI the wrong way.

如果我的判断是对的，那么接下来我们就能真正开始激动人心的工作——围绕AI的能力重新设计一切，这对我们的个人生活和职业生活都将产生深远影响。

If I am right, what we'll be left with is a situation where we can actually begin the exciting work of really redesigning around the capabilities of AI, with all of the implications for both our personal and our professional lives.

好了，今天的《AI每日简报》就到这里。

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

一如既往，感谢你的收听或收看，下次再见，祝好。

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