为何“缪斯探索”正在分化AI用户

看来我们正式迎来了前沿AI竞赛的新玩家。

It appears we officially have a new player in the frontier AI race.

本周早些时候，Meta刚刚发布了一系列新模型，叫做Wonder。

Earlier this week, Meta has just announced a new series of models that it calls Wonder.

与过去几个月发布的Spark模型不同，Wonder的目标是与那些大公司正面竞争，正如Meta首席AI官亚历山大·王所说，与那些“大块头”一较高下。

And unlike the Spark models it released over the last several months, Wonder is meant to compete head-on with, as Meta chief AI officer Alexander Wang put it, the big boys.

但它能撑得住吗？

But can it hang?

今天我们要来看看Wonder和它更大的兄弟Wonder Full，并思考Meta重新坚定地回到前沿AI竞赛意味着什么。

Today we're going to look at Wonder and its bigger sibling, Wonder Full, and ask what it means that Meta is now firmly back in the frontier AI race.

《AI每日简报》是一个每日播客和视频节目，讨论AI领域最重要的新闻和话题。

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

好了，朋友们，在开始之前先快速说几个通知。

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

首先，感谢今天的赞助商：KPMG、cData、Granola和Blitzy。

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

想要收听无广告版本，请访问 patreon.com/aidailybrief，或者你也可以在 Apple Podcasts 上订阅。

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

如果想了解赞助节目的详情，请发送邮件到 sponsors@aidailybrief.ai。

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

今天我们先从最新一家强调速度的公司说起。

We kick off today with the latest company to put some emphasis on speed.

过去几个月里，我们越来越频繁地看到这种情况。

This is something we've seen increasingly throughout the last couple of months.

显然，谷歌的 Flash 模型一直强调这一点，但最近 OpenAI 也在力推他们的超快模式，而 Anthropic 也找到了办法，通过新增的快速模式，让 Opus 5.5 以惊人的速度运行。

Obviously, it's something that Google has emphasized with their Flash models, but more recently, OpenAI has also been pushing their ultra-fast mode, and Anthropic has now figured out how to serve Opus 5.5 at a blistering pace with a new addition to fast mode.

这次他们似乎加快了节奏，现在 Opus 的最高速度已经验证达到每秒 285 个 token。

They seem to have stepped up the pace for this one, now serving Opus at a verified 285 tokens per second at max.

OpenAI 还没有公布 GPT 61 Sol 在快速模式下的速度测试结果，但他们的文档显示理论最高速度是每秒 280 个 token。

OpenAI hasn't published a speed test for GPT 61 Sol on their fast mode, but their docs suggest a theoretical maximum of 280 tokens per second.

而且，OpenAI 对高级速度的收费要高出 50%。

What's more, OpenAI is charging 50% more for the premium speed.

虽然大家在争论哪个更好，但实际上，这两个都是价格不菲的产品版本，应用场景也比较有限。

And while people debate which is better, realistically, both are very expensive versions of the product with fairly niche applications.

至于Anthropic的版本，使用量并不包含在常规订阅里，这个周末有些人就因此吃了亏。

For Anthropic's version, usage isn't included in regular subscriptions, as some found out the hard way this weekend.

不过，对于那些想在预算内快速消耗大量token的人来说，Chubby指出，Anthropic每美元能提供的token数多出50%。

Still, for those looking to burn through a ton of tokens quickly on a budget, Chubby noted that Anthropic can serve 50% more tokens per dollar.

我觉得目前这些快速和超快速模式对大多数人来说还有点新鲜。

I think that right now these fast and ultra-fast modes are a little bit of a novelty to most people.

但随着它们变得越来越普遍和流行，模型延迟的大幅变化会改变我们与它们的互动方式，这种变化可能要到真正体验时才能完全理解。

But as they become more common and de rigueur, big shifts in the latency of models will change how we interact with them in ways that I don't think that we'll quite know until we get there.

这最终在AI竞争中到底有多重要，我们拭目以待，但可以肯定的是，目前所有实验室都在探索这一点。

How important this ultimately is as a vector of AI competition, we will see, but it's certainly something that all the labs are exploring at the moment.

说到Anthropic，最近有新的Anthropic模型在测试中被发现了。

Now, speaking of Anthropic, new Anthropic models have been spotted in testing.

有用户报告说，他们的查询被路由到一个新模型，而不是Opus 5.5，而且结果非常出色。

Users have reported queries being routed to a new model instead of Opus 5.5 with very impressive results.

大家都在展示用 Opus 5.5 推出的代码动画，但后来好像都转到这个新模型上了。

People were showing off code-based animations that were introduced with Opus 5.5 but then seemingly pushed to this new model.

还有人报告说，Fable 的提示词也被路由到了新模型，这暗示可能会同时发布两个模型。

Others reported seeing Fable prompts routed to a new model as well, opening the possibility of a double-headed release.

Chubby 在 X 上写道：“发布应该快了。

Chubby on X wrote, "A release should be near.

因为 Anthropic 的 IPO 也快到了，他们希望新模型能提前几周发布，这样市场就能真正把它计入估值。

Since Anthropic's IPO is also pretty close, they want their new model to be released a few weeks ahead so the market can really count that into the valuation.

长话短说，Fable 5.5 应该很快就会到，而且看起来比我们预期的进步要大得多。”

Long story short, Fable 5.5 should arrive soon, and it looks to be a much bigger leap than we even expected."

除了竞争动态，这次发布也可能是一个重大考验，看看各大实验室未来愿意发布什么样的模型。

In addition to the competitive dynamics, this release could also be a big test of what the major labs are comfortable releasing moving forward.

OpenAI 和 Anthropic 都限制了对它们最先进内部模型的测试，并且可能通过专注于发布像 Opus 和 Sol 这样的中型模型来承诺控制前沿速度。

Both OpenAI and Anthropic have restricted testing of their most advanced internal models and could commit to pacing the frontier by focusing on midsize model releases like Opus and Sol.

或者，我们可能会看到新一轮的加速，因为新的更大的前沿模型被释放出来。

Or we could see another round of acceleration as new larger frontier models are unleashed.

至于AI业内人士对此有什么预期，谷歌的帕特里克·托尔梅写道：“昨晚我在纽约参加了一个AI晚宴。

As to what AI insiders expect from this, Google's Patrick Tolme wrote, "I was at an AI dinner last night in New York City.

讨论了一些有趣的话题。”

Some interesting topics discussed."

其中一个话题，用他的话说就是：“大家一致认为Fable 5.5会惊艳亮相，但OpenAI会在12月用Belle击败它。”

One of those topics was, as he put it, "Consensus was Fable 5.5 will be stunning, but OpenAI will beat it in December with Belle."

换句话说，如果你以为假期能歇口气，那可就大错特错了，朋友们。

If you thought, in other words, about slowing down for the holidays, think again, my friends.

