Gemini 4 Argon, Sonnet 5.5 and What Matters with AI Models — 文稿

M10:00

进入2026年，谷歌在AI竞赛中看起来相当不错。

Coming into 2026, Google was looking pretty good in the AI race.

2025年是个好年头。

2025 Had been a good year.

DeepMind内部很多问题都得到了解答。

A lot of the questions inside DeepMind had been answered.

他们推出了有竞争力的Gemini模型。

They were putting out competitive Gemini models.

他们在推进有趣的新产品。

They were pushing forward with interesting new products.

而且他们在消费者分发、数据等方面的天然优势，让很多人非常看好他们作为一个竞争者。

And all the natural advantages that they had always had in terms of consumer distribution and data and all those things had a lot of people very bullish on them as a contender.

但后来“vibe coding”出现了。

But then vibe coding happened.

随着编码能力的提升，加上Claude Code和Codex这类新工具的强大功能，这些新能力以前所未有的方式解锁了智能体。

And with the increase in coding capabilities paired with the power of the new harnesses like Claude Code and Codex, Those new capabilities unlocked agents in a way that hadn't been possible before.

突然间，谷歌发现自己非常落后。

And all of a sudden, Google found itself very behind.

客气点说，2026年这家公司一直在追赶，但即使这么说也不完全准确。

Charitably, you would say that in 2026, the company has been playing catch-up, but even that's not really accurate to what's been happening.

今年大部分时间里，谷歌作为一个顶级模型实验室，已经完全不在讨论之列。

For most of this year, Google has been firmly outside of the conversation as a top-model lab.

然而我认为，那些对其他实验室没有党派偏见的人，永远不会完全放心地把谷歌排除在外。

And yet I think that those who didn't have a partisan bias towards one of the other labs would never be fully comfortable writing Google off.

本周，公司发布了Gemini 4，这是他们六个多月来第一款新的前沿模型。

This week, the company announced Gemini 4, their first new frontier model in more than 6 months.

从基准测试来看，谷歌似乎强势回归了。

By the benchmarks, it looks like Google is so back.

但这就是全部吗？

But is that the whole story?

那么让我们深入看看Gemini 4 Argon，以及它在当前AI竞赛中的位置。

So let's dig into Gemini 4 Argon and where it lands in the current AI race.

《AI每日简报》是一档每日播客和视频节目，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，在深入之前先快速说几条公告。

Alright friends, quick announcements before we dive in.

M11:59

特朗普总统看着OpenAI开发者日，心里想，不，绝对不行。

President Trump really looked at OpenAI Dev Day and said, nope, absolutely not.

我不想让那成为本周AI界最大的事，于是邀请了几乎所有大型AI公司的CEO到白宫，参加后来被大卫·萨克斯称为“AI的布雷顿森林体系”的会议。

I don't want that to be the biggest thing happening in AI this week, and invited basically every big AI CEO to the White House for what David Sacks would later call the Bretton Woods of AI.

这次会议产生了许多讨论，但人们注意到的第一件事是，特朗普让Anthropic的CEO达里奥·阿莫迪在会后向媒体发言。

Now, there was a lot of chatter that came out of this meeting, but one of the first things that people noticed was that Trump pulled Anthropic CEO Dario Amodei to be the one to speak to the press following the meeting.

达里奥坚称他只是在说他一贯的观点，即AI有惊人的好处，也有非常真实的风险。

Now, Dario insisted that he's just saying what he's always said, that AI has some incredible benefits and some very real risks.

然而，这一举动有人视为公开顺从，也有人视为达里奥终于开始参与游戏，他评论说：“正如总统所说，谁赢了AI谁就赢了。”

However, in an act which some saw as public deference, others saw as Dario finally playing the game, he commented, as the president has said, whoever wins AI wins.

我认为这非常重要。

I think that's very important.

处理风险的机制仍在讨论中。

The mechanism, how we address the risks, is still under discussion.

但我们都必须共同努力，确保我们能赢，而且能安全地赢。

But we all need to work together to make sure that we can win and we can win safely.

如果我们做得对，并与总统和在座各位合作，我们都能安全地获胜。

If we do this right and work with the president and everyone here, we can all win safely.

当时X上满是人们在剖析这一时刻，分析其他科技领袖的表情、达里奥紧张的抽搐，甚至研究他的穿着品味。

Now, X was absolutely filled with people psychoanalyzing the moment, breaking down the expressions from other tech leaders, Dario's anxious tics, and even dissecting his fashion sense.

虽然有人认为特朗普和其他科技领袖可能在达里奥上前发言时欺负他，但总统后来在新闻发布会上说了些好话，称“达里奥一直很棒。”

While some thought that Trump and the other tech leaders might be bullying Dario as he stepped forward to speak, The president followed up with some kind words later in the press conference, saying, Dario has been fantastic.

我们前几天晚上一起吃了晚餐，他同意所有人的看法。

We had dinner the other night and he agrees with everyone.

不管这是不是 Dario 真的在向特朗普低头，对于一个之前告诉员工 Anthropic 因为拒绝给特朗普“独裁者式的赞美”而成为目标的 CEO 来说，这确实是个重要时刻。

Whether or not this was Dario truly bending the knee, it was a significant moment for a CEO who previously told staff that Anthropic had been targeted for refusing to give, quote, dictator-style praise to Trump.

至少在眼前这段时间里，这场新闻发布会完全盖过了会议本身的话题——也就是前沿实验室负责人签署的那份超级智能协议。

At least in the immediate wake, the press conference completely overshadowed the subject of the meeting, which was the head of the Frontier Labs signing an accord on superintelligence.

这份协议只有一页纸，承诺为前沿模型的测试和部署制定内部管控措施，并且与外部审计机构合作来验证这些管控措施。

The accord is a one-page commitment to develop internal controls for frontier model testing and deployment, as well as partnering with external auditors for verification of those controls.

特朗普对媒体说，这份文件在道义上具有约束力，这暗示了AI安全测试的志愿性路线会延续下去。

Trump told the press that the document is morally binding, suggesting a continuation of the voluntary approach to AI safety testing.

特朗普还说：“世界上最顶尖的人都签了，我作为总统也签了，这确实是一种保护形式。”

Trump added, the biggest people in the world signed that, and I signed it as president, and it really is a form of protection.

他还在新闻发布会上提出了一个十人行业监督委员会的想法。

He also floated the idea of a 10-member industry oversight committee during the press conference.

但眼下他似乎挺满意，评论说：“我看到的是大量的自我约束。”

But for now he seems satisfied, commenting, I'm seeing tremendous self-policing.

