The Most Important AI News This Week Is Getting Overlooked — 文稿

M10:00

这周有很多值得关注的模型发布。

We got a lot of notable model releases this week.

Muse带着Wonder和Wonderful模型重新回到了前沿。

Muse jumped back to the frontier with Wonder and Wonderful.

谷歌实际上发布了Gemini 4 Argon。

Google actually released Gemini 4 Argon.

然而，看起来周四发布的两个不起眼的小模型——也就是今天节目的主题——却是所有这些发布中最不引人注目的，但它们实际上可能是最重要的，至少在AI在企业内部实际应用方面是这样。

And yet it seems like the 2 little unassuming models that were announced on Thursday that are the subject of today's show, the ones that are getting the least fanfare of all of these, might actually be the most significant, at least when it comes to the practical reality of using AI inside businesses.

《AI每日简报》是一个每日更新的播客和视频节目，涵盖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，在正式开始之前先快速通知几件事。

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

M11:14

Anthropic已经开始将第三方安全评估机构嵌入到他们的组织中。

Anthropic has begun the process of embedding third-party safety evaluators into their organization.

上个月，作为暂停前沿模型开发呼声的一部分，Dario Amodei承诺将第三方评估机构整合到Anthropic中，给予他们员工级别的权限，以监控安全流程和事故报告。

Last month, as part of calls to pause the frontier, Dario Amodei pledged to integrate third-party evaluators into Anthropic, giving them staff-level access to monitor safety procedures and reports on incidents.

在周四发布的更新中，Anthropic表示他们已经整合了第一批评估机构，也就是来自METR的六人团队。

In an update published on Thursday, Anthropic said that they've now integrated the first batch, which is of course a 6-person team from METR.

他们表示，METR的人员拥有“类似员工”的权限，这是他们的说法，包括使用公司电脑、办公空间和内部工具。

They said the METR folks have employee-like access, which is their term, including use of company computers, office space, and use of internal tooling.

Anthropic补充说，他们正在与SecureBio和Accenture作为评估机构合作，并且正在与其他组织进行早期讨论，包括Hugging Face。

Anthropic added that they're working with SecureBio and Accenture as evaluators and are in early discussions with other organizations, including Hugging Face.

在解释为什么支持这种安全监控模式时，Anthropic写道：“我们希望对我们做出的决策承担公共责任，但我们不能公开分享所有与安全工作相关的信息。”

Explaining why they support this model of safety monitoring, Anthropic wrote, "We want to be publicly accountable for the decisions we're making, but we can't publicly share all the information relevant to our safety work."

其中一些信息可能被恶意行为者利用，或者具有商业敏感性，比如计算训练信息。

Some of it could be useful to malicious actors or is commercially sensitive, like compute training information.

相反，我们给予第三方评估机构对我们信息的深度访问权限，并让他们分享关于构建模型所涉及风险的证据和合理结论。

Instead, we're giving third-party evaluators deep access to our information and are having them share evidence and reasoned conclusions about the risks involved in building our models.

我们的目标是缩小评估者对前沿AI公司系统的了解与公司自身了解之间的差距。

We aim to close the gap between what evaluators understand about a frontier AI company's systems and what the company itself understands.

现在，随着这套安全机制的推出，有两个大的讨论领域——坦白说也是批评的焦点——值得我们了解。

Now, as this safety regime gets rolled out, there are 2 big areas of discussion and frankly criticism that are worth understanding.

第一个是外部安全评估机构的独立性。

The first is the independence of the outside safety evaluators.

与一些人的说法不同，Anthropic声称他们并不资助METR，坦白说，这其中的人脉关系网有点超出了本期《头条》所能详细解读的范围。

Unlike what some have said, Anthropic claims they do not fund METR, and frankly, the web of connections is a bit beyond the scope of this particular Headlines to unpack.

然而，METR确实与Anthropic共享一些主要资助者，这些资助者都围绕有效利他主义领域。

However, METR does share major backers with Anthropic aligned around the effective altruism space.

这倒不完全是阴谋论，只是它们源于相同的哲学根源，因此最终有很多共同的资助者。

It's not necessarily full tinfoil hat, it's just that they come from the same philosophical roots and end up having a lot of the same funders because of that.

对于这种安排，人们可以持有一系列合理的看法，但这也解释了为什么Anthropic迅速推进这种特定合作，而OpenAI似乎更加犹豫。

Now, there are a range of valid opinions that could be held about the arrangement, but it's context for why Anthropic is quickly moving forward with that particular partnership while OpenAI appears a little more hesitant.

另一个大问题是实验室对安全审计报告发布有多少控制权。

The other big issue is how much control the labs have over the publication of safety audits.

本周早些时候，路透社发布了一篇报道，称Anthropic和OpenAI都在寻求对报告进行严格的编辑控制。

Earlier this week, Reuters published a report claiming that both Anthropic and OpenAI were seeking strict editorial control over reports.

在报道时，这似乎是一个灵活的谈判过程，双方似乎正在走向一个妥协方案，即删除机密信息，并注明何时进行了重大删减。

This seemed to be a fluid negotiation at the time of reporting, and the parties seemed to be moving toward a compromise that involved redacting confidential information and noting when major redactions had occurred.

关于METR，Anthropic写道，他们“不能否决发布，但如果信息属于狭窄的预定义类别，我们可以要求他们删除。”

With METR, Anthropic wrote that they, quote, can't veto publication, but we can require them to redact information if it falls into narrow predefined categories.

METR可以自由地公开描述我们的删减内容，并说明它认为这些删减是否合理，只要不透露被删减的信息即可。

METR is free to characterize our redactions publicly and say whether it believes the redactions were justified, as long as it does so without revealing the redacted information.

提到的类别包括客户数据、法律特权信息以及模型架构信息。

The categories mentioned were data about customers, legally privileged information, and information about model architectures.

Anthropic还为嵌入的审计员定义了三个层级，METR是唯一处于最高层级的，拥有员工级别的访问权限。

Anthropic has also defined 3 tiers for embedded auditors, with METR the only one in the top tier, defined by employee-level access.

SecureBio目前处于最低层级，他们开始初步工作，审查Anthropic在八月份风险报告中已经公开披露的现有化学和生物武器风险。

SecureBio is currently in the bottom tier as they begin initial work reviewing existing chemical and biological weapons risk already publicly disclosed in Anthropic's August risk report.

在该层级中，Anthropic要求审计员在发布前寻求批准。

In that tier, Anthropic requires auditors to seek approval before publishing.