说到OpenAI，据报道该公司正在探索通过合作来复兴Sora。

Speaking of OpenAI, the company is reportedly exploring a Sora revival as part of a partnership.

哦，那是很久以前的事了，这款TikTok风格的视频生成应用曾经风光一时。

Oh, so long ago, the TikTok style video generation app enjoyed a brief moment in the sun.

2025年底，Sora在App Store排行榜上登顶，把ChatGPT挤了下去。

In late 2025, Sora grabbed the number one spot on the App Store chart, displacing ChatGPT.

大约五分钟的时间里，Sora的梗成了热门话题。

For about five minutes, Sora memes became the hot trend.

不过，到了今年年初，面对算力部署的艰难抉择，像编程这样的企业应用场景胜出，Sora就被搁置了。

However, by the beginning of this year, when faced with difficult choices around where to deploy compute, enterprise use cases like coding won out, and Sora was mothballed.

四月份的关闭是OpenAI践行其所谓“停止支线任务”使命的最佳例证。

The shutdown in April was OpenAI's best example of its mandate to end what it called side quests.

但现在，复出可能正在酝酿之中。

Now, however, a comeback could be in the works.

据《The Information》报道，OpenAI正在与一家未具名的初创公司洽谈，计划授权Sora模型并围绕它重新推出服务。

The Information reports that OpenAI is in talks with an unnamed startup to license the Sora model and relaunch services around it.

这家隐秘的初创公司据称由前OpenAI研究员、曾领导Sora项目的Bill Peebles在近几个月内创立，联合创始人包括前迪士尼董事长、好莱坞传奇人物Jeffrey Katzenberg，以及前Dropbox首席财务官Sujay Jaswa。

The stealth startup was reportedly founded in recent months by Bill Peebles, the former OpenAI researcher who led the Sora project, and co-founders include former Disney chairman and Hollywood legend Jeffrey Katzenberg, as well as former Dropbox CFO Sujay Jaswa.

消息人士称，这笔交易仍在推进中，尚不清楚这家初创公司是否能获得Sora的完整权重，但他们将获得足够的权限来对模型进行深度定制。

Sources said the deal is still coming together, and it's unclear whether the startup would get full access to Sora's weights, but they would have enough access to heavily customize the model.

所以，那些怀念疯狂猫咪表情包的人，你们可能很快就能重拾乐趣了。

So for those who miss those crazy cat memes, you might be back in business soon.

在收购新闻方面，亚马逊正在洽谈收购Decart.ai，这可能是他们在AI时代最大的一笔收购。

In acquisition news, Amazon is in talks to acquire Decart.ai in what could be their largest acquisition of the AI era.

《华尔街日报》报道说，这笔交易正在以70亿美元的估值进行谈判，可能很快就会达成。

The Wall Street Journal reports that the deal is being negotiated at a $7 billion valuation and could come together soon.

Decart对亚马逊来说是个有意思的收购目标，这也能在很大程度上说明他们如何看待自己在AI竞赛中的位置。

Decart is an interesting target for Amazon and could say a lot about how they view their position in the AI race.

这家初创公司成立于2023年，一直专注于开发用于实时视频的世界模型。

The startup was founded in 2023 and has focused on developing world models for use in live video.

他们的模型可以生成实时生成视频，比如用于直播中的角色实时变身，或者生成式游戏。

Their models can produce live generative video for use cases like live character transformations for streams or generative gaming.

虽然技术很酷，但Decart并没有获得太多市场关注。

While the tech is cool, Decart hasn't found all that much traction.

与此同时，亚马逊自去年12月以来就没有发布过新系列模型，而且在上一轮裁员中解雇了大部分AI研究人员。

Amazon, meanwhile, hasn't released a new series of models since last December and let go most of their AI researchers in the last round of layoffs.

在大多数人看来，他们似乎无意在模型层面与前沿或通用大语言模型竞争。

It appeared to most that they had no interest in competing on the model layer for either frontier or commodity LLMs.

所以，他们的策略似乎更侧重于硬件优化。

The play then seems to be about hardware optimization.

去年，Decart 提前获得了亚马逊 Trainium3 芯片的使用权，并开发了一套他们称之为 Decart 优化系统（简称 DOS）的技术栈。

Decart was given early access to Amazon's Trainium3 chips last year and developed a technology stack they called the Decart Optimization System, or DOS.

他们声称这项技术能让推理速度提升八倍，成本降到原来的八分之一。

They claim this technology allows inference to operate up to eight times as fast and at one-eighth the cost.

DOS 与硬件无关，因此可以帮助为 Nvidia GPU 或 Google TPU 优化的 AI 模型在亚马逊的硬件上高效部署。

DOS is hardware agnostic, so could help AI models optimized for Nvidia GPUs or Google's TPUs deploy efficiently on Amazon's hardware.

到目前为止，DOS 似乎只用于 Decart 自己的世界模型，所以还不清楚这项技术能否普遍适用于大语言模型。

So far, it seems like DOS has only been used for Decart's world models, so it's unclear whether the technology can apply generally to LLMs.

但亚马逊似乎愿意赌一把。

But Amazon appears to be willing to take the gamble.

上个月，有消息称亚马逊曾考虑以 60 亿美元收购，但在进行尽职调查后放弃了。

Last month, Amazon was said to be exploring an acquisition at $6 billion but walked away after performing due diligence.

Decart 上次融资是在五月份，估值达到 40 亿美元，这意味着亚马逊愿意支付比上次估值高出 75% 的溢价。

Decart last raised funds in May, reaching a valuation of $4 billion, pointing to Amazon paying a 75% premium on their last mark.

如果这笔交易达成，显然重点在于推理优化。

If the deal comes together It's clearly about inference optimization.

如果这项技术能大规模应用，那AWS就会在竞争对手面前占据巨大优势。

If the technology works at scale, it would give AWS a huge edge over their rivals.

我们看到行业关注点已经从单纯的产品和模型，转向了如何以最低成本、最高效率地提供token服务。在这种背景下，亚马逊似乎正在为这样一个世界做准备——在那里，硬件效率是唯一重要的衡量标准。

Given how much we're seeing industry attention shift from models and products alone to the business of serving tokens as cheaply and efficiently as possible, Amazon potentially appears to be preparing for a world where hardware efficiency is the only benchmark that matters.

还有一笔有趣的收购正在进行中：据报道，英伟达正在洽谈收购美国开源模型开发商Reflection AI。

One other interesting acquisition in the works, Nvidia is reportedly in talks to acquire US-based open model developer Reflection AI.

你可能还记得，经过漫长的等待，Reflection AI上周终于发布了他们的首个模型，名叫Beam。

You might remember that after a long wait, Reflection AI unveiled their first model, called Beam, last week.