整体基调——至少参会的人这么觉得——是这是重要的第一步。

The general tone, at least among the people at the meeting, was that this was an important first step.

马克·扎克伯格对媒体说：“这个想法的重点不是我们只做这一件事。”

Mark Zuckerberg told the press, the idea isn't that this is the only thing we will ever do.

而是这是一个起点，是整个行业能够达成的一份协议。

It's that this is a start and an accord the whole industry could come to.

媒体的反应很大程度上成了一次罗夏测试，测试的是人们对特朗普的态度，以及对那些科技CEO的态度——也就是说，你是不是倾向于用“寡头”这个词来形容他们。

Mostly the media response to this was a complete Rorschach test about how one feels about Trump and how one feels about the tech CEOs, i.e., whether you're inclined to use the word oligarchs to describe them.

但我确实觉得，对这份协议最持怀疑态度的评价，可能来自媒体评论员 Chuck Todd，他写道：“如果你相信科技公司能自我监管，并且希望它们决定AI如何影响社会而不用听取公众意见，那你就会觉得这份协议没问题。”

But I do think that the skeptical take on this one might be best summed up by media commentator Chuck Todd, who wrote, if you trust the tech companies to police themselves and you want them to decide how AI will impact society without any input from the public, then you'll be fine with this accord.

想从中找到AI温和派或AI现实派的立场，我觉得人们可以认同：这些人经常和政府关键领导人同处一室，而不只是从竞争的角度来看待自己的角色，这是好事。

Trying to find the AI moderate or AI realist position in this one, I think a person could agree that it's better that these folks are frequently in rooms together with key leaders in the government rather than exclusively viewing their role through the lens of competition.

但同时，也希望看到更多外部参与，尤其是得要明确那个独立的外部审计或评估机构——做独立评估的——在实际中到底是个什么样子。

While also wanting to see more external involvement, not least of which an articulation of what the independent external auditor or evaluator to carry out independent assessments is actually going to look like in practice.

也许既然这份协议已经签了，这个细节接下来就能逐步完善起来。

Perhaps now that this accord has been signed, that particular detail is one that can get built out.

午餐会议之后，特朗普总统签署了两项行政命令，标志着超级智能时代的开启。

Following the lunchtime meeting, President Trump signed a pair of executive orders marking the beginning of the superintelligence age.

第一项命令正式确认了特朗普的改名主张，其中写道：“随着这些能力的持续提升，它们不再仅仅是人工智能，而是代表着一个超级智能的新时代。”

The first made Trump's name change official, stating, as these capabilities continue to improve, they increasingly represent not merely artificial intelligence, but a new era of superintelligence.

联邦政府使用的术语应该反映这些技术的变革性能力，以及它们为美国人民创造的无限机遇。

The terminology used by the federal government should reflect the transformative capabilities of these technologies and the limitless opportunities they create for the American people.

更具功能性的是第二项命令，它建立了 America.gov 作为一个由AI驱动的政府服务单一门户。

Much more functional was the second order, which established America.gov as an AI-powered single portal for government services.

特朗普说：“有了 America.gov，联邦政府不再挡你的路，它只为你服务。”

Trump said, With America.gov, the federal government no longer stands in your way, it only stands at your service.

我们在简化它，在美化它，在把它做成该有的样子。

We're simplifying it, we're glamorizing it, we're making it what it should be.

与第二项行政命令一起，我们还看到一个类似苹果产品发布会的主题演讲，由国务卿 Marco Rubio 扮演 Steve Jobs 的角色。

Alongside that second executive order, we got an Apple-style product unveiling keynote, complete with Secretary of State Marco Rubio playing the role of Steve Jobs.

Rubio 演示了计划中的未来功能，比如能用这个门户申请护照、婚后改名、或者注册 Medicare。

Rubio showed demos of planned future functionality, like being able to use the portal to apply for a passport, change your name after getting married, or enroll for Medicare.

这些功能有点像是政府版的 MCP——用户只需提供一次信息，然后 America.gov 的智能体就会在多个政府部门之间找到并填写所有必要的表格。

The features kind of seem like an MCP for government— a user can provide their information once, then the America.gov agent will track down and complete all the necessary forms across numerous government departments.

这个效果如何还有待观察，但确实这是个真实存在的问题。

How well this works remains to be seen, but certainly it is a real problem.

比如婚后改名的难度——取决于你住在哪个州——简直离谱到难以形容。

The insane difficulty, for example, of changing your name after getting married, depending on which state you live in, is hard to overstate.

Rubio 说这些功能将从明年开始可用，行政命令要求政府部门在90天内将它们的服务整合到门户中。

Rubio said that the functionality will be available from next year, with the executive order giving government departments 90 days to integrate their services into the portal.

目前，America.gov 只作为一个政府服务的知识库，搜索29000个政府网站，根据官方来源回答问题。

For now, America.gov only functions as a knowledge base for government services, searching across 29,000 government websites to answer questions based on official sources.

这方面很多讨论都集中在安全和隐私问题上。

Now, a lot of the chatter on this centered around security and privacy concerns.

尤其是总统说了“理论上它不可能被黑进去”之后。

Particularly after the president said, it cannot, in theory, be hacked into.

等他们找到办法这么做的时候，我们就结束了。

And when they figure out a way to do that, we'll end it.

因为他们总能找到办法，对吧？

Because they always figure out a way, right?

负责设计网站的国家设计工作室在六月份因为在其设计的其他网站中加入了活动追踪工具而受到批评。

The National Design Studio, which designed the website, came under fire in June for including activity tracking tools in other websites they've designed.

虽然有些人对于提供申请政府服务所需的个人信息提出了质疑，但AI越狱者Pliny the Liberator还是对这个聊天机器人进行了反复试探，发现它的防护措施相当严密。

And while some raised questions about providing the personal information required to apply for government services, AI jailbreaker Pliny the Liberator poked and prodded at the chatbot and found its guardrails were pretty airtight.

他写道：有趣。

He wrote, Interesting.

America.gov聊天机器人会标记任何看起来像个人信息的内容，比如任何稍微像地址或密码的文字，并在被标记的文字被删除之前拒绝接受查询。

The America.gov chatbot will flag anything that appears to be personal information, like any text vaguely resembling an address or password, and refuse to accept the query until the flagged text is removed.

以前从没见过这种情况。

Never seen that before.

与此同时，如果你认为他们一起出席这次会议意味着这些公司就能逃脱所有政府审查，那可能值得注意的是，联邦贸易委员会已经对OpenAI和Anthropic的“ rogue agents”展开了调查。

Meanwhile, if you think that them showing up together at this meeting means these companies are going to escape all government scrutiny, it might be worth noting that the FTC has opened an investigation into OpenAI and Anthropic's rogue agents.