Anthropic 表示，他们会评估将 SecureBio 移到第二层级。这样的话，在完成第一份报告之后，SecureBio 就能获得更多信息访问权限，而且不需要经过审批。

Anthropic said that they will evaluate moving SecureBio to the second tier, which would get them more access to information and no approval required following the completion of their first report.

Accenture 还没有开始相关工作，但 Anthropic 表示，他们希望与 Accenture 建立合作伙伴关系，涉及红队测试模型、护栏测试和对齐评估。

Accenture has not begun work, but Anthropic says they are aiming to develop a partnership with them that involves red teaming models, guardrail testing, and alignment assessments.

围绕这件事的大部分讨论，都是在说 METR 的独立性。Embryo 的创始人 Zach Korman 总结了一种很主流的观点，他写道：对于那些推动像 METR 和 AISI 这样独立评估机构的 AI 安全人士来说，这种做法的名字早就有了——安全剧场。

Most of the chatter coming out of this is about the independence of METR, with Embryo founder Zach Korman summing up a big strand of thought when he writes, for all the AI safety people pushing for independent evaluators like METR and AISI, there's already a term for this: security theater.

但我觉得对很多人来说，重要的不是起点，而是终点。

And yet I think for a lot of folks, it's less about where it starts and more about where it ends.

换句话说，如果只有 METR 一个评估机构，那就像摩根大通挑几个它喜欢的银行评估机构，然后说这就是足够独立的审计了。

In other words, if METR is the only evaluator, it's kinda like JPMorgan picking a handful of bank evaluators that it likes, and saying that's a good enough independent audit.

但最终情况似乎不太可能这样。

It seems like that's unlikely to ultimately be the scenario.

就连 METR 团队内部的一些人，比如 Parv Mahajan，也提醒说，我对此作为风险评估的原型感到非常兴奋，但重要的是要记住，这不能替代正式的监督。

Even some from the METR team, like Parv Mahajan, reminded, I'm pretty excited about this as a prototype for risk assessments, but it's important to remember that this is not a substitute for formal oversight.

METR 是在自愿协议下工作的评估者，而不是审计者，因为目前并没有既定的安全标准可供审计。

METR is an assessor working under a voluntary agreement, not an auditor, because there are no set safety standards to audit.

我认为最终这必须演变成某种更外部的审计机构。事实上，现在这可能是四大会计师事务所的一项业务。

I think ultimately this is going to have to end up in some sort of much more external auditor, In fact, this is probably a line of business for the Big 4 right now.

当然，限制在于，真正有资格做这类评估的人并不多，而这些人几乎都不在那些真正独立的机构工作。

Of course, the constraint is that there just aren't that many people qualified to actually do this sort of assessment, and pretty much none of them work for those truly independent organizations.

那么，关于前沿实验室对这一举措的立场，很多人注意到，埃隆·马斯克在 X 上回应了 Anthropic 关于这一举措的帖子，说 SpaceX 可能也会这样做。

Now, in terms of Frontier Lab alignment around this as a move, many noted that Elon Musk responded to Anthropic's post about the move on X, saying SpaceX will probably do the same.

Cognition 正在推出一个新系统，帮助组织更高效地把 token 用到上限。

Next up, Cognition is launching a new system to help organizations token max more efficiently.

今年到目前为止，我们大致看到了两种 token 花费的模式。

So far this year, we've broadly seen 2 settings for token spend.

第二季度是 token 最大化时代的盛宴，预算无限，还要求尽可能多用。

There was the feast of the token-maxing era in Q2, with unlimited token budgets and a mandate to use as much as possible.

然后是 token 紧缩时代的匮乏，组织开始对每月使用量设置上限。

Then there was the famine of the token austerity era, when organizations started placing caps on monthly usage.

实际上，大多数组织在 token 上限上仍然相当慷慨，但这也传递出一个信息：token 预算是未来的方向。

Realistically, most organizations were still pretty generous with their token caps, but it sent the message that token budgeting was the way of the future.

Cognition 的客户工程副总裁 Calvin Giroud 认为，这两种做法都是有问题的。他说，用户应该有的限制不应该是静态的。

Calvin Giroud, Cognition's VP of Customer Engineering, believes both approaches are wrongheaded, commenting, the limits that a user should have should not be static.

它们应该根据你的使用效率和你从中获得的价值来动态调整。

They should be dynamic based on the efficiency of your usage and the value that you're getting out of it.

Cognition 想出了一种有趣的方法，来为使用他们的编码代理 Devin 的工程师动态调整预算。

Cognition has come up with an interesting way to dynamically adjust budgets for engineers using their coding agent Devin.

Cognition 表示，他们可以用 AI 来检查一次会话，并估算出人类完成同样工作量所需的小时数。

Cognition says that they can use AI to examine a session and estimate the equivalent number of hours it would have taken a human to achieve the same amount of work.

当工程师到达他们的上限时，系统会评估他们使用 token 预算的效率，并给出评分。

When engineers hit their limits, the system scores how efficiently they're using their token budget.

如果分数令人满意，上限就会自动提高。

If they have a satisfactory grade, the limits automatically increase.

Giroud 说，这就像信用卡的工作原理一样。

Giroud said, this is how credit cards work, for example.

你有一个固定的信用额度，当你证明自己是一个良好的信用卡使用者，并且按时还款，你的额度就会越来越高。

You have a set credit limit, and as you prove you are a good steward of the credit card and you're paying back things on time, you get higher and higher and higher limits.

Cognition 认为，在代理式编程这个竞争极其激烈的领域里，给客户提供效率工具来建立信任，比鼓励最大程度消费能带来更好的回报。

Cognition believes that in the hypercompetitive landscape of agentic coding, giving customers efficiency tools to build trust is going to have a better payoff than encouraging maximum spend.

Giroux 评论说，无节制的 token 最大化——用他的话来说——不是一种可持续的长期方法。

Giroux commented that unrestrained token maxing is, quote, not a sustainable long-term approach.

而且我们的很多客户都是大型上市公司，他们必须就所用工具的效率，向公开市场上的股东负责。

And many of our customers are large public companies that have to answer to their shareholders in public markets around the efficiency of the tools they're using.

Cognition 还在推出一个名为 Devin Coach 的新功能，这是一个 AI 系统，可以对工程师使用代理的方式提供反馈。

Cognition is also rolling out a new feature called Devin Coach, an AI system to give feedback on how engineers are using their agent.