这个模型被定位为与中国开发者最好的开源权重模型竞争，虽然它在基准测试上还差一点，但Beam显然已经领先于其他美国开源模型，比如英伟达的Nemotron。

The model was positioned to rival the best open-weight models from Chinese developers, and while it wasn't quite there on the benchmarks, Beam was still clearly ahead of other US open models, like Nvidia's Nemotron.

《金融时报》现在报道说，在模型发布后，英伟达想加大对Reflection AI的投资。

The "Financial Times" now reports that Nvidia wants to double down on Reflection AI following their model reveal.

英伟达已经是Reflection最大的投资者之一，持有8亿美元的股份，据报道他们正在探索各种选项，包括以现金或算力形式追加投资，以及全面收购并整合团队。

Nvidia is already one of the largest investors in Reflection, with an $800 million stake, and are reportedly exploring options, including further investment in cash or compute, as well as a full acqui-hire.

Reflection在3月份的最新估值是250亿美元，所以这笔交易可能会超过英伟达在12月以200亿美元收购Groq的金额，成为他们历史上最大的一笔收购。

Reflection was last valued at 25 billion in March, so this deal could exceed the $20 billion Nvidia paid for Groq in December to become their largest acquisition ever.

《金融时报》说，这笔交易可能会在几周内达成，但也可能谈崩。

The "FT" said a deal could be reached in the coming weeks but could still fall apart.

现在很明显，英伟达想在西方打造最好的开源模型体系。

It is very clear at this point that Nvidia wants to build the best open model stack in the West.

金玉晨写道：“英伟达收购了Hugging Face，收购了Poolside和Essential AI的团队，现在又盯上了Reflection AI。”

Yuchen Jin wrote, "Nvidia bought Hugging Face, acqui-hired Poolside and Essential AI, and is now eyeing Reflection AI.

黄仁勋正在组建这些‘Neo实验室’，来训练美国最好的开源模型。”

Jensen is assembling these 'Neo labs' to train the US' best open source models."

关于竞争的问题，Zephyr写道：“很明显，英伟达现在要和前沿实验室正面竞争了。”

As to the question of competition, Zephyr writes, "It's pretty clear that Nvidia will compete head on with the frontier labs now.

黄仁勋很恼火，因为那些实验室在开发自己的ASIC芯片，抢英伟达的生意。”

Jensen is pissed because labs are developing their own ASICs and undercutting Nvidia."

不过，Suthrini的Jukon写道：“我觉得英伟达收购的目的不是要和OpenAI或Anthropic竞争。”

Suthrini's Jukon, though, writes, "I don't think Nvidia's intent with this acquisition is to compete with OpenAI or Anthropic.

如果是那样的话，英伟达为什么还要投资OpenAI的这轮融资呢？”

If it were, why would Nvidia invest in OpenAI's current funding round?

这大概是为了改善Nemotron那糟糕的表现。

It's probably about improving Nemotron's terrible performance."

他接着写道：“假设这次收购是为了让英伟达自己变成一个人工智能实验室。”

Continuing, he wrote, "Let's say this acquisition is intended to turn Nvidia into an AI lab itself.

那只会促使OpenAI和Anthropic加速摆脱对英伟达的依赖，这对英伟达来说更不利。

That would only push OpenAI and Anthropic to accelerate their shift away from Nvidia, which would be even less in Nvidia's interest."

不管出于什么逻辑，很明显英伟达对市场的开源方面非常感兴趣。

Whatever the logic, it is very clear that Nvidia is very interested in the open source side of the market.

而且，就在这个话题正热的时候，微软恰好发布了他们下一代自研芯片，并计划大规模部署。

And with perfect timing to add intrigue to exactly this conversation, Microsoft has unveiled the next generation of their homegrown chips with a plan to deploy them at scale.

首席执行官萨蒂亚·纳德拉周一发布了新的Maia 300芯片，称这是微软迄今为止在定制硅片上最大的一次押注。

CEO Satya Nadella announced the new Maia 300 chip on Monday, calling it Microsoft's biggest bet on custom silicon to date.

他声称，新芯片每瓦特能处理的令牌数比领先的GPU硬件多40%。

He claimed the new chips can deliver 40% more tokens per watt than leading GPU-based hardware.

从明年开始，他写道：“我们将把Maia系列扩展到千兆瓦级部署，使其成为我们自家产品和服务的主要推理系统。”

Starting next year, he wrote, "We will scale our Maia family to multi-gigawatt deployments, making it the primary inference system for our own product and services."

Maia 200芯片已经在小规模部署了，但这一代预计性能会有大幅提升。

The Maia 200 chips are already deployed at a smaller scale, but this generation is expected to be a big leap in performance.

投资人Chamath Palihapitiya写道：“你应该预期芯片公司会向上游发展，而应用公司会向下游发展。

Wrote investor Chamath Palihapitiya, "You should expect the silicon chip companies to move up the stack as the application companies move down the stack.

在AI领域，越来越明显的是，每个人都在和所有人竞争。

In AI, it is increasingly true that everyone competes with everyone.

很快每个人都会推出自己的LP、芯片、云和应用。

Everyone will soon offer their own LPs, silicon, cloud, and applications.

问题是，你能围绕什么建立长期市场和利润壁垒？

The question is, what will the constraint be that you can build a long-term market and margins around?

是分销渠道吗？

Is it distribution?

是数据护城河吗？

Is it a data moat?

还是像电力这样的物理限制？”

Is it a physical constraint, like power?"

未来几个月，这类讨论会越来越多，但今天的热点新闻就先到这里。

Expect a lot more discussion along those lines in the months to come, but for now, that is going to do it for today's headlines.

接下来是本期的主节目。

Next up, the main episode.

大家好。

Hello, everyone.

AI带来的一个重大变化是，我们的思路从如何给网页排名，转向了如何成为AI足够信任、愿意用来回答问题的信息来源。

One big change around AI is we've shifted our thinking from how we rank our pages to how do we become the source that AI trusts enough to answer with?

在毕马威，他们对此深有体会。

At KPMG, they're seeing this firsthand.

AI生成的结果现在直接给出答案，往往连一次点击都不需要。

AI-generated results now surface answers directly, often without a single click.

因此，他们越来越专注于生成式引擎优化，也就是GEO，通过结构化内容，让AI系统能够检索、理解，并将其作为可信的权威来源引用。

That's why they are increasingly focused on generative engine optimization, or GEO, structuring content so AI systems can retrieve it, understand it, and cite it as trusted authority.

这不仅仅是SEO的演变，更是关乎可见度的必然要求。

This is not just an SEO evolution but a visibility mandate.

确实，KPMG提出的GEO要求很简单：如果AI在影响决策，你的专业知识就需要出现在答案里。

And indeed, the GEO mandate from KPMG is simple: If AI is shaping decisions, your expertise needs to show up inside the answer.

详情请访问kpmg.com/us/geo。

Read all about it at kpmg.com/us/geo.