据机构消息人士向《纽约邮报》透露，联邦贸易委员会几周前对这两家领先的AI实验室展开了调查。

According to agency sources speaking with the New York Post, The FTC opened an investigation into the 2 leading AI labs a few weeks ago.

调查范围将涵盖Hugging Face事件以及自那以来披露的数十起所谓的 rogue agent事件。

The scope will cover the Hugging Face incident and the dozens of so-called rogue agent events that have been disclosed since.

调查已经进展到联邦贸易委员会正在起草民事调查要求，类似于传票。

The investigation has advanced to the stage that the FTC is drafting civil investigation demands, which are similar to subpoenas.

一位联邦贸易委员会消息人士说，该机构计划强制这些公司的高管就他们的产品以及他们声称产品可能对消费者、对美国人民构成的危险作证。

An FTC source said, the agency has plans to compel the executives at these firms to testify about their product and about the dangers they allege their products may have to consumers, to Americans.

有报告称，第三方安全研究实验室METR以及前沿实验室都可能收到这类要求。

Reports state that third-party safety research lab METR can expect a demand alongside the Frontier Labs.

从AI讨论很早开始，一个主要的讨论焦点就是认为不需要专门针对AI的新法规——政府只需执行已有的产品安全立法即可。

Now, from very early on in the AI narrative, one big thread of discourse has been focused on the idea that new AI-specific regulation wasn't necessary— that government could simply enforce product safety legislation that's already on the books.

前联邦贸易委员会主席Lina Khan持这种观点，她上个月写道：执法机构已经有权力起诉那些创造并发布危险、未经审查或存在缺陷产品的公司及其CEO。

Former FTC Chair Lina Khan was of this view, writing last month, law enforcers already have authority to charge companies and their CEOs for creating and releasing dangerous, unvetted, or defective products.

我们不应让关于新法律制度的讨论分散注意力，从而忽视了既定法律对AI并没有豁免权这一事实。

We shouldn't let discussions about new legal regimes distract from the fact that there's no AI exemption from laws already on the books.

这篇帖子得到了前AI主管David Sacks的联署，这又是一个AI安全辩论催生奇怪联盟的例子。

This post was co-signed by former AI czar David Sacks, in yet another example of the AI safety debate creating very strange bedfellows.

不过，根据联邦贸易委员会消息人士的评论，目前还不清楚该机构是否一定会寻求某种严厉的惩罚。

Still, based on comments from the FTC source, it's not clear that the agency is necessarily pushing for some sort of harsh penalty.

他们告诉《纽约邮报》：我们绝对要赢得这场AI竞赛，而且我们正在获胜，这太棒了。

They told the New York Post, we need to win this AI race absolutely, and we are winning and that's fantastic.

消息人士又带了一句政治攻击：显然另一方，就是民主党，想要摧毁这项技术，想要向我们的敌人和竞争对手投降。

Getting a political jab in, the source continued, and obviously the other side, the Democrat Party, wants to destroy this technology, wants to surrender to our enemies and competitors.

这既不是主席想干的事，也不是总统想干的事。

And that's not something that the chairman's interested in doing, and that's not really something the president's interested in doing.

话虽如此，法律还是要遵守的。

With that being said, the laws have to be followed.

人们在谈论我们需要对这些公司制定新法律，但主席一直很明确：我们已经有很多法律在册了。

People talk about how we need new laws for these companies, and the chairman's been very clear: we have plenty of laws on the books.

所以，一次调查。

So, an investigation.

这挺好的，对吧？

That's good, right?

这能安抚那些认为这些实验室在逃脱监管的人吗？

That'll appease people who think that the labs are getting a free pass?

呃，讨论的实际情况并不完全是那样。

Eh, that's not exactly what the discussion suggests.

反垄断活动家Matt Stoller写道：这是一个保护伞。

Anti-monopoly activist Matt Stoller wrote, it's a protection racket.

思路是特朗普先调查然后洗白他们，这样其他调查者就会遇到障碍。

The idea is Trump investigates and clears them so other investigators have hurdles in their probes.

前联邦贸易委员会公共事务主任Douglas Farrer写道：如果联邦贸易委员会能妥善调查这些公司就好了，最好是从两年前开始，但这次新调查的公信力是0%，尤其是在昨天AI CEO与特朗普的“爱之聚会”之后。

Former Director of Public Affairs at the FTC, Douglas Farrer, wrote, it would be good if the FTC investigated these companies properly, better starting 2 years ago, but the public credibility of this new investigation is 0.0%, especially after yesterday's AI CEO and Trump lovefest.

乐观的观点来自Joel Thayer，他写道：我有点喜欢这种“信任但核实”的策略。

The optimistic take comes from Joel Thayer, who writes, I kind of love this trust-but-verify strategy.

尽管这基本上是自我监管，但如果有任何签署方未能履行这些承诺，就可能触发FTC根据其UDAP权限第5条进行执法。

Even though it's mostly self-regulatory, if any signatories fall short of these commitments, it could trigger FTC enforcement under Section 5 of its UDAP authority.

这使得FTC主席Andrew Ferguson出席这次会议显得非常恰当。

Makes FTC Chairman Andrew Ferguson's presence at the meeting very apt.

和现在任何政策一样，看待这件事时，还是得多留个心眼。

As with anything policy right now, it's worth having a whole bowl of salt when you look at this.

但事情总归在朝某个方向前进。

But things continue to move in some sort of direction.

头条新闻就到这里。

For now, that's going to do it for the headlines.

接下来是主节目。

Next up, the main episode.

M113:04

欢迎回到AI每日简报。

Welcome back to the AI Daily Brief.

朋友们，看起来猪会飞了，地狱也结冰了——随便选一个表示难以置信的比喻——因为经过数月的等待，谷歌终于宣布了Gemini 4。

And friends, it appears that pigs are flying, hell has frozen over— choose your metaphor for incredulity— because after months of waiting, Google has announced Gemini 4.

不过，也许那些猪还没有完全离地，因为尽管他们发布了这个模型，我们实际上还无法使用它。

Although maybe those pigs aren't completely off the ground because although they announced the model, we don't actually have access to it yet.

周三，谷歌发布了Gemini 4 Argon。

On Wednesday, Google unveiled Gemini 4 Argon.

这是他们六个多月来首个前沿模型。

It's their first frontier model in over 6 months.

谷歌在2026年的困境已经被充分记载。

Now, the 2026 struggles of Google have been well documented.

我们最终没有等到Gemini 3.5 Pro，因为谷歌基本认为它还不够好，不值得发布。

We never got a Gemini 3.5 Pro, with the company basically deciding that it just wasn't good enough to release.