这个教练系统旨在识别几十种最常见的错误，比如提出模糊的请求，或者为普通任务选择昂贵的模型。

The coach is designed to pick up a few dozen of the most common errors like making ambiguous requests or selecting an expensive model for a mundane task.

开发者的主要反应是：去你的吧。

The main response from developers was screw this.

软件工程师 Jesse 写道，软件工程现在就是一个工厂岗位了。

Software engineer Jesse wrote, Software engineering is now a factory role.

而 Amisha 写道，这有种很苏联的感觉。

While Amisha wrote, there's something very Soviet about this.

Brian Chong 写道，纠结人类需要多少小时才能构建这个东西，本身就衡量错了。

Brian Chong wrote, how many hours would a human have needed to build this is measuring the wrong thing.

这其实是在鼓励人们去搭建最复杂、最令人费解的东西，而不是最有价值的东西。

It's incentivizing folks to build the most complex and bewildering thing possible, not the most valuable thing.

不过也有人意识到，就算这种具体的方法论不对，这类系统也会成为我们这个世界的一个常态。

And yet others got that even if this particular methodology is wrong, this type of system is just going to be a feature of the world that we live in now.

Token 分配将成为未来十年里最重要的技能。

Token allocation will be the most important skill in the next decade.

在政府这边，白宫人工智能特别工作组在周四召开了首次会议，为政府内部的人工智能监管制定了议程。

Over in the government realm, the first meeting of the White House's AI Task Force was held on Thursday, setting the agenda for AI oversight within the administration.

首要任务是监督一份人工智能安全报告，根据总统的行政命令，这份报告要在 120 天内发布。

The primary task is overseeing an AI safety report to be published within 120 days according to the president's executive order.

总统的科学技术顾问委员会受命准备这份报告，内容涵盖人工智能在网络安全和关键基础设施、生物安全及双重用途应用，以及模型控制与问责方面的风险。

The president's Council of Advisors on Science and Technology has been commissioned to prepare the report, which will cover AI risk in cybersecurity and critical infrastructure, biosecurity and dual-purpose applications, as well as model control and accountability.

工作组还同意成立一个工作组，来实施各个前沿实验室和 Nvidia 签署的白宫协议。

The task force also agreed to set up a working group to implement the White House Accord signed by each of the frontier labs and Nvidia.

该协议要求实验室在董事会层面监督下实施安全控制，并与外部审计机构合作。

The Accord commits the labs to implement safety controls with board-level oversight, as well as partnering with external auditors.

国防部副部长 Emil Michael 将牵头一项工作，建立报告机制，让实验室在发生安全事件时能够通知政府。

Defense Undersecretary Emil Michael will lead an initiative to establish reporting protocols to enable the labs to inform the government when security incidents occur.

与此同时，FTC 主席 Andrew Ferguson 负责在司法部和 FTC 之间创建一个执法与消费者保护协调委员会。

FTC Chair Andrew Ferguson, meanwhile, is tasked with creating a Law Enforcement and Consumer Protection Coordination Committee across the DOJ and FTC.

最后，人事管理办公室主任 Scott Kapoor 将与公共和私营部门合作，加强关键系统的安全，同时协调劳工部的工作，研究人工智能对美国劳动力队伍和经济的影响。

Finally, Office of Personnel Management Director Scott Kapoor will work with the public and private sector on the hardening of critical systems, as well as coordinating work at the Department of Labor to study the impact of AI on the American workforce and economy.

目前还悬而未决的一个问题是，政府将如何确保合规，因为白宫协议只是自愿性的，或者用总统的话说，是道义上的约束。

One thing that's still up in the air is how the government will enforce compliance, given that the White House Accord is only voluntary or, as the president put it, morally binding.

一位白宫官员告诉 Semafor，一个可能的强制手段是出口管制的威胁——今年早些时候就对 Anthropic 使用过。

A White House official told Semafor that one stick could be the threat of export controls, which was used against Anthropic earlier this year.

与此同时，现在 X 上每个人都在谈论新的特斯拉 Roadster，它基本上能飞起来，多亏了 SpaceX 的推进器。

Meanwhile, everyone on X right now is talking about the new Tesla Roadster, which basically kind of flies thanks to some SpaceX thrusters.

但这不是本周特斯拉和 SpaceX 合作的唯一大事。

But that's not the only big Tesla-SpaceX collaboration this week.

巴克莱银行的一份研究报告称，期待已久的合并可能在今年年底前出现。

A research note from Barclays claims that a long-awaited merger could be in the cards before the end of the year.

分析师 Dan Levy 在周三的一份报告中写道，合并可能会在特斯拉 12 月 9 日的年度股东大会之前宣布。

Analyst Dan Levy wrote in a Wednesday note that a merger could be announced ahead of Tesla's annual shareholder meeting on December 9th.

两家公司都没有发布任何公告，也没有向 SEC 提交任何文件。

Neither company has announced anything and nothing has been filed with the SEC.

不过，Levy 写道，虽然合并的产业逻辑曾受到质疑，但我们认为近期的管理层评论表明他们看到了合并的关键好处。

Still, Levy wrote, while the industrial logic of a combination had been in question, we believe recent managerial comments indicate they see key benefits of a combination.

Levy 说两家公司都没有聘请银行家或设立董事会委员会来启动谈判。

Levy said that neither company has hired bankers or created a board committee to begin talks.

然而，他认为合资的 TeraFab 芯片厂将是一个关键驱动力。

However, he believes the joint TeraFab chip plant would be a key driver.

他建议合并可能以全股票交易的方式进行，以大约 0.7 到 1 的比例用特斯拉股票换取 SpaceX 股票。

He suggested the merger could be structured as an all-stock deal, swapping Tesla shares for SpaceX stock at around 0.7 to 1.

巴克莱认为，这将使 Musk 拥有合并后公司 37% 的所有权和大约 56% 的投票权。

Barclays suggests that this would leave Musk with 37% ownership of the combined company and roughly 56% voting power.

到目前为止，预测市场并不相信，年终前合并的概率仍然只有 24%。

So far, the prediction markets don't believe it, with odds of a merger before the end of the year still at 24%.

不过，尽管缺乏明显的准备，但这毕竟是 Elon 和 Elon 之间的交易，而 Elon 本人已经同意了。

And yet, despite the lack of obvious preparation, this is another deal between Elon and Elon, and Elon is already in.

在上个月的一次采访中，他说，想象一下当合作如此紧密时，你会采取什么行动。

During an interview last month, he said, imagine what action one might take when there's so much close collaboration.