再说一遍，网址是kpmg.com/us/geo。

Again, that is kpmg.com/us/geo.

本期节目由cData赞助播出。

This episode is brought to you by cData.

正在部署AI的公司都遇到了同样的难题。

Companies rolling out AI keep hitting the same wall.

Token消耗不断攀升，安全问题拖住了AI代理的后腿，而且没人真正了解AI到底是怎么被使用的。

Token spend is climbing, security is holding agents back, and no one knows how AI is truly being used.

根本原因是什么？

The root cause?

AI与业务运行所依赖的系统和上下文脱节了。

AI is cut off from the systems and context your business runs on.

这就是为什么cData打造了Connect AI Gateway，一个连接你的AI和系统的控制点。

That's why cData built Connect AI Gateway, one control point between your AI and your systems.

它建立在为Palantir、Google、Salesforce以及数百家公司提供支持的连接层之上，每个代理都能获得它所需的数据和工具。

Built on the connectivity layer powering Palantir, Google, Salesforce, and hundreds more, each agent gets exactly the data and tools it needs.

它能精确到记录级别进行管控，让你终于能在同一个地方看到整个组织的AI使用情况。

It governs down to the record and lets you finally see your org's complete AI usage in one place.

准确的答案，更低的token成本，以及一个每次交互都能学习你业务的网关。

Accurate answers, a fraction of the token cost, and a gateway that learns your business with every interaction.

详情请访问cdata.ai。

See how at cdata.ai.

当我开会时，我会全身心投入。

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

我思考的是迭代和创意上的来回碰撞，这些才能真正推动目标前进。

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

我不会去想记录要点、追踪待办事项这些事，而这就是Granola发挥作用的地方。

What I'm not thinking about is capturing takeaways, tracking to-dos, or any of that, and that's where Granola comes in.

Granola 是一个由人工智能驱动的记事本，它能捕捉你会议中的内容，并将其转化为清晰、结构化的笔记，把决策和行动项单独提取出来，方便查找。

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.

无需设置，也无需配置。

There's no setup and no configuration.

它完全融入你现有的工作方式。

It just fits into how you already work.

对我来说，这意味着我可以一直保持创意状态，而 Granola 确保这些想法真正转化为行动。

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

一旦你在第一次会议上试用 Granola，就很难再离开它了。

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

你可以在 granola.ai/brief 上完全免费试用。

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

访问 granola.ai/brief，把时间拿回来。

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

Blitzy 在写代码之前，会深入理解你的代码库。

Blitzy deeply understands your code base before it writes code.

这是第一个见效的地方，AI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.

它们深藏在数百万行相互关联的代码中，修补一个地方往往会悄悄破坏另外三个地方。

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

这就是为什么表面扫描会失败。

That's why surface level scans fail.

Blitzy从整个应用的知识图谱出发，识别并展示整个系统中的CVE，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，所以不会引入新的问题。

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形势。

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

一位Blitzy客户在四天内解决了六个核心微服务中的21个活跃CVE。

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

零编译错误，所有验证扫描都干净通过，原本计划几个月的修复工作不到一周就完成了。

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

安全修复基于真实的架构背景，速度堪比计算速度。

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

去Blitzy.com加固你的代码库。

Harden your codebase at Blitzy.com.

就是B-L-I-T-Z-Y点com。

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

欢迎回到AI每日简报。

Welcome back to the AI Daily Brief.

就在不久之前，肯定不到一年，大多数关注此事的人基本上都把Meta从AI竞赛中排除了。

It wasn't very long ago at all, less than a year certainly, that most people paying any sort of attention were more or less writing Meta out of the AI race.

事实上，在2025年，关于Meta和AI的头条新闻大多是马克·扎克伯格花巨资吸引研究人员加入Meta，当然，还有这些举动之后引发的内部纷争。

In fact, in 2025, most of the headlines around Meta and AI were about Mark Zuckerberg spending an extraordinary amount of money to lure researchers to Meta, and then, of course, all of the internal strife that would follow those moves.

然而，在2026年这个智能体之年，当然，2026年秋季是个人智能体的季节，Meta完全重新回到了前沿实验室的行列。

And yet, in 2026, the year of the agent, with fall of 2026, of course, being the season of the personal agent, Meta is fully back in the mix as a frontier lab.

我们从Muse Spark模型的发布开始看到了这一点，这些模型虽然不是最顶尖的，但也绝对不丢人，而且足以支撑他们消费者应用Muse的许多功能。

We started to see this with the release of the Muse Spark models, which were, if not state of the art, certainly not embarrassing, and good enough to power many of the functions for their consumer app, Muse.

然而，周一Meta推出了Muse Wonder，这是一个更大的新模型系列。

On Monday, however, Meta introduced Muse Wonder, a new larger class of models.

这次发布包括两个模型变体：Wonder和Wonder Full。

The release included two model variants, Wonder and Wonder Full.

Wonder是精简版，优化了速度，而Wonder Full是通过API提供的完整版本。

Wonder is the slimmed-down version optimized for speed, and Wonder Full is the full version available through the API.

今天，我们要讨论这些模型的表现如何，社区的初步评价和体验，这如何改变AI竞赛，以及这对前沿节奏意味着什么。

Today, we're going to talk about how the models perform, early takes and experiences from the community, how this changes the AI race, and what it says about pacing the frontier.

在基准测试中，Wonder的两个版本都接近排行榜的顶端。

On the benchmarks, both versions of Wonder show up near the top of the rankings.

Wonder Full在DeepSuite 1.1上得了79%，比GPT-6 Astra和Opus 5.5高出近5个百分点，而在Terminal Bench 4.0上，只比那些前沿竞争对手低2分。

Wonder Full scored 79% on DeepSuite 1.1, which is almost five points ahead of GPT-6 Astra and Opus 5.5, and on Terminal Bench 4.0, it was behind those frontier rivals by just two points.

对于代理任务，Wonder Full似乎略有优势。

For agentic tasks, Wonder Full seemed to have a slight lead.

它在Meta内部的个人超级智能指数上排名第一，比Astra和Opus高出大约四分。

It topped Meta's internal personal super intelligence index, beating Astra and Opus by around four points.

在Harvey Lab的法律工作、Val的金融代理以及Automation Bench上，它也达到了最先进的水平。

It was also state of the art across Harvey Lab for legal work, Val's Finance Agent, and Automation Bench.

在测试计算机使用的OSWorld上，Wonder比前沿竞争对手落后几分，但依然稳稳地处于竞争行列。

On OSWorld, which tests computer use, Wonder was a few points behind frontier rivals, but still decidedly in the mix.

Artificial Analysis证实了Meta的结果，给Wonder Full打了56分的智能指数，Wonder则是52分。

Artificial Analysis backed up Meta's results, giving it a 56 on the intelligence index for Wonder Full and 52 for Wonder.