几周以来，外界一直在猜测Gemini 4即将发布，而Google DeepMind首席执行官Károly Kéveszoglu表示，这款模型确实不负众望。

And for a few weeks now, speculation has been coming that Gemini 4 was on the horizon, and Google DeepMind CEO Károly Kéveszoglu says that the model delivers.

在真实世界的软件工程、法律和金融等企业知识工作，以及网络安全防御等复杂工作流中，表现出前沿水准。

Frontier performance in complex workflows across real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense.

尽管在很长一段时间里，谷歌作为前沿实验室的未来看起来有些不确定，但仅从报道的基准测试来看，谷歌似乎又重新回到了竞争行列。

Now, despite a long period where Google's future as a frontier lab has seemed somewhat tentative, to judge only by the reported benchmarks, Google looks to be right back in the game.

Gemini 4 Argon是许多基准测试中新的最先进模型。

Gemini 4 Argon is the new state-of-the-art model across many benchmarks.

在智能体知识工作方面，Gemini 4在VALS Index上得分68.9%，击败了Fable 5.1、GPT-6 Astra，并且比当前领先者Opus 5.5的分数高出几个百分点。

In agentic knowledge work, Gemini 4 scored 68.9% on the VALS Index, Beating Fable 5.1, GPT-6 Astra, and exceeding the score from the current leader Opus 5.5 by a couple of percentage points.

在Automation Bench和Val's Finance Agent上，差距甚至更大。

The gap was even larger on Automation Bench and Val's Finance Agent.

在Harvey's Legal Agent基准测试中，Gemini 4的得分是当前领先者Fable 5.1的三倍多，达到19.6%。

And on Harvey's Legal Agent benchmark, Gemini 4 more than tripled the score of current leader Fable 5.1, scoring 19.6%.

智能体编码得分则参差不齐一些，但Gemini 4绝对与前沿竞争对手处于同一水平。

Agentic coding scores were a little more patchy, but Gemini 4 is definitely in the ballpark of frontier rivals.

在DeepSWE上，它达到了新的最先进水平，得分77.9%，而Opus 5.5为74.2%。

It's the new state of the art on DeepSWE, scoring 77.9% compared to Opus 5.5's 74.2%.

在Frontier Sui上，Gemini 4得分55%，比当前领先者Astra 6低10分，比Opus 5.5低7分，比Fable 5.1低1.3分。

On Frontier Sui, Gemini 4 scored 55%, which placed it 10 points behind current leader Astra 6, 7 points behind Opus 5.5, and 1.3 points behind Fable 5.1.

然而，在Frontier Sui上，Gemini 4得分55%，比当前领先者Astra 6低10分，比Opus 5.5低7分，比Fable 5.1低1.3分。

On Frontier Sui, however, Gemini 4 scored 55%, which placed it 10 points behind current leader Astra 6, 7 points behind Opus 5.5, and 1.3 points behind Fable 5.1.

在Terminal Bench 4.0上，它也垫底，得分57.4%。

It was also bottom of the pack on Terminal Bench 4.0 with a score of 57.4%.

比当前领先者Opus 5.5低9分。

9 Points behind current leader Opus 5.5.

你可以从两个角度来看待这个问题。

Now, you can look at this in 2 ways.

第一，智能体编码即便不是最重要的用例，也是最重要的用例之一，所以未能达到最先进的分数就有点成问题。

One, agentic coding is one of the most important, if not the most important use case, so failing to achieve a state-of-the-art score is kind of problematic.

或者，你可以把这看作是谷歌的一次巨大进步，而这是不可否认的。

Or you can look at this as a huge improvement for Google, which it undeniably is.

编程在过去一年一直是 Gemini 最大的弱点，所以 Gemini 4 能和 OpenAI、Anthropic 的顶尖模型处于同一水平，绝对代表着一次大追赶。

Coding had been Gemini's biggest weak point for the past year, so to have Gemini 4 be in the same ballpark as the frontier models from OpenAI and Anthropic absolutely represents a big catch-up.

在计算机使用方面，Gemini 4 紧追 Astra，而 Astra 刚刚为这个类别设立了新标准。

For computer use, Gemini 4 is nipping on the heels of Astra, which set a new standard for the category.

在 OSWorld 上，Gemini 4 得了 69.2%，Astra 是 72.6%。

On OSWorld, it scored 69.2% against Astra's 72.6%.

Artificial Analysis 确认谷歌已经重回领先模型开发者的行列。

Artificial analysis confirmed that Google is back in the mix as a leading model developer.

Gemini 4 Argon 在智能指数上得了 53 分，与 GPT-6 Astra 和 Fable 5.1 并列第三，比 GPT-6.1 Sol 高 1 分，比 Sonnet 5.5 低 3 分，比 Opus 5.5 低 5 分。

Gemini 4 Argon scored 53 on the Intelligence Index, putting it tied for 3rd place with GPT-6 Astra and Fable 5.1, 1 point ahead of GPT-6.1 Sol, and 3 and 5 points behind Sonnet 5.5 and Opus 5.5 respectively.

在成本和效率方面，它很接近帕累托前沿，在 AA 基准测试中每次任务花费 1.99 美元，而 Astra 是 3.26 美元，Fable 5.1 是 7.63 美元。不过，这是谷歌 50% 的发布折扣价，折扣持续多久没说，所以原价会比 Astra 更贵。

In terms of cost and efficiency, it's pretty close to the Pareto frontier, costing $1.99 per task on the AA benchmark run, compared to $3.26 for Astra and $7.63 for Fable 5.1, however, that is with Google's 50% launch discount, which will be available for an unstated period, meaning that the full price would make it more expensive than Astra.

它处在一个尴尬的中间地带，即使打折也比 GPT-6.1 Sol 贵一倍多，但基准测试的提升又不够明显。

It's also in that uncomfortable middle ground where it's more than twice as expensive as GPT-6.1 Sol, even with the discount, without a super clear improvement on the benchmarks.

不过，目前这些都无关紧要，因为可惜的是谷歌并没有向公众开放 Gemini 4 Argon。

For the moment, however, none of this really matters, because unfortunately Google is not making Gemini 4 Argon generally available to the public.

他们给出的理由是关于网络安全的。

Their stated reasoning has to do with cybersecurity.

这个模型在 CWE Bench 网络安全基准上得了 68%，与 Grok-47 和 GPT-6 Astra 并列，略高于 Opus 5.5、Fable 5.1 和 Mythos 5.1。

The model scored 68% on the CWE Bench cybersecurity benchmark, tied with Grok-47 and GPT-6 Astra, and slightly ahead of Opus 5.5, Fable 5.1, and Mythos 5.1.