最后，今天还有一个很酷的后续跟进，关于我们昨天报道的——最新一代模型在影响力和扩展能力上开始变得更像 Opus 4.5 了。

Lastly today, a really cool follow-up to our story yesterday about how this latest generation of models is starting to turn out to look more like Opus 4.5 in terms of their impact and expanding capabilities.

YouTuber Max Blade 和主播朋友们用 vibe coding 搞出了一款大逃杀游戏，叫 Kong Royale。

YouTuber Max Blade and his streamer buddies have vibe-coded a battle royale game called Kong Royale.

Blade 称这是全球首款 AAA 级 vibe coding 大逃杀，看了截图，确实挺不错的。

Blade called it the world's first AAA vibe-coded battle royale, and looking at the screenshots, it seems pretty good.

游戏里有一个巨大的岛屿地图，有载具，玩家角色是一只大猩猩。

The game has a huge island setting, vehicles, and a giant gorilla as the player character.

Blade 早在 vibe coding 刚兴起的时候就有了这个想法，但他说当时的模型还做不了。

Now, Blade worked on this idea during the early days of vibe coding, but said the models just weren't ready for it.

不过有了 Opus 5.5 和几个 Claude Max 账户，他花了几周时间就搞出来了。

But armed with Opus 5.5 and a couple of Claude Max accounts, he was able to put it together with a few weeks' work.

Blade 把他们的方法叫做 hive coding，就是在他的直播聊天室里扔一堆 AI agent，真人开发者也会进来参与。

Blade calls the methodology that they used hive coding to do the work, which is basically throwing a bunch of agents into his daily livestream chat with human developers who tune in.

到了晚上，那些 agent 会回顾聊天记录，选出最好的改动，然后打包上线，第二天早上就能玩到新版本。

Overnight, the agents would review the chat, select the best changes, and ship them for a morning build.

Blade 基本上是用观众的想法来众包开发这个游戏。

Blade essentially crowdsourced the development of the game using ideas from his audience.

距离 Peter Levels 那款飞机游戏已经过去一年半了，那款游戏开启了用 vibe coding 做完整游戏的风潮。

We're now about a year and a half out from Peter Levels' plane game, which kicked off the idea of vibe coding full games.

一个独立开发者，只要有点毅力、买个 AI 订阅，现在就能做出这么厉害的东西，真的很了不起。

It is pretty amazing what a solo dev can now achieve with a bit of perseverance, an AI subscription.

Riley Brown 觉得这种社交创作模式明显是下一步。

Riley Brown thinks that this social creation is the obvious next step.

他写道，这显然是游戏的未来——主播和他们的社区一起搭建游戏，然后一起玩。

He wrote, this is obviously the future of video games— streamers and their community building the game together, then playing it.

随着模型每个月都在进步，这些游戏看起来会越来越专业。

Every month that goes by, as models improve, these games will look more and more professional.

这就是新时代的兄弟之夜。

This is the new era of bros nights.

别再在 Fortnite 里骂那些比你小一半的孩子了，去买点激浪、披萨，再拿块白板。

Instead of cussing at kids half your age on Fortnite, go get Mountain Dew, pizza, and a whiteboard.

你们一起规划自己的游戏，或者想想怎么重制或者改版你十几岁时玩过的游戏。

You map out your own game or think of a refresh or mod of a game from your teenage years.

交给 Opus 5.5 每晚循环迭代，再买一个你们平分的 Claude Max 订阅。

Give it Opus 5.5 to loop on it every night and a Claude Max sub you all split.

兄弟之夜又回来了。

Bros nights are so back.

好了，新闻快讯就到这儿。

That, however, is going to do it for the headlines.

接下来是本期主要内容。

Next up, the main episode.

KPMG 和德克萨斯大学奥斯汀分校的一项新研究发现，和 AI 一起工作时，技能相似的人结果却大不相同。

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多名早期职业人士，发现表现最好的人通过引导和评估来持续放大 AI 的价值。这些顶尖人才被称为 AI 放大器，他们的优势不在于知识本身，而在于与 AI 协作的方式。

Researchers studied more than 500 early career professionals and found that the best performers consistently amplified the value of AI by guiding, evaluating. These top performers, called AI amplifiers, weren't defined by what they knew alone, but by how they worked with AI.

M115:25

欢迎回来，这里是 AI Daily Brief。

Welcome back to the AI Daily Brief.

你敢信吗？

Well, would you believe it?

本周最重要的 AI 模型发布，其实是大家聊得最少的那几个。

The most important AI model announcements of this week are the ones that people are talking the least about.

周四，Meta 和微软都发布了新模型。

On Thursday, we got model announcements from Meta and Microsoft.

但这些模型没有一个在基准测试上刷新了行业最佳水平。

Nowhere among them are you going to see benchmarks that change the state of the art.

Artificial Analysis 也不会花大篇幅去分析它们。

Artificial Analysis isn't going to dedicate pages and pages to these.

然而，跟 X 上那些激动人心的炒作完全无关，而是跟企业使用 AI 方式的变化有关——微软和 Meta 的新模型发布其实更值得关注。

And yet, for reasons that have nothing to do with the excited hype discourse on X and everything to do with how business usage of AI is changing, Microsoft and Meta's latest model announcements deserve a lot more attention.

我们先来说说微软。

Let's talk about Microsoft first.

周四，微软的AI首席执行官穆斯塔法·苏莱曼发布了一个新的最先进的基础模型，叫做MAI Base 2。

On Thursday, Microsoft's AI CEO Mustafa Suleyman announced a new state-of-the-art base model called MAI Base 2.

这是一个2.8万亿参数的模型，和KIMI K3的规模一样。

It's a 2.8 trillion parameter model, making it the same size as KIMI K3.

它可能比SOL或Opus稍微大一点，但没Fable或Astra那么大，不过我们也不太清楚那些模型具体有多大。

It's likely a little larger than SOL or Opus, but not quite as large as Fable or Astra, although we don't have good information on how large any of those models are.

它使用了混合专家架构，有800亿活跃参数，这一点和KIMI K3的架构非常相似。

It uses a mixture-of-experts architecture with 80 billion active parameters, which is again fairly similar to the architecture of KIMI-K3.

而且这是一次完全从头开始的预训练，完全没有从其他模型蒸馏。

This is also a completely fresh pre-training run with zero distillation from other models.

现在这只是一个基础模型，也就是说它还没有经过强化学习。

Now, this is just a base model, meaning that it hasn't gone through reinforcement learning.