这使得Wonder Full与Sonnet 5.5并列第二，比Opus 5.5低两分。

That makes Wonder Full tied for second against Sonnet 5.5 and two points behind Opus 5.5.

它排在Claude Fable 5.1和GPT-6 Astra之前。

It places it ahead of Claude Fable 5.1 and GPT-6 Astra.

更精简的Wonder则稍微落后，排在第七位，与GPT-6-1 Sol和Gemini-4 Argon并列，后者当然还没有发布。

The slimmer Wonder is a little further behind in seventh place, tied with GPT-6-1 Sol and Gemini-4 Argon, which is, of course, yet to be released.

Artificial Analysis强调，这是对Muse Spark 1.3的实质性提升，而且距离该模型发布仅六周时间。

Artificial Analysis highlighted that this is a, quote, "material gain over Muse Spark 1.3 and comes just six weeks after that model's release."

最大的进步是在代理型专业工作和编程任务上，包括在GDP Val AA上排名第二，在AA Briefcase上排名第三。

The big gains were on agentic professional work and coding tasks, including a second place ranking on GDP Val AA, and third place on AA Briefcase.

Muse Wonder的价格相当高，每个任务4.52美元，这几乎是Muse Spark 1.3的三倍。

Muse Wonder was pretty pricey at $4.52 per task, which represents almost a 3x increase from Muse Spark 1.3.

这比GPT-6 Astra或Grok 4.7每个任务贵大约一美元，但比Opus 5.5每个任务便宜大约1.5美元。

That puts it about a dollar per task more expensive than GPT-6 Astra or Grok 4.7, and about $1.50 cheaper per task compared to Opus 5.5.

如果单看这些基准测试，Meta多年来首次回到了前沿水平。

If then we went purely from the benchmarks, Meta is back at the frontier for the first time in years.

对一些人来说，这本身就令人震惊。

For some, that was jaw-dropping all on its own.

Oli Lehmann写道：“我们现在处在什么时间线，扎克伯格竟然在碾压Astra和Fable？”

Oli Lehmann writes, "What timeline are we in where Zuck is mogging Astra and Fable now?

无限的资金加上无限的渠道，现在还有前沿智能？

Infinite money plus infinite distribution and now frontier intelligence?

长期看好Meta。”

Long Meta."

Shopify的Toby Lutke写道：“这个模型非常非常厉害。

Shopify's Toby Lutke writes, "This model is very, very good.

我绞尽脑汁想找点它做不了的事，结果发现其他顶级模型都搞不定的，它都能搞定。

I was struggling to find anything to throw at it that it couldn't do, even when other top models failed.

Meta回来了。”

Meta is back."

但基准测试从来都不是全部，没过多久，大家就开始给这股兴奋劲儿泼冷水了。

And yet the benchmarks are never the full story, and it wasn't long before folks started poking some holes in the excited press.

Max Blade给Muse Wonder和Opus 5.5出了同样的题，让它们做一个AAA级的割草机模拟器，他写道：“很明显，Muse在3D任务上，比如电子游戏或者可打印的CAD设计，离前沿水平还差得远。”

Giving Muse Wonder and Opus 5.5 the same prompt to build AAA-level lawnmower simulator, Max Blade wrote, "It's very clear that Muse is not anywhere near close to frontier in 3D tasks like video games or even printable CAD designs."

他还预示了另一个我们后面会看到的抱怨，接着说：“更吓人的是，它烧光了我50美元的计划，还额外花了4.11美元的API等价token。

Presaging another complaint that we will see, he also continued, "Even scarier is it burned my $50 spending plan, plus $4.11 in API equivalent tokens.

我不得不升级到200美元的计划才跑完这一轮。”

I had to upgrade to the $200 plan just to finish this run."

我就不再举更多Muse Wonder在那类测试中表现不佳的例子了，但想让你看看Theo的这条评论，他写道：“已经看到很多人不看好Muse Wonder了，因为它在3D方面没有像Astra和Opus那样做强化学习。”

I will spare you the million more examples of Muse Wonder underperforming on that type of test, but point you to this comment from Theo, who writes, "Already seeing a lot of people discount Muse Wonder because it wasn't RL'd in 3D stuff like Astra and Opus.

3D游戏演示看起来很酷，但它们根本不能代表真实的代码工作。

3D game demos are cool, but they don't represent real code work at all."

我完全同意这个观点，再同意不过了。

I could not agree more strenuously with this.

实际上，我认为这对整个行业和整个世界都是极其负面的，因为3D游戏和渲染在社交媒体上看起来很好，前沿实验室就觉得他们都必须针对这种用例进行优化，因为他们知道X上的人会根据这种用例来评判他们。

In fact, I think that it is actually extremely net negative to the entire industry and world around it that because 3D games and renderings look good on social media, the frontier labs feel like they all have to optimize for that use case because they know that people on X are going to judge it based on that use case.

Meta的首席AI官亚历山大·王甚至补充说：“整个团队现在都觉得自己很蠢，但这是事实。”

Meta's chief AI officer, Alexander Wang, even added, "Whole team feels dumb now, but it's true.

我们并没有明确地针对3D演示游戏来训练我们的模型。

We did not explicitly train our model on 3D demo games.

显然，事后看来，我们可能应该做一些这方面的训练，但也没什么好后悔的。

Obviously, in retrospect, we probably should have done a bit of that, but no regrets.

请在你的实际编码工作中试试Muse Wonder吧。”

Please try Muse Wonder on your actual coding work."

不幸的是，不仅仅是那些3D任务，有些人在其他方面也发现了问题。

Unfortunately, it wasn't just those 3D tasks where some people found issues.

AI研究员梅胡·莫尔汉写道：“抱歉，但Muse Wonderful完全没让我印象深刻。

AI researcher Mehu Molhan writes, "Sorry, but not impressed with Muse Wonderful at all.

离Opus 5.5的性能差远了，我觉得甚至比不上GPT-6-1 Sol。

Nowhere near Opus 5.5 performance, I'd argue not even GPT-6-1 Sol.

我不确定别人用的是什么模型，但那些让它和Opus相提并论的基准数字看起来很奇怪。

I'm not sure what model others are using, but the benchmark numbers making it comparable to Opus seem weird.

不是价格问题，光是质量本身就很差。

Not pricing, just quality itself is bad.

这些都是真实的编程任务，用的是我自己的智能体。”

These are real programming tasks and my own agents."

Every的丹·希珀写道：“Muse Wonder在我们高级工程师基准测试上的表现令人失望。

Every's Dan Shipper wrote, "Disappointing show for Muse Wonder on our senior engineer benchmark.

它要求模型从头重写一个用Vibe编码的垃圾数据库。

It asks the model to rewrite a Vibe-coded slop database from scratch.

这是衡量它对现实工程有用程度的一个好指标。”

A good measure of its usefulness for real-world engineering.