谷歌认为，这个理由足以让他们只对所谓的可信网络安全防御者开放访问权限，以确保模型不被误用。

This was reason enough, according to Google, to limit access to what they call a set of trusted cyber defenders to ensure the model is not misaligned.

他们表示正在参与美国政府的自愿测试和发布前访问计划，并打算逐步扩大访问范围，先从 API 和 Google Ultra 订阅用户开始。

They said that they're engaging with the U.S. government's voluntary testing and pre-release access program and intend to gradually expand access over time, beginning with API and Google Ultra subscribers.

在博客中，谷歌说这个模型已经在为内部工作流提供支持，并且在长周期编码任务（比如大规模代码库迁移）上表现出色。

In their blog post, Google said that the model is already powering internal workflows and boasting strong performance on long-horizon coding tasks like large-scale codebase migrations.

至于选择宣布模型却不实际发布的原因，我不太确定。

Now as to the choice to announce the model without actually releasing it, I'm not really sure.

说实话，对于我们这些关注已久的老手来说，这有点让人想起谷歌第一次落后时的情形——2023 年 12 月，他们觉得有必要宣布 Gemini，尽管 Pro 版本几个月后才会推出。

Honestly, for us old heads who have been watching this for a while, it kind of has echoes of the first time Google fell behind and felt the need back in December of 2023 to announce Gemini even though the Pro version of the model wouldn't be available for a number of months.

不过也许这是个正确的决定，因为有很多人看到它都很兴奋。

Yet maybe it was a good call because there were plenty of people who were excited to see it.

在 X 上有一整类帖子，基本意思就是：谷歌，回来了。

There was an entire genre of posts on X that basically came down to Google, so back.

Daria Anutmaz 博士写道：“哇哦，哇哦，Google DeepMind 带着 Gemini 4 Argon 回来了。”

Dr. Daria Anutmaz wrote, wow, wow, Google DeepMind is back with Gemini 4 Argon.

基准测试成绩惊人。

Insane benchmarks.

一下子，他们就跳到了 AI 前沿的最顶端。

In one fell swoop, they've jumped to the very top of the AI frontier.

就像我之前说的：永远不要在 AI 时代与谷歌对着干。

As I've said before, never bet against Google in the age of AI.

不过他又补充了一句——在我看来这是关键细节——：“迫不及待想尽快试试 Gemini 4。”

Although he did add, and this seems to me to be the key detail, can't wait to try Gemini 4 ASAP.

Ethan Mollick 写道：“又变成三强争霸了。”

Ethan Mollick writes, and it's a 3-way race again.

Nathan Lambert 是做开放模型研究的，他几乎不可能是个炒作派，他写道：“很高兴看到谷歌用 Gemini 4 给大家带来惊喜。”

Nathan Lambert, who does open model research and is about as far away as you could be from a hypester, writes, love to see Google surprising people with Gemini 4.

更多实验室跻身前沿，对消费者来说是好事，因为有了竞争；对世界也是好事，因为权力更分散了。

More labs at the frontier is wonderful for consumers because of competition and wonderful for the world because of reduction in concentration of power.

很期待它在真实场景中的表现。

Excited to see how it goes in real-world scenarios.

I rule the world 写道：“我过去一直批评谷歌的模型，但 Gemini 4 在基准测试上看起来很强。”

I rule the world writes, I've been critical of Google models, but Gemini 4 looks strong on benchmarks.

我们以前在谷歌身上见过这种情况，但我乐观地认为他们能像 Codex 那样，发布一个一流模型并配上第一流的配套工具。

We've seen this before with Google, but I'm optimistic they can release a first-class model with a first-class harness like Codex.

向参与 Gemini 4 的所有人致敬。

Great work to all involved with Gemini 4.

为你们加油。

Rooting for you.

但硬币的另一面是Angel，他写道，抱歉，但我受不了这些“谷歌回归”的帖子。

But then on the other side of the coin is Angel, who writes, I'm sorry, but I can't with all these Google is so back posts.

我们根本都没法用这个模型，而且谷歌模型的基准分数从来就不靠谱。

We literally can't even use the model, and it's not like benchmark scores for Google's models have ever been reliable.

我们真的需要记住Gemini 3 Pro发生了什么吗？

Do we really need to remember what happened with Gemini 3 Pro?

说真的吗？

Seriously?

Ishu Agarwal提供了一些数据，分享了一组较旧的基准测试，并补充道，顺便提一下，这些是Gemini 3.1 Pro发布时的官方基准。

Ishu Agarwal put some numbers around that, sharing an older series of benchmarks and adding, by the way, these were the official benchmarks for Gemini 3.1 Pro when it released.

几乎全面碾压Opus 4.6。

Practically destroying Opus 4.6 across the board.

我不是说Gemini 4 Argon会很差，但考虑到谷歌过往的记录，竟然没有人有丝毫怀疑，这太疯狂了。

I'm not saying Gemini 4 Argon will be bad, but it's crazy that no one has the slightest bit of skepticism given Google's track record.

公平地说，Logan Kilpatrick——他确实应该获得谷歌今年的MVP，因为他在所有这些公关挑战中始终坚守——实际上回应了这一点，他写道：我们现在在谷歌内部已经大大改进了大规模测试模型的能力，所以假设大多数新的Gemini版本在发布前都会经过数千名软件工程师数周的测试，这有望真正缩小基准测试与现实之间的差距。

To their credit, Logan Kilpatrick, who honestly should get Google's MVP for hanging in there throughout all of these PR challenges this year, actually engaged with this, writing, We've gotten much better at testing our models at scale across Google now, so assume most new Gemini revs go through thousands of software engineers for weeks before getting released, hopefully has helped close the benchmark-to-reality gap by a real margin.

到目前为止，我们所依据的只是报告的基准测试，以及公司内部的一些抱怨声音。

Now, so far, all we have to go on is the reported benchmarks, along with some reports from grumbling inside the company.

一篇彭博文章《谷歌应对员工对新Gemini 4的怀疑》称，虽然Gemini 4在广泛用于衡量模型效率的基准测试中表现良好，但当员工实际将其投入工作时，表现就不那么好了。

A Bloomberg piece called Google Grapples with Employee Skepticism About New Gemini 4 says, while Gemini 4 has performed well on benchmarks widely used to gauge model efficiency, it does less well when employees actually put it to work.

据能够直接接触该模型的人士称，该模型在处理某些编码任务时存在困难。

The model struggles to handle certain coding tasks, according to people with direct access to the model.