但微软对目前看到的结果还挺兴奋的。

But Microsoft is pretty excited about what they've seen so far.

在发布前的一次全体会议上，苏莱曼对员工说，我们现在有一个比当前驱动GPT-5.6 SOL的模型更强的基础预训练模型。

During a town hall prior to the announcement, Suleyman told staff, We now have a stronger pre-trained base model than the model that currently powers GPT-5.6 SOL.

而且实际上，我们只用了OpenAI训练他们模型所用算力的四分之一。

And in fact, we used one quarter of the compute that OpenAI used to train their model.

和OpenAI模型的对比是基于微软作为投资协议的一部分所获得的内部信息，并没有公开分享。

Now the comparison against OpenAI's models was done using proprietary information Microsoft has access to as part of their investment deal and wasn't shared publicly.

不过在公开场合，微软把MAI-2的性能和所有超过一万亿参数的开源模型做了比较。

However, in public, Microsoft compared the MAI-2's performance against all open models over a trillion parameters.

在预训练基准测试中，MAI-2在每个指标上都超过了所有模型，包括KIMI K3，而且有些指标还领先很多。

Across pre-training benchmarks, MAI-2 outperformed on each benchmark against every model, Including KIMI-K3, and by a wide margin in some cases.

微软表示，他们实现这一结果的预训练效率比DeepSeek V4 Pro高出12倍以上。

Microsoft said that they managed the result with more than 12x the pre-training efficiency as DeepSeek V4 Pro.

这个模型是在16000块GB200 Blackwell GPU上训练了38天。

The model was trained on 16,000 GB200 Blackwell GPUs across 38 days.

之后发布的技术文章有点专业，但核心信息是微软在前沿规模上成功完成了一次非常出色的预训练。

From there, the announcement post gets a little technical, but the big takeaway is that Microsoft pulled off a very successful pre-training run at frontier scale.

Prime Intellect的Elie Bukausch写道，哇，微软这篇关于新3T预训练的技术博客文章太棒了，里面提供了关于如何训练这种规模模型的非常罕见而清晰的见解。

Prime Intellect's Elie Bukausch writes, Wow, amazing technical blog post about Microsoft's new 3T pre-training with very rare and clear insights on how to train models of this size.

至于MAI-2为什么重要，对微软和对客户来说，答案是不一样的。

Now, in terms of why MAI-2 matters, there's the answer for Microsoft and the answer for customers.

在所有训练AI模型的公司中，可以说微软走向前沿的历程最曲折，至少也是最复杂之一。

Among all the companies training AI models, one could argue that Microsoft has had the most torturous journey to the frontier, or at least one of the most complicated.

最初他们决定买而不是自己造，在2019年到2022年底ChatGPT刚推出后不久，向OpenAI投资了130亿美元。

Initially, they decided to buy rather than build, investing $13 billion into OpenAI between 2019 and shortly after the launch of ChatGPT in late 2022.

那笔投资当时看起来很有先见之明，当其他大公司还在摸索的时候，微软就已经搭上了OpenAI的势头，一上来就推出了高质量的AI产品。

That investment looked prophetic, and as the other big companies bumbled around, Microsoft was able to take advantage of OpenAI's momentum and jump right out of the gate with quality AI.

但没过多久，路上就出现了一些不小的坎坷。

However, it wasn't long before there were some fairly significant bumps in the road.

最引人注目的可能是2023年底，Sam Altman被赶下台然后又重新上任OpenAI CEO，中间有一段时间看起来整个OpenAI团队都要搬到微软去。

Probably the most notable came in late 2023 when Sam Altman was ousted and then reinstated as OpenAI CEO, with a short period where it looked like basically the entire OpenAI team would go move into Microsoft.

我觉得在那时候微软领导层就很清楚了，不管他们多喜欢OpenAI、多喜欢和OpenAI模型合作，单从受托责任的角度来看，他们也得有更多的独立性。

I think at that point it became pretty clear to Microsoft's leadership that no matter how much they liked OpenAI and liked working with OpenAI models, simply from a fiduciary responsibility standpoint, they were going to need to have a bit more independence.

2024年3月，微软聘请了DeepMind的联合创始人穆斯塔法·苏莱曼来担任一个新AI部门的CEO。

In March of 2024, Microsoft hired DeepMind co-founder Mustafa Suleyman to become CEO of a new AI division.

接下来一年里，我们没怎么听到这个内部部门的消息。

For the next year, we didn't get much from that internal division.

我们只看到一些关于关系紧张的报道，比如2024年10月《纽约时报》的一篇报道说，苏莱曼在一次视频通话中对一位OpenAI员工大喊大叫，基本上是在指责OpenAI隐瞒技术突破。

All we got was reports of acrimony, like an October 2024 report from the New York Times that had Suleyman yelling at an OpenAI employee on a video call, essentially accusing OpenAI of sandbagging technological breakthroughs.

即使微软和OpenAI有这么大的合作优势，2025年他们还是过得很艰难。

Even with all of its big access to OpenAI, Microsoft struggled across 2025.

今年情况开始变了，或者说至少微软的新策略开始变得清晰了，那就是微软发布了MAI-1模型系列。

Things started to shift this year, or at least Microsoft's new strategy started to become clear when Microsoft released their MAI-1 family of models.

表面上看，这些模型没什么特别值得夸耀的。

Now on the surface, these models were nothing to write home about.

基础模型比当时的前沿模型要小，性能也只是马马虎虎。

The base model was smaller than frontier models of the era, and it was only okay on the benchmarks.

这些模型作为日常工作的通用型模型来说够用了，微软也基于这个目的把它们部署到了整个 Copilot 生态系统里。

The models were adequate as commodity models for doing routine work, and Microsoft started to deploy them across the Copilot ecosystem for that purpose.

但跟 Anthropic 差不多半年前发布的 Opus 4.5 相比，它们肯定还是差了一大截。

But they were certainly a shadow of what Anthropic had achieved with Opus 4.5, which had been released almost a half year earlier.

不过，对于任何密切关注的人来说，这些模型只是开胃菜，真正的主菜是那个新的商业模式。

However, to anyone watching closely, the models were just the appetizer to the entree, which was the new business model.

这个模式叫做 Frontier Tuning，它允许客户在微软的基础模型之上微调自己的模型。

That model was called Frontier Tuning, and it allowed customers to fine-tune their own models on top of a Microsoft base.