Muse Wonder 只拿到52%，远远落后于GPT Sol 6.1的71%。

Muse Wonder gets 52%, far behind even GPT Sol 6.1 at 71%."

Riley Brown写道：“我用了Muse Wonder五个小时。

Riley Brown wrote, "Used Muse Wonder for five hours.

本来打算做个视频的。

Was going to make a video.

没必要了。

No need.

还是用回Codex和Claude Code吧。”

Back to Codex and Claude Code."

有人猜测，问题其实出在Muse Wonder和Muse Wonder Full之间差距太大。

Some speculated that the issue was actually the fairly significant gap between Muse Wonder and Muse Wonder Full.

Adam Holter写道：“Muse Wonder Full看起来确实不错，但Muse Wonder在我所有的测试里表现都不怎么样。”

Adam Holter wrote, "Muse Wonder Full looks really good, but Muse Wonder isn't doing very well on any of my tests.

所以现在大家用Muse Wonder都得到不好的结果，这反而连累了Muse Wonder Full，因为它们名字一样。

So now everybody is getting bad results from Muse Wonder, and it reflects badly on Muse Wonder Full because they have the same name.

说实话，我觉得要是分开发布可能会更好。

I honestly think it would've been better to release these at different times."

我得说，从这件事能看出团队有多上心，亚历山大·王回应说：“反馈很中肯。”

I will note, giving a sense of how engaged the team is, Alexander Wang responded to that saying, "Fair feedback.

我私信你，想了解更多情况。

DM-ing you to understand more."

我还看到不少其他用户也有类似的看法，他们用Wonder Full体验不错，但用精简版Wonder就感觉一般。

And I saw a number of other users suggesting something similar, where they had good experiences with Wonder Full but bad experiences with the slimmer Wonder.

另一个主要的抱怨是关于套餐额度消耗得太快。

The one other big complaint was around how fast the plans got consumed.

阿拉法特·舒沃写道：“我用Muse Wonder都没法好好干活。”

Arafat Shuvo wrote, "I can't even do work properly with Muse Wonder.

每周的限额一下子就烧完了。

It burns through the weekly limit so fast."

马特·西尔维写道：“我刚升级到Muse每月50美元的套餐，想试试Wonder Full。”

Matt Silve wrote, "I just upgraded to the Muse $50 a month plan to try Wonder Full.

即使使用限额一个小时前重置了，它在不到10分钟内就用掉了我五小时额度的13%。

Even after the usage limit reset an hour ago, it used 13% of my five-hour usage in less than 10 minutes."

亚历山大·王确实提到可能有个bug在错误计算人们的使用量，但不管原因是什么，这都没能改善整体的第一印象。

Now, Alexander Wang did report that there might have been a bug that was miscounting how much people were using of their plans, but still, whatever the cause, it didn't help the overall initial impressions.

然而，Muse Wonder，特别是Muse Wonder Full，确实也赢得了一些支持者。

And yet, Muse Wonder, or specifically Muse Wonder Full, did win some adherence as well.

开发者兼企业家Theo再次将Muse Wonder Full列入了他的S级，与Opus 5.5并列，其他所有模型都远低于这个级别。

Developer and entrepreneur Theo again put Muse Wonder Full in his S tier alongside Opus 5.5 and everything else being far below that.

他举了个例子写道：“我让Opus 5.5、GPT-6.1-Soul和Muse Wonder Full修复一个bug。”

Giving an example, he wrote, "Asked Opus 5.5, GPT-6.1-Soul, and Muse Wonder Full to fix a bug.

Opus找到了根本原因并提交了一个PR。

Opus found the root cause and filed a PR.

GPT-6.1-Soul问了个问题，被异步问题工具中断搞崩溃了，然后找到了bug，修复了它，写了300行测试，并用难以理解的术语解释了一遍。

GPT-6.1-Soul asked a question, crashed from the async question tool interrupt, found the bug, fixed it, wrote 300 lines of tests, and explained it in impossible to understand terms.

Muse Wonder Full发现别人提交的一个小PR，然后说你应该合并这个。

Muse Wonder Full found a small PR someone else had filed and said you should probably merge this instead.

我真的很喜欢这个模型。

I really like this model."

主持人查理·霍尔茨也很喜欢这个模型。

Conductors Charlie Holtz also liked the model.

他写道：“快速印象：一次就发现了我们代码库里的一个真实安全漏洞，其他代理都没发现。”

He wrote, "Quick impressions, found a real security vulnerability in our code base in one shot that other agents missed.

它在设计和可视化方面很出色。

It's great at design and visualizations.

它非常喜欢用子代理，而且在命令行界面里有很多很棒的细节。

It really likes sub-agents and lots of great details in the CLI.

这里的进展让我印象非常深刻。”

Super impressed by the progress here."

BoneGPT写道：“我提前体验了Muse Wonder，它是个让人惊喜的模型。”

BoneGPT wrote, "I had early access to Muse Wonder, and it is a surprising model.

在我们跑的每一个技术基准测试里，它都和Astra、Fable、Opus不相上下，甚至更好。

On every technical benchmark we ran, it was neck and neck or better than Astra, Fable, and Opus.

它处理得很细腻。

It has a delicate touch.

它做创意决策时很有灵感。

It makes inspired creative decisions.

其实没有衡量灵魂感的基准，但Muse有。

There aren't really benchmarks for soulfulness, but Muse has it.

它是一个温暖的模型，一个能干的模型，而且有大模型的那种感觉。

It's a warm model, a capable model, and it has big model smell.

它的输出总是快乐、天真又友好。

Its outputs are happy, innocent, and friendly.

它不会像其他一些模型那样居高临下地跟你说话。

It never talks down to you like some of the other models.

它想帮你做出很酷的东西。

It wants to help you make dope stuff.

它希望你赚钱。

It wants you to make money.

它跟你是一条心的。

It's aligned with you.

Meta 将会以压倒性优势赢得消费者代理市场。

Meta is going to win consumer agents in a landslide."

克莱尔·沃按类别拆解了她的测试，发现它在前端设计方面表现相当糟糕。

Claire Vo broke her tests down by category, finding that on front-end design, it was quite bad.

除了写邮件，其他写作任务它都不太行。

Writing anything other than emails, it was bad.

它在 SVG、视频和创意设计方面表现不佳，但正如她所说，在编码和代理任务上，它跟大模型们并驾齐驱，甚至碾压对手，在她测试的八个前沿模型中排名第一。

That it wasn't great for SVGs, videos, and creative design, but that, as she put it, it rolls with the big boys on coding and on agentic tasks, it crushed, coming in first out of eight frontier models she tested.

现在，一个有趣的问题是，这将在 Meta 的战略中扮演什么角色。

Now, one interesting question is where this is going to fit in Meta's strategy.

当然，Meta 比其他实验室更注重消费者体验，因为他们拥有消费者应用的分发渠道和网络。

Meta is, of course, much more focused on the consumer experience than other labs are, given their distribution and network of consumer apps.