谷歌方面否认了这一报道，评论称，声称Gemini 4在编码等某些领域表现不佳是不准确的。

Google, for their part, denied the reporting, commenting that it would be inaccurate to claim that Gemini 4 is underperforming in some areas such as coding.

彭博社的另一位消息人士表示，这些抱怨者是少数，并声称公司内部有“广泛的共识”，认为Gemini 4是前沿模型。

And one of Bloomberg's other sources said that these gripers are a minority, claiming that a, quote, large consensus internally at the company believe that Gemini 4 is the frontier.

谷歌的处境之所以如此具有挑战性，是因为在他们不在顶峰的这段时间里，竞争的性质已经发生了根本性的变化。

What makes the position of Google so challenging is that in the time that they've been away from the top, the nature of the race has fundamentally changed.

Peter Yang非常乐观地写道，谷歌在Gemini 4上大显身手。

With much optimism, Peter Yang wrote, Google cooked on Gemini 4.

现在他们只需要在Coding Harness、Antigravity和personal agent Spark上更具竞争力。

Now they just need to be more competitive on Coding Harness, Antigravity, and personal agent Spark.

换句话说，竞争已经不仅仅是模型本身了。

In other words, the competition has become more than just models.

而且，这也不是说模型竞争已经不那么激烈了。

And it's not like there already isn't a ton of model competition.

在OpenAI Dev Day之前发布的一个我们还没来得及讨论的模型是Claude Sonnet 5.5。

One model released in advance of OpenAI Dev Day that we didn't have a chance to talk about yet is Claude Sonnet 5.5.

如果你觉得自己有点想翻白眼，或者至少是质疑自己是否真的该关心，那当然可以理解，因为Sonnet和Opus系列的近期历史确实提供了一些值得怀疑的理由。

And it's certainly if you feel yourself giving a bit of an eye roll, or at least a raised eyebrow wondering if you should actually care, The recent history of the Sonnet and Opus series certainly created some reason for skepticism.

你还记得吗，Opus 5很不受欢迎，而Sonnet 5似乎以更高的价格提供了更差的结果，因为它消耗的代币太多了。

Opus 5, you remember, was strongly disliked, and Sonnet 5 seemed to deliver even worse results for a higher price because it was so token hungry.

然而，如果你必须从过去几周的节目中选出一个主题，那就是Anthropic借助Opus 5.5回归正轨，受到大家喜爱。

However, if you had to pick just one theme from the last couple of weeks on this show, it's the beloved return to form for Anthropic with Opus 5.5.

那么Sonnet 5.5是否延续了这一趋势？

So does Sonnet 5.5 continue that trajectory?

简短的回答似乎是，在大多数人看来，是的。

The short answer seems to be, in most people's estimation, yes.

Anthropic表示，Sonnet 5.5比Sonnet 5快30%，便宜30%，并且该模型在基准测试上展现了一些显著的提升。

Anthropic said that Sonnet 5.5 is 30% faster and 30% cheaper than Sonnet 5, and the model demonstrates some significant jumps on the benchmarks.

在agent decoding方面，它在TerminalBench 4.0上得分70.6%，而Sonnet 5仅为10.3%，甚至超过了Opus 5.5的66.4%。

For agent decoding, it scored 70.6% on TerminalBench 4.0, which is up from just 10.3% for Sonnet 5, and even beats Opus 5.5 at 66.4%.

在FrontierCode和CursorBench上的得分比Opus 5.5略低。

Scores on FrontierCode and CursorBench were very slightly behind Opus 5.5.

在知识工作基准测试GPT-Val-AA和NAA-Briefcase上也是如此，Sonnet 5.5缩小了之前Sonnet与OpenClass模型之间的巨大差距。

The same was true for knowledge work benchmarks GPT-Val-AA and NAA-Briefcase, where Sonnet 5.5 closed the previously massive gap between Sonnet and OpenClass models.

Artificial Analysis证实了这些基准测试，给Sonnet 5.5打出了56的智力指数得分。

Artificial Analysis affirmed the benchmarks, giving Sonnet 5.5 an intelligence index score of 56.

这使得它排在Opus 5.5之后，位居第二，却领先于Fable 5.1和GPT-6 Astra。

That puts it in second place behind Opus 5.5 and somehow ahead of Fable 5.1 and GPT-6 Astra.

不过，他们在测试中发现，Anthropic 声称的成本节省并没有实现。

However, they didn't find that Anthropic's claims about cost savings stood during their testing.

Sonnet 5.5 每次任务花费 7.60 美元，而且消耗的 token 数量比 Sonnet 5 多得多，但每个 token 的价格是一样的。

Sonnet 5.5 cost $7.60 per task, and used significantly more tokens than Sonnet 5 at the same per-token price.

这意味着 Sonnet 5.5 的基准测试成本几乎和 Fable 5.1 一样高，比 Opus 5.5 贵 27%，比 GPT-6 Astra 贵一倍以上。

This meant that Sonnet 5.5's benchmark run was almost as expensive as Fable 5.1, 27% more expensive than Opus 5.5, and more than twice as expensive as GPT-6 Astra.

平心而论，Anthropic 声称在类似任务上相比 Sonnet 成本降低了 30%。

Now, to give Anthropic the benefit of the doubt, they claimed a 30% cost reduction on similar tasks compared to Sonnet.

主要的 AA 基准测试是在最大推理设置下运行的，但如果把设置调到超高，成本就能降低三分之二。

The main AA benchmark is run at max inference settings, but turning down the settings to extra high cut the cost by two-thirds.

不过，等大家真正上手之后，这个模型还是让人印象深刻的。

Still, the model impressed once people got their hands on it.

在社交媒体上很受关注的渲染测试中，它比 Sonnet 5 有了巨大提升。

It is a massive improvement over Sonnet 5 in the rendering tests that grab attention on social media.

Matthew Berman 做了一系列视觉测试和游戏克隆，他写道：Sonnet 5.5 基本上就是 Opus 5.5，但便宜 50%，而且快得多。

Matthew Berman made a series of visual tests and game clones, writing, Sonnet 5.5 basically Opus 5.5 but 50% cheaper and much faster.

我提前测试了一下，效果太棒了。

I've been early testing it and it's incredible.

如果这就是前沿发展的节奏，那算我一个！

If this is pacing the frontier, sign me up!

但就实际使用而言，开发者 Kun Chen 发现，只有在任务合适的时候它才表现最好。

But for real-world use, builder Kun Chen found it performed best when it's the right tool for the job.

他写道：Sonnet 绝对不等同于 Opus，不只是便宜 50% 那么简单。

He wrote, Sonnet is definitely not the same as Opus, just 50% cheaper.