这件事之所以重要，是因为就在同一时间，企业通过实际应用发现，单纯订阅最前沿的模型公司、用最顶尖的模型来满足所有智能需求，根本不可能做到成本高效。

Now, why this mattered is that this was happening at exactly the same time that enterprises were realizing by seeing it in practice that simply subscribing to the most state-of-the-art company and trying to run all of their intelligence needs through the top models had no chance of being anywhere near cost efficient.

事实证明，AI 能用在做很多事情上，尤其是当我们进入智能体时代，开始用代码和软件解决工程以外的各种问题时，最终要付的 token 账单数额相当可观。

Turns out that you can use AI for a lot of things, and especially as we entered the agentic era where we began to use code and software to solve all sorts of problems even outside of engineering, the total token bill that was coming due was quite significant.

而这还是在全企业大规模真正采用这些新的智能体方法之前的情况。

And this was even before you had true mass-scale adoption of these new agentic methods across the whole enterprise.

从那以后，在商业层面，大问题就变成了如何解决这个成本问题。

Since then, on the business side, the big questions have been about how you address that.

一部分是政策上的，比如对不同人员设置新的 token 上限，以及他们需要满足什么条件才能获得更多额度。

Some of it's policy in terms of new token caps on how much different people can use and what it takes for them to get more access.

另外，现在企业变得像模型杂食动物，希望可以根据具体任务的需求，在不同模型之间切换。

Then there's the fact that enterprises are becoming model omnivores now, wanting the ability to move in and out of different models based on the needs of the particular task at hand.

而不是被锁定在某一家供应商的生态系统里。

Rather than be locked into one vendor ecosystem.

而这就是 Frontier Tuning 模式开始显得更有意思的地方。

And that's where the frontier tuning model starts to look a little more interesting.

微软 CEO Satya Nadella 一直在敲边鼓，强调企业必须拥有智能模型之上的智能层，不能把这层交给他们合作的智能公司。

Microsoft CEO Satya Nadella has been hammering the drum that it is imperative for enterprises to own the intelligence layer around the intelligence models and to not give that layer to the companies they work with.

这就需要构建更复杂的基础设施，可能包括模型路由器、开源模型或本地部署的模型。正是在这种环境下，微软推出了它的 MAI 模型和新的 Frontier Tuning 商业模式。

That requires building a more sophisticated infrastructure that involves some combination of model routers, or open models, or locally hosted models, and it's into that environment that Microsoft dropped its MAI models and its new frontier tuning business model.

然而，如果没有一个最先进的基础模型，这个方案能走多远是有限的。

And yet, without a state-of-the-art base model, there's only so far that program could go.

而这正是 MAI-2 开始重要起来的地方。

And that's where MAI-2 starts to matter.

如果这个后训练模型真的能接近最先进水平，那我们可能会看到 6 家美国公司在前沿领域竞争，而几个月前还只有 2 家。

If the post-train model actually brings it up close to the state-of-the-art, we could have 6 U.S. companies competing for the frontier, up from just 2 a few months ago.

更重要的是，在这些模型中，微软凭借其 Frontier Tuning 产品，为它认为的一个关键且日益严峻的企业挑战提供了答案。

And what's more, of those models, Microsoft, with its Frontier Tuning product, is offering an answer to what it perceives to be a key and growing enterprise challenge.

虽然其他实验室比如 Anthropic 和 OpenAI 只是通过提供更快、能力较弱但效率更高的模型来关注成本问题，但微软的 Frontier Tuning 模型也隐含地回应了数据主权问题。

While the other labs like Anthropic and OpenAI focus simply on the cost side of the equation by offering faster, less capable, but more efficient models, Microsoft's Frontier Tuning model has an implicit answer to the sovereignty question as well.

当然，在开源模型里也有针对数据主权和成本问题的其他答案。

Now, of course, there are other answers to both the sovereignty and the cost question in open models.

但考虑到微软已经深深地嵌入商业生态系统，你可以想象很多公司会更愿意和微软合作来微调模型，而不是自己去搭建基础设施，部署一个基于中国模型的开源本地模型。

But given how embedded Microsoft already is in the business ecosystem, you have to imagine that there are a lot of companies that are going to be much more comfortable working with Microsoft to fine-tune a model than going out and figuring out the infrastructure to host an on-premise model built on a Chinese base.

不过，中国开源模型可能不是那些公司的唯一选择。

Except that Chinese open models may not be those companies' only option.

Meta 这周已经发布了 MuseWonder 和 MuseWonder Full，这是一个大新闻。

Meta already made a big announcement this week with MuseWonder and MuseWonder Full.

虽然这些模型绝对让他们重新回到了美国实验室的前沿讨论中，并大大提升了他们面向消费者的 Muse 智能体的能力，但精明的人都在关注他们周四的公告，认为 Muse Spark Open 可能才是更重磅的消息。

And while those models absolutely put them back in the frontier conversation among U.S. labs, and significantly increase what their consumer-focused Muse agent can do, the smart money is looking at their announcement from Thursday and thinking that Muse Spark Open might be an even bigger deal.

Meta 称 Muse Spark Open 是目前能力最强的美国开源权重模型。

Meta is calling Muse Spark Open the most capable US open-weight model to date.

该模型总参数量为 8160 亿，激活参数 420 亿，上下文窗口为 100 万 token。

The model has 816 billion total parameters, 42 billion active, and a million-token context window.

作为对比，GLM-5-3 是一个 7530 亿参数的模型，规模差不多。

GLM-5-3, by way of comparison, is a 753 billion parameter model, so on a similar scale.

AI CEO Alexander Wang 表示，该模型专为编码和智能体任务设计，比如长周期任务、复杂软件工程和多步推理。

AI CEO Alexander Wang said the model is built for coding and agentic work like long horizon tasks, hard software engineering, and multi-step reasoning.

在基准测试方面，Spark Open 介于美国顶级开源模型（如 Inkling 和 Nemotron 3 Ultra）和中国开源前沿模型（如 GLM-5-3 和 Kimi K3）之间。

In terms of benchmarks, Spark Open is somewhere between the top US open models like Inkling and Nemotron 3 Ultra and the Chinese open frontier represented by GLM-5-3 and Kimi K3.

在编码基准上，它的分数略低于 Kimi K3，但 Meta 声称在智能体基准（如 JobBench 和 MCP Atlas）上取得了开源模型的最先进成果。

Its scores are a bit behind Kimi K3 on the coding benchmarks, but Meta claimed open state-of-the-art results for agentic benchmarks like JobBench and MCP Atlas.