大约一个月前，我注意到 Muse 的一个独特之处在于，它是第一个大多数用户根本不关心背后是什么模型在驱动的 AI 产品。

About a month ago, I had observed that one of the things that was distinct about Muse was that it was the first AI product where most of its users just didn't care what model was powering it.

然而，尽管我相信这是真的，但确实有不少用户开始遇到Muse能力上的某些限制。

And yet, even though I believe that that is true, there were a fair number of users that started to run up against some limits in what Muse could do.

我在X上发布那篇文章一周后，Sergito写道：“有没有人也觉得Muse不如Instinct聪明？

A week after I posted that on X, Sergito wrote, "Anyone else finding Muse to be less smart than Instinct?

我给它们俩布置了类似的多步骤任务，Muse不仅产生了一些幻觉，而且在尝试修正时还不断加重错误。”

I gave them both a similar multi-step task, and Muse hallucinated some stuff and then kept compounding its error when trying to fix it."

事实上，我们现在的处境有点像之前讨论过的，重度用户会为不同任务选择不同模型，而很多早期采用者也在同时使用多个不同的个人助理。

And in fact, we've kind of been in a spot where in the same way that I discussed power users using different models for different tasks, a lot of early adopters are using multiple different personal assistants.

Brycent最近在X上发帖说：“我意识到，没有哪个个人助理是万能的。

Brycent recently posted on X, "What I've realized is that there is no one optimal personal assistant.

目前，我正在使用Instinct、Muse、Grok Bot和Dots。

At the moment, I'm currently using Instinct, Muse, Grok Bot, and Dots.

我确实喜欢它们各自在特定场景下的表现。”

I genuinely enjoy using them all for specific use cases."

这也许就是未来的趋势，但我觉得对很多人来说，在多个个人助理之间切换的成本，会远高于在其他场景中切换模型的成本。

Now, that may be how it plays out, but I think for a lot of folks, the context switching cost of using multiple personal assistants is going to be significantly higher than the context switching cost for using models in other types of ways.

换句话说，对于LLM和编程代理来说，有很多使用场景，即使它们对你一无所知，也能提供巨大的价值。

In other words, there are many use cases for LLMs and coding agents in general, where they can provide a ton of value, even if they don't know anything about you.

而个人助理则不同，它们越了解你，越能访问你的系统，就越有用。但大多数人连给一个个人助理提供所有这些信息都很难，更不用说给六七个了。

Personal assistants, on the other hand, get more useful in proportion to what they know about you and the systems of yours they have access to, and most people are going to have a hard time giving even one personal assistant access to all that information, much less a half dozen.

考虑到这一点，像Muse Wonder或Muse Wonder Full这样的需求在最近的社交媒体上随处可见。

And with that in mind, the need for something like Muse Wonder or Muse Wonder Full is written all over the recent pages of social media.

投资人和企业家Satya Mallick最近写道，虽然Muse在处理即时任务（比如订机票和酒店）上很擅长，但用他的话说，它在长期任务上表现得很糟糕。

Investor and entrepreneur Satya Mallick recently wrote that while Muse is very good at immediate tasks like booking flights and hotels, it is, in his words, terrible for long-term tasks.

比如，他说：“它今天提醒我，下午2点的会议还没安排好。”

For example, he said, "It reminded me today that a meeting has not been set for 2:00 PM.

那是个误报，因为会议是昨天下午2点安排的。

That was a false alarm because the meeting was set for 2:00 PM yesterday.

这种情况发生过好几次，它搞错了日期，就慌慌张张地告诉我一些事情。”

This has happened several times where it panics and tells me something because it gets the day wrong."

投资人Sam Lessin写道：“昨晚我试着用Muse订酒店。”

Investor Sam Lessin wrote, "Tried to book a hotel with Muse last night.

希尔顿的网站让我的智能体陷入了死循环。

Hilton's site put my agent in a doom loop.

重试了三次后，我放弃了，自己订了。

Third retry, I gave up and booked it myself."

很明显，Meta团队在开发Wonder时，心里想的就是这类挑战。

And it's quite clear that this type of challenge was on the mind of the Meta team in building Wonder.

X上的Chubby写道：“一个有趣的见解，可能被发布新闻淹没了。”

Chubby on X wrote, "An interesting insight that was probably drowned in the release news.

Meta构建了Wonder，用于代表你行动的智能体，首先应用于其Muse应用。

Meta built Wonder for agents that act on your behalf, starting with its Muse app.

有两个功能非常重要。

Two features are really important.

主动记忆，能跨对话跟踪对你重要的事情；主动专注，能在任务遇到困难时继续推进，尝试其他方法，并从错误中恢复。”

Active memory, which keeps track of what matters to you across conversations, and active focus, which keeps a task going when it gets hard, tries other approaches, and recovers from errors."

不过，我看了发布媒体后，还有一个疑问：通过Wonder模型系列，Meta是否真的开始认真对待不仅个人智能体用例，还有专业用例呢？

Another question I had, though, looking at the launch media, is whether with the Wonder family of models, Meta is actually getting serious about going after not just personal agentic use cases, but professional ones as well.

当然，马克·扎克伯格称其为个人和专业代理工作的最先进性能。

Certainly, Mark Zuckerberg called it state-of-the-art performance for personal and professional agentic work.

亚历山大·王在介绍时写道：“在个人代理和专业代理方面都是最先进的。”

Alexander Wang, in introducing it, wrote, "State of the art at personal agents and professional agents."

而且Meta的AI账号也明确提到了个人代理和专业代理，这意味着他们进行了一场协调一致的宣传，确保提到专业代理。

And the AI at Meta account also explicitly mentioned both personal and professional agents, meaning that there was a coordinated messaging campaign to make sure to mention professional agents.

这可能意味着Meta这次真的重视专业使用场景，或者也可能意味着他们明白，个人代理中隐含了专业使用场景，而从公司战略角度来看，划清界限可能合理，但这并不符合消费者对自己的认知。

Now, that could mean that Meta is serious about professional use cases this time, or it could mean that they understand that implicit in a personal agent are going to be professional use cases, and that drawing a clear line might make sense from a company strategy standpoint, but doesn't match how consumers think about themselves.

换句话说，可能并不是Meta突然想和Claude Code在企业交易上竞争，而是他们认为，在个人层面上，如果消费者真的选择并投资于一个个人代理，那么这个个人代理在很多情况下就必须处理专业事务。

In other words, it may not be that all of a sudden Meta wants to compete with Claude Code for enterprise deals, but that they believe that on an individual level, if consumers really choose and invest in a single personal agent, that personal agent is going to have to do professional stuff in many cases.

当然，还有另一类专业人士对Meta来说已经非常重要，那就是小企业主，他们已经在Facebook和Instagram等其他应用上使用了大量功能。

There is, of course, another category of professional that is already deeply, clearly important to Meta, which is small business owners who already use tons and tons of features on their other apps like Facebook and Instagram.