只有对那些不需要太多智慧的问题来说才成立。

That's true only for problems that don't need much wisdom.

当要求它们为模糊的产品问题提出方案时，两者的区别就很明显了。

There's a clear difference when they are asked to propose plans for ambiguous product problems.

Opus 能够以更具战略性的思维来解决问题，比如思考真正的目标是什么，而 Sonnet 则更倾向于战术层面，只想着怎么完成任务。

Opus is able to approach problems with more strategic thinking, i.e., what's the real goal here, while Sonnet is more just looking at tactically, how do I get this done?

现在，Kun 的做法是把 Anthropic 的模型当作一个完整的系统来用：用 Opus 做规划，Sonnet 做执行，出了问题再找 Fable。

Now, Kun's current approach is to use Anthropic models as a complete system, with Opus for planning, Sonnet for implementation, and calling on Fable when things go sideways.

Pavel Huren 用他的 bug 修复基准测试了 Sonnet，发现它的表现超过了包括 Fable 和 Astra 在内的所有其他模型。

Pavel Huren tested Sonnet on his bug fix benchmark and found it outperformed all other models including Fable and Astra.

他写道：秘诀是什么？

He wrote, the secret?

Sonnet 5.5 Max 对于大多数任务来说可能既便宜又快，但它是我测试过的最不偷懒的模型。

Sonnet 5.5 Max might be cheap and fast for most tasks, but it's the least lazy model I tested.

它在 bug 排查中领先，是因为愿意多花几个回合。

It leads in bug hunting by spending turns.

有意思的是，博主兼创业者 Theo 的评价是：Sonnet 5.5 是个很棒的模型，但你可能不该用它。

Fascinatingly, YouTuber and entrepreneur Theo's review was that Sonnet 5.5 is an incredible model, that you probably shouldn't use.

在他看来，这个模型相比 Sonnet 5 有了很大提升，但在最大推理模式下会过度思考，反而容易出错。

In his view, the model is a big improvement over Sonnet 5, but on max reasoning it overthinks and risks getting things wrong because of it.

而在较低的设置下，Sonnet 又因为太能吃 token，根本找不到比 Opus 更划算的点。

And on lower settings, there's no point where Sonnet is more cost-effective than Opus because of how token-hungry it is.

当把 Sonnet 当成子代理来用时，这个情况会略有改观，在某些任务上效率更高，但通常它不适合作为独立模型使用。

This changes a little when using Sonnet as a subagent, where it can be more efficient on certain tasks, but it's usually not optimal as a standalone model.

那么，经过几周的测试，这些结论还会成立吗？

So will these results stand after a couple weeks of testing?

Sonnet 5.5 本身堪称技术奇迹，毕竟它能在发布仅四个月后就达到 Fable 5 的水平，但 token 消耗太大，对大多数情况来说都不划算。

With Sonnet 5.5 being a technical marvel given that it can produce the results of Fable-5, just 4 months after the release, but being too token hungry to make sense for most?

这当然是我会持续关注的点，等大家用得多了，我再回来汇报。

That's certainly what I'll be watching and I'll report back as people get more reps in.

不过，回到 Gemini 4 现在要参与竞争的新局面，Peter Yang 一直在强调的一点是：光模型好是不够的。

Still, going back to the new world that Gemini 4 has to compete in, the point that Peter Yang again was making is that just being good on the model isn't enough.

实际上，Muse 的成功引出了一个问题：一款用户体验好但模型并非最先进的 AI 产品，是否比一款用户体验一般但模型最先进的产品更强？

In fact, one really interesting question brought up by the success of Muse is whether an AI product with a great user experience but a less-than-state-of-the-art model beats a product with a less good user experience but a state-of-the-art model.

Muse 的月活用户已经达到了 300 万。

Muse has hit 3 million weekly active users.

每天至少发送一个提示词的日活跃用户已经达到一百万。

Daily users, who send at least one prompt per day, have reached 1 million.

记住，这个产品才上线三个多星期，所以这是一个很强的早期积极信号。

Now remember, the product has only been available for a little over 3 weeks, so this is a strong positive early indication.

第一周结束时，Muse 有五十万周活跃用户和二十五万日活跃用户。

At the end of the first week, Muse had 500,000 weekly active users and 250,000 daily users.

一方面，这只是 Meta 分发潜力的一小部分——Meta 能触达全球一半的人，而且这跟 ChatGPT 的增长完全没法比，但更好的对比可能是 Codex 的采用曲线。

Now on the one hand, this is a tiny fraction of Meta's distribution potential, which reaches half the people on Earth and is nowhere near as strong as ChatGPT's growth, but a better comparison might be the adoption curve for Codex.

Codex 大约花了三个月才达到三百万周活用户，所以 Muse 目前看起来是个增长更快的 AI 应用。

Codex took about 3 months to reach 3 million weekly active users, so Muse currently looks like a faster-growing AI app.

你可能会觉得这也正常，毕竟 Muse 是个人代理，而 Codex 面向的是非常开发者导向的受众，但不管怎样，Muse 的早期成功还在继续。

Now one might expect that, given that it is a personal agent versus Codex's very developer-focused audience, but whatever the case, the early success of Muse continues on.

然而现在 OpenAI 的 DOTS 来了，我们开始会对这个问题——好的体验加差的模型，还是差的体验加好的模型——有一些直接答案。

And yet now that OpenAI's DOTS is here, we're going to start to have some direct answers to this question of good UX plus worse model or worse UX plus better model.

Eureka 联合创始人 Sina 写道，用了一段时间 Muse 并且很喜欢之后，我开始注意到底层模型不够聪明。

Eureka co-founder Sina writes, After using Muse for a while and loving it, I'm starting to notice the underlying model isn't smart enough.

我不确定是不是上下文过载，但它最近经常出错。

I'm not sure it's context overload, but it gets stuff wrong a lot recently.

也许是因为他们太成功了，所以不得不换成一个不同的、更便宜也更差的模型。

Maybe they've had such huge success that they've had to switch to a different, cheaper, worse model.

总之，Muse 很棒，产品形态也很好，但作为消费者，我对谁都没有忠诚度。

Anywho, Muse is great, form factor is great, but as a consumer, my loyalty is to none.

如果 OpenAI 推出一个带更好模型的 Muse，那我就用最好的那个。

If OpenAI will drop Muse with better models, I will use whatever is best.

智能仍然是护城河。

Intelligence is still the moat.

然而反过来，过了几天，我确实看到不少对 DOTS 工作方式的抱怨。

And yet on the flip side, a couple days in, I'm definitely seeing some gripes with the way that DOTS works.