Artificial Analysis给MuseOpen打了39分，这让它和DeepSeek V4-1 Flash并列总排名第17位，在开源模型里排第7。

Artificial Analysis scored MuseOpen at 39, which put it tied with DeepSeek V4-1 Flash in 17th place overall, and 7th place among open models.

它比开源模型里最先进的水平大概低了7分，但比Nemotron和Inkling足足高了13分。

It's about 7 points off the state-of-the-art for open models, but a solid 13 points ahead of Nemotron and Inkling.

Artificial Analysis还没完成对Reflection AI的Beam的测试，所以很难做直接的对比，不过在Reflection已经公布的几个基准上，MuseOpen看起来不相上下，甚至还稍微领先一点点。

Artificial Analysis hasn't completed their testing of Beam from Reflection AI, so it's difficult to get an apples-to-apples comparison there, but on the few benchmarks that Reflection has published, MuseOpen looks on par if not ever so slightly ahead.

而且跟Reflection的模型不一样，MuseOpen在主要推理提供商那里都可以直接使用，完整的权重也能下载。

And, unlike Reflection's model, MuseOpen is fully available across major inference providers with full weights available to download.

对Meta来说，这算是他们在开源权重模型训练上的一次胜利回归。

Now, for Meta, this is a triumphant return to open-weight model training.

在AI周期的早期，Meta就是开源权重模型的代名词，发布了曾经备受喜爱的LLaMA系列。

Early on in the AI cycle, Meta was synonymous with open-weight models, releasing the once-beloved LLaMA family.

特别是在2024年LLaMA 3的时代，可以说Meta在前沿领域已经紧追OpenAI了。

Particularly during the LLaMA 3 era in 2024, Meta was, one could say, nipping at OpenAI's heels on the frontier.

当时很多人都认为开源模型对Meta来说是最佳策略。

At the time, many viewed open models as the optimal strategy for Meta.

他们不需要向消费者收20美元一个月的聊天机器人订阅费，因为他们已经拥有了科技行业最赚钱的商业模式之一。

They didn't need to charge consumers $20 for a chatbot subscription, given that they already had one of the most profitable business models in tech.

Meta要做的就是扮演搅局者的角色，免费提供AI，阻止OpenAI取胜。

All Meta had to do was play spoiler by giving away AI for free, preventing OpenAI from winning.

然后就到了Llama 4。

Then came Llama 4.

2025年4月发布的时候，这个模型立刻就让人失望了，而且一周后OpenAI发布了O3，就更显得它不行了。

Released in April of 2025, the model was disappointing to people right away and started looking a whole lot worse when OpenAI released O3 the following week.

Meta从未发布旗舰版、拥有2万亿参数的Llama 4（代号Behemoth），到那年夏天，Mark Zuckerberg就开始全面重组Meta的AI部门了。

Meta never released the flagship 2 trillion parameter version of Llama 4, dubbed Behemoth, And by the summer of that year, Mark Zuckerberg had embarked on a full teardown rebuild of Meta's AI division.

当Meta在今年4月再次发布模型，也就是MuseSpark 1.0时，有些人以为他们不做开源模型了。

When Meta started releasing models again in April of this year with MuseSpark 1.0, some assumed that they were done with open models.

那时Meta已经在AI建设上投入了数千亿美元，显然不可能全部免费送出去。

By then, Meta had committed hundreds of billions of dollars to the AI buildout, and they clearly couldn't give it all away for free.

Alexander Wang承诺会回归开源模型，但有些人对此持怀疑态度，觉得他们可能不会真的去做。

Alexander Wang pledged to return to open models, but some were skeptical that they would get around to it.

但看起来，他们并不是光说不练。

It seems, however, they weren't just blowing smoke.

现在Meta已经回到了开源模型排名的榜首，至少在美国是这样。开源模型研究员Nathan Lambert写道：各位，这对我来说就像超级碗一样。我等了很久，终于等到美国的开源权重模型再次变得重要了。

And Meta has now returned to the top of the open model ranking, at least in the U.S. Open model researcher Nathan Lambert writes, guys, this is like my Super Bowl.

在今天之前，美国的模型还落后于中国2026年2月就发布的那些模型。

I waited so long for U.S. open weight models to be relevant again.

虽然还有更多工作要做，但这至少把差距缩小了一半。

Before today, the U.S. models were still behind Chinese models from as early as February 2026.

他后来补充说：如果实际表现跟分数显示的大致一样，那这个模型的使用量会爆炸式增长。

Still more work to do, but this cut the gap by at least half.

这是美国第一个具有接近前沿智能体能力的开源权重模型。

He later added, if it performs roughly like what the scores indicate, adoption of this model is going to go gangbusters.

Nathan提醒说，GLM-5-2和KIMI-K3是中国那边带来的突破，但他补充说美国模型的采用率会非常疯狂。

First open-weight model in the US with near-frontier agentic capabilities.

这会进一步加速AI的普及，因为企业可以混合搭配前沿性能用于编排和高端任务，同时还能用低成本、定制化的后训练模型来处理可重复的工作流程。

Nathan reminded that GLM-5-2 and KIMI-K3 was that step change from China, but added that US models get crazy adoption.

总的来说，这会极大地增加整个技术栈中的token消耗和工作负载。

This is going to further accelerate AI diffusion because enterprises will be able to mix and match frontier performance for orchestration and high-end tasks while being able to have low-cost and tailored post-trained models for repeatable workflows.

所以现在我们又回到了之前讨论微软时的话题：那些不太愿意把所有数据和依赖都交给OpenAI或Anthropic的企业，现在有了比几个月前多得多的选择。

Overall, this dramatically increases token consumption and workloads across the stack.

另外，关于MuseSpark Open还有一个有趣的小细节。

And so now we are right back again to the same conversation we were having about Microsoft, where enterprises who are not so keen to just hand over all their data and all their dependencies to OpenAI or Anthropic are starting to have a heck of a lot more choices than they did just a few months ago.

信不信由你，这个模型得到了苹果的第一方支持。

Now, one more interesting nugget on the MuseSpark Open story.

苹果和Meta不仅不是朋友，我甚至觉得你都不能说他们只是关系不好。

The model, believe it or not, has first-party support from Apple.

现在，Apple和Meta不仅不是朋友，我觉得你甚至不能说他们是亦敌亦友的关系。

Now, Apple and Meta not only aren't friends, I don't even think you could call them frenemies.