这正是Satrini Research看到可行性的地方，他们写道：“我们的结论是，Meta在吸引小企业代理方面有真正的机会。”

That's where Satrini Research saw the viability, writing, "Our takeaway is that Meta has a real shot at capturing agents for small business.

这是能力和执行力的很好平衡。

It's a really good balance of capability and execution."

对很多人来说，重点不在于Wonder能做什么的最终结果，而在于变化的速度和幅度。

For a lot of folks, the big story is not the endpoint in anything Wonder can do, but the delta and the rate of change.

Pulsia的Ben Sarah写道：“Meta正在以闪电般的速度前进。

Pulsia's Ben Sarah wrote, "Meta is moving at lightning speed.

Muse发布才34天，现在他们已经跻身前沿了。”

Muse launched 34 days ago, now they've joined the frontier."

Ethan Mollick说，从有限的测试来看，这似乎是个不错的模型，但引用他的话说：“也许更令人印象深刻的是，对比一下进步的速度。

Ethan Mollick said that it seems to him like a good model from a limited test, but, quote, "Perhaps more impressive, compare the progress.

Meta在短时间内取得了非常快的进展。”

Very fast progress from Meta in a short time."

Yuchen Jin写道：“一年前，很多人嘲笑Meta花数十亿美元聘请顶尖AI人才并收购Alexander Wang。

Yuchen Jin wrote, "A year ago, many dunked on Meta for spending billions to hire top AI talent and acquire Alexander Wang.

说实话，我当时也持怀疑态度。

I was skeptical too, to be honest.

但Muse应用和模型的成功证明了一件事：聘请顶尖人才，组建一支精英团队，给他们一个雄心勃勃的目标和足够的资源，他们就能创造奇迹。”

But the success of the Muse app and models proves one thing: hire top talent, assemble a crack team, give them an ambitious goal and enough resources, and they can work wonders."

当然，市场对看到的东西很满意，消息一出，Meta股价上涨了4%。

Certainly, markets liked what they saw, with Meta up 4% on the news.

但这对整个“引领前沿”的讨论意味着什么呢？

But what does this mean for the whole pacing-the-frontier conversation?

在AI首席执行官中，扎克伯格与其他人的信息传递并不特别一致。

Among the AI CEOs, Zuckerberg has not been particularly aligned with the messaging coming from the others.

他没有签署《引领前沿》公开信，而且在这封信发表几周后，他通过自己的文章《未来属于每个人》发出了相反的信号。

He did not sign the Pacing the Frontier letter, and a few weeks after it was published, he counter-signaled with his essay, "The Future is for Everyone."

他的核心论点是，开放和广泛获取必须成为AI时代的核心原则，即使这会带来一些额外的风险。

His core thesis is that open and widespread access needs to be the core tenet for the AI era, even if that carries some extra risk.

在扎克伯格看来，将未发布的模型仅限内部使用所导致的权力集中，可能比发布强大AI模型所带来的风险更具灾难性。

In Zuckerberg's view, the concentration of power that comes from keeping unreleased models for internal use only could be even more disastrous than the risks associated with releasing powerful AI models.

现在，显然，这一系列Wonder模型并没有推动前沿，而且似乎有非常强大的网络防护措施，这意味着与现有模型相比，它们并不危险。

Now, clearly, this set of Wonder models isn't pushing the frontier and do seem to have very strong cyber guardrails, meaning that they are not dangerous compared to what's already available.

但从发展轨迹来看，扎克伯格对放慢脚步毫无兴趣。

But from a trajectories perspective, Zuckerberg is showing no interest in slowing down.

在接受《金融时报》采访时，他宣称Meta已经达到了前沿水平，并且，引用他的话说，“已经在着手下一步的提升了。”

In an interview with the "Financial Times," he declared that Meta has reached the frontier and is, quote, "already working on that next step up."

他并没有打算放慢脚步，反而补充说，模型开发“从现在开始可能会感觉快很多。”

Rather than slowing down, he added that model development will, quote, "probably feel like it'll go quite a bit faster from here."

他拒绝透露自己的“P doom”数值，也就是AI杀死所有人的概率，这个概率在过去一个月里让全世界都炸开了锅。

He declined to give a number for his P doom, the odds of AI killing everyone that got the entire world riled up over the past month.

相反，他说：“我更倾向于相信，AI是一种可以通过正常防护措施来工程化、并且负责任地构建的技术，而不是那种拟人化的AI愿景，即AI会自己去做自己的事情。”

Instead, he said, "I'm much closer to the belief that AI is a technology that can be engineered with normal guardrails and built responsibly than the anthropomorphic vision of AI, which is that it's just going to go do its own thing."

不管你是否同意这种可能性，Meta对他们正在构建的东西的看法，与其他实验室确实存在根本性的差异。

Whether you agree with whether it's possible or not, there does seem to be a fundamental difference in how Meta views what they're building from the other labs.

简而言之，OpenAI和Anthropic要么是在试图构建，要么是觉得他们不可避免地会构建出令人敬畏的机器神，而Meta则是在试图构建他们认为只是一种普通技术的东西。

In short, OpenAI and Anthropic are either trying to or feel like they inevitably are building machine gods that inspire awe, while Meta is trying to build what it views as just a normal technology.

谁对谁错还有待争论，但这就是根本的分界线。

Who's right or wrong about that remains the debate, but that is the fundamental dividing line.

总的来说，退一步看，我们进入2026年时，有三个前沿模型实验室，分别是OpenAI、Anthropic和谷歌。

Overall, when you take a step back, we headed into 2026 with three frontier model labs in OpenAI and Anthropic and Google.

随着编程智能体冲到最前面，显示出它们比其他一切都重要，至少在短期内是这样，那个数字降到了两个。

As coding agents forced their way to the top and showed that they were more important than everything else, at least in the short term, that number went down to two.

但现在，尤其是如果Gemini 4 Argon能做到谷歌所说的那样，我们又回到了美国有大约五个前沿实验室在竞争的局面。

And yet now, especially if Gemini 4 Argon can do what Google says it can do, we are back to having five-ish frontier labs competing in the US.

不过话说回来，在幕后，正如我们今天在头条新闻中听到的那样，可能还有更强大的模型即将出现，所以也许那个前沿双头垄断很快就会重新出现。

Then again, behind closed doors, as we heard in the headlines today, there might be much more powerful models just on the horizon, so maybe that frontier duopoly will be right back soon.

目前看来，Meta似乎完全回到了游戏中。

For now, it appears that Meta is fully back in the game.

今天的AI每日简报就到这里。

That's going to do it for today's AI Daily Brief.

感谢你的收听或观看。

Appreciate you listening or watching.

一如既往，下次再见，祝一切安好。

As always, and until next time, peace.