X 上的 Habibi Slop 写道，整体感觉不太好。

Habibi Slop on X writes, overall, the vibes are not good.

这感觉方向就错了。

This feels directionally wrong.

我只想和模型对话。

I just want to talk to the model.

这个工具是想让你在一间不允许碰管道的房子里当水管工。

This tool wants you to be a plumber in a house where you're not allowed to touch the pipes.

DOTS 内置了太多摩擦，让人觉得它压根没打算被使用。

DOTS has so much friction built in that it feels like it's not meant to be used.

除非你觉得这只是一般的吐槽账号，不然他们接下来还会举大概十几个例子。

And unless you think that this is just a general griper account, They then go on to include about a dozen examples.

我看到被反复提及的最大一点出现在他们的最后感想里，他们写道，他们发布了一个新产品，想帮你管理在 ChatGPT 和 Codex 里的工作，却没有足够权限访问它所需要的来自那些平台上的工具和数据。

The big one that I've seen repeated comes in their final thoughts when they write, they've launched a new product that wants to help you manage your work in ChatGPT and Codex without having sufficient access to any of the tools and data it needs from those surfaces.

我当然远远没到愿意去宣布，专注模型还是专注产品用户体验，哪一边才是制胜策略。

I am certainly at least not even close to being willing to declare which is the winning strategy between focusing on the model versus focusing on the product user experience.

最后一点，当我们纵观谷歌准备发布 Gemini 4 时所面临的竞争态势。

One last note as we look across the state of the competition that Google finds itself in as it gets ready to release Gemini 4.

这周有一个意外发布，DoorDash 宣布他们的 AI 代理也得到了 Muse 的待遇，现在可以通过短信接单了。

In one surprise launch this week, DoorDash has announced that their AI agent is getting the Muse treatment and can now take orders via text.

DoorDash 网站上本来就有代理，但他们利用 Muse 这类个人代理的一些功能，把这事又推进了一步。

DoorDash already had an agent on their site, but they're using some of the features of personal agents like Muse to take it a step further.

顾客现在可以用自然语言提示词给 DoorDash 代理发短信，剩下的就交给代理自己搞定。

Customers can now text the DoorDash agent with a natural language prompt, and the agent will figure out the rest.

DoorDash 举了个例子，一位顾客发短信说，把我平时点的蛋白质碗送到办公室。

DoorDash gave the example of a customer texting, order my usual protein bowl to the office.

然后代理会把手机号和用户档案匹配起来，查到常点的订单和送货地址，再处理支付。

The agent then matches the phone number to their user profile, looks up the usual order and delivery address, then handles the payment.

现在，鉴于亚马逊决定屏蔽 Muse 代理，一个巨大的悬而未决的问题是，最好的代理策略到底是与品类领导者合作，还是构建一个平台专属代理。

Now, in light of Amazon's decision to block Muse agents, there's a huge open question of whether the best agent strategy is to partner with the category leaders or build a platform-specific agent.

目前来看，DoorDash 做代理的逻辑已经相当清晰了。

The logic for an agent from DoorDash seems pretty clear at this point.

该公司表示，在购买杂货时，代理订单的购物车金额高出50%。

The company has said that agentic orders have a 50% higher basket value when buying groceries.

而通过自建代理，DoorDash可以尝试保持与用户的紧密关系。

And by building their own, DoorDash can attempt to hold the user relationship close.

当然，存在一个风险，即像Muse这样的外部代理可以直接从餐厅下单，从而绕过DoorDash这个中间商。

There's a risk, of course, that external agents like Muse could order directly from restaurants and cut DoorDash out as a middleman.

与此同时，用户偏好可能会拒绝平台专属代理，因为用户对它们的控制力弱得多。

At the same time, user preference might reject platform-specific agents because they have much less control.

目前还不清楚什么是最优策略，但眼下DoorDash似乎两方都在做——既保持平台对第三方代理开放，同时也在自建内部代理——正因如此，这值得关注，即使你不太在意最终谁胜出、你如何订午餐到办公室这件事。

It's unclear what the optimal strategy will be, but for now, DoorDash seems to be doing a bit of both— keeping their platform open to third-party agents while also building an internal agent— and for that reason it's worth paying attention to, even if you don't much care who wins how you order lunch to the office.

总之，回到Gemini 4 Argon——我赞同Nathan Lambert的观点，他认为谷歌表现出色既对消费者更有利，也对竞争更有利，所以我坚定地站在希望这次发布大获成功的人群中。

Anyways, bringing it back to Gemini 4 Argon— I'm with Nathan Lambert when he argues that Google doing well is both better for consumers and better for competition, So put me firmly in the camp of people who are hoping that this release is great.

我们就希望它快点发布，这样我们就能亲自判断，而不是只能从那些自报的基准数据和向Bloomberg抱怨的员工口中捡点残羹冷炙。

Let's just hope we get it soon so we can actually decide for ourselves, rather than dining from the table scraps of self-reported benchmarks and gripey employees talking to Bloomberg.

今天的AI Daily Brief就到这里。

That's gonna do it for today's AI Daily Brief.

一如既往感谢你的收听或观看，我们下次再见。

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

已剔除 2 处广告（点击展开查看）

M11:18广告 · 已剔除

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

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].

Also note that we have our next cohorts of Superintelligence Agent Training coming up.

These are paid programs, which in very short order will get you far ahead when it comes to your agentic understanding and your ability to use agents in your daily work.

We have both the Executive Catch-Up Program and the agent intensive, which we call the Executive Agent Leadership Program.

The next cohorts for those start next week, and you can find links to all of that at the very top of aidailybrief.ai.

M110:18广告 · 已剔除

If you're leading AI inside an enterprise, you already know that the gap right now isn't capability but execution.

That's why KPMG's You Can with AI is back with a new season featuring conversations with leaders like Surojit Chatterjee of Emma.

Habib of Ryder, McKesson CIO Ellery Fisher, and others focused on practical execution, what's working, what's not, and what it actually takes to move from pilots to real scaled impact across strategy, data readiness, governance, workforce, and value.

And of course, it's co-hosted by me, Nathaniel Whittemore.

Go listen and subscribe at www.kpmg.us/aipodcasts.

That's www.kpmg.us/aipodcasts.

The best teams don't have a single star carrying everyone else.

They know their own strengths and each other's weaknesses and play to both.

That's the team Robots and Pencils has built on purpose.

Nobody there is grinding through busywork to pad a headcount number.

People come for the hard problems, and they stay because everyone around them is leveling up at the same time.

In a market full of companies that are just trying to hire fast, that's worth a look.

Check out robotsandpencils.com/careers.

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

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 SpaceX AI.

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 Ecosystems 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. 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.