作为CEO，马克·扎克伯格生涯中的很大一部分时间都在努力摆脱苹果公司对移动用户和App Store的掌控。

A fairly big chunk of Mark Zuckerberg's life as CEO has been spent trying to get out from under the thumb of Apple based on its ownership of mobile users and the App Store.

但这个新模型带来了一项新的合作。

And yet this new model comes with a new partnership.

作为合作的一部分，苹果更新了他们的MLX软件工具包，以确保Spark Open能在苹果硬件上流畅运行。

As part of that partnership, Apple has updated their MLX software toolkit to ensure Spark Open will run smoothly on Apple hardware.

两家公司的CEO都特意谈了谈对方。

And both CEOs of these companies took time to talk about each other.

扎克伯格发布了一段Muse在Mac Studio上运行的视频，并说：“MuseSpark Open，今天发布权重，得益于与John Turnus和苹果的合作，在新Mac Studio上运行得非常好。”

Zuckerberg posted a video of Muse running on Mac Studios and said, MuseSpark Open, weights releasing today, runs great on new Mac Studios thanks to work with John Turnus and Apple.

Turnus回复说：“恭喜团队！”

Turnus responded, congrats to the team!

“能在你的桌面上如此高效地运行MuseSpark Open这种级别和规模的模型，真是太神奇了。”

Amazing to run a model of MuseSpark Open's caliber and scale this efficiently right from your desk.

“Mac Studio不断推动设备端AI的边界。”

Mac Studio keeps pushing the boundaries of on-device AI.

这可真是宣布你作为苹果新CEO要做出改变的一种方式啊。

Talk about a way to announce that you are going to do things differently as the new CEO of Apple.

不过，尽管在苹果硬件上运行这个模型是可能的，但这并不意味着很容易。

Now still, just because running this model on Apple hardware is possible doesn't mean it's easy.

Spark Open的完整版本需要1TB的内存，所以你需要两台顶配的M5 Ultra Mac Studio。

The full version of SparkOpen requires a terabyte of memory, so you're talking about a pair of tricked-out M5 Ultra Mac Studios.

那个型号的最高内存配置还没有发货，两台的价格大约在4万美元。

The top memory spec for that model hasn't shipped yet, and a pair is going to cost around $40,000.

所以虽然对于最发烧的专业用户来说也不太容易够得着，但对于关心私有推理的小型初创公司来说，也并非贵得离谱。

So while that's not really in reach for even the most devoted of prosumers, it's not unobtainably expensive for small startups that care about running private inference.

而且因为它是开源模型，你当然可以运行一个量化版本，塞进Mac Mini里，只要你接受其中的权衡。

And because it's an open model, you can of course run a quantized version and cram it into a Mac Mini, as long as you're happy with the trade-offs.

两家公司对于过去的问题并没有遮遮掩掩。

The companies weren't coy about their past issues.

实际上，在接受Axios采访时，Alexander Wang感谢了前苹果高管Daniel Gross（现在在Meta超智能实验室）为消除过去的隔阂、让双方良好合作所做的努力。

In fact, in an interview with Axios, Alexander Wang credited former Apple executive Daniel Gross, now at Meta Superintelligence Labs, for smoothing over the past grievances to work nicely together.

Wang说：“我们很高兴爸爸妈妈能好好相处。”

Said Wang, we're excited for mommy and daddy to be playing nicely together.

认为大多数企业会放弃那种“给我一个现成解决方案”的简单方法——比如订阅Claude或ChatGPT——是不现实的，至少短期内是这样。

Now, the idea that most enterprises are going to opt away from the easy, hand-me-a-solution sort of approaches that just getting a Claude or a ChatGPT subscription offer is unrealistic, at least in the short term.

但我确实认为，越来越多的公司，特别是那些有一定技术能力的公司，将倾向于设计更复杂的模型架构，以便根据需求和约束（可能是性能和能力需求，也可能是成本和效率需求）在不同模型之间切换。

But I do think that a growing portion of companies, especially those that do have some technical capabilities, are going to be looking to design more complex model architectures that allow them to move in and out of different models based on their needs and constraints, which could be performance and capability needs, or they could be cost and efficiency needs.

本周，Meta和Muse都向这个领域推出了大幅升级的模型，我认为美国企业将因此受益。

This week, both Meta and Muse threw majorly upgraded models into that fray, and I think businesses in America are better off for it.

这是我们当然会继续关注的趋势，但今天就到这里，这是今天的AI Daily Brief。

This is a trend we will of course continue to watch, but for now, that's going to do it for today's AI Daily Brief.

一如既往感谢你的收听或观看。

Appreciate you listening or watching as always.

下次再见，平安。

And until next time, peace.

和平。

Peace.

已剔除 2 处广告（点击展开查看）

M10:38广告 · 已剔除

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

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

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

One more shout-out, we have just announced the latest Superintelligent Premium Training Program.

We are soon kicking off our first AI Champions Bootcamp.

This builds not only your skills to use AI, but to support adoption across your enterprise.

It's anchored by 6 live sessions and has a ton of value built in, and you can find a link to it from the top of the AI Daily Brief website.

M112:49广告 · 已剔除

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

Every episode, I talk about the competition between OpenAI, Anthropic, xAI, Google, and Meta.

And if you've been listening for a while, you might have a favorite.

Maybe you think OpenAI and Anthropic can stay ahead, or perhaps Meta's open-source strategy can win out.

Whatever your view, every AI lab creates a different investment opportunity.

Harbor Capital Advisors' AI Lab Ecosystem ETF suite lets you invest in the ecosystem behind the AI lab you believe in.

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.

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

This is a paid advertisement and not personalized investment advice.

Investing involves risk, including possible loss of principal.

When I'm in a meeting, I'm fully in it.

I'm thinking about the iteration and creative back and forth it takes to actually push a goal forward.

What I'm not thinking about is capturing takeaways, tracking to-dos, or any of that.

And that's where Granola comes in.

Granola is an AI-powered notepad that captures what happens in your meetings and turns it into clean, structured notes with the decisions and action items pulled out and easy to find.

There's no setup and no configuration.

It just fits into how you already work.

For me, it means I get to stay in idea mode and Granola makes sure those ideas actually become action.

Once you try Granola on a first meeting, it is hard to go without.

You can try it totally free at granola.ai/brief.

That's granola.ai/brief to get your time back.

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

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

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

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

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

Hundreds of thousands of lines of production-ready code.

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

Another shipped an entire application with no frontend engineer.

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

Build what's next at blitzy.com.

That's B-L-I-T-Z-Y dot com.