AI Model Month Is Off to a Blistering Start — 文稿

F10:00

整个夏天，我们在AI每日简报里一直在探讨的一个大主题，就是从单一模型范式——也就是你挑一个最好的模型，然后就一直用那个——转向更复杂的模型架构。在这种架构下，无论是个人还是团队，都能灵活地在不同模型甚至不同工具之间切换，从而根据具体的应用场景，最大限度地发挥AI的潜力。

Throughout the summer, the big theme we've been exploring at the AI Daily Brief is all about the move from a single model paradigm where you pick the best model overall, and that's the one you stick with, to a more complex model architecture where we are both as individuals and as teams able to navigate nimbly between different models and even different harnesses to get the most out of AI based on whatever particular use case we might have.

而且，这个夏天我们非常清楚了一点：要最大化AI的效用，不光要看模型或工具的能力，还得考虑效率和成本，尤其是在我们转向更复杂的智能体工作负载时。

And what's more, this summer we got really clear on the fact that getting the most out of AI is not just a question of model or harness capability, but also a question of efficiency and cost, especially as we move to more complex agentic workloads.

所以，九月初涌现了一大波新模型，从Fable 5.1到GPT-6 Astra，再到我们今天要看的Muse Spark 1.3和ChatGPT Images 2.5，真是应景。

And so it's fitting that the beginning of September has been just a cavalcade of new models from Fable 5.1 to GPT-6 Astra to the models that we're looking at today, including Muse Spark 1.3 and ChatGPT Images 2.5.

所有这些加起来，让我们可用的工具更加多样化了，你可以设计出最适合自己实际生活和工作的AI组合。

All of these add up to way more diversity in the tools we have access to for you to design the perfect AI stack for your actual life and work.

AI每日简报是一档日播播客和视频节目，专注AI领域最重要的新闻和讨论。

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

好了朋友们，在正式开始之前，先快速说几个通知。

Alright friends, quick announcements before we dive in.

F11:56

今天我们先从一个故事说起，这事儿本来完全可以做成一期主节目，因为围绕它有好多的戏剧性事件。

We kick off today with a story that very easily could have been the main episode, given how much drama is surrounding it.

周二，OpenAI发布了一个纳维-斯托克斯问题的解法，这是2000年选出的七个千年大奖问题之一。

On Tuesday, OpenAI published a solution to the Navier-Stokes problem, one of the 7 problems selected for the Millennium Prize in the year 2000.

华尔街日报把这些难题称为数学界的圣杯，这个说法相当准确。

The Wall Street Journal characterized these problems as the holy grail of math, and that's fairly accurate.

每个千年大奖问题都附带一百万美元的奖金，而在该奖项设立的26年里，只有一个问题被解决。

Each Millennium Prize problem has a million-dollar reward attached, and only one has been solved in the 26 years since the prize was established.

其他问题包括数学中最著名的未解难题，比如黎曼猜想和P/NP问题。

The other problems include the most famous unsolved problems in math, such as the Riemann hypothesis and P versus NP.

你知道的，就是大家聚餐时都会聊的那些话题。

You know, the things we all talk about when we get together for dinner.

不过，就这期节目而言，我其实不打算深入探讨问题本身，或者争论它有没有什么重要的实际应用。

Now, for the purposes of this particular episode, I'm actually not going to get into the details of the problem itself, or debates around whether it has any significant real-world applications.

我来读一下OpenAI对问题的描述，给你们一点感觉。

I'll read OpenAI's description of the problem just to give you a flavor.

他们写道：纳维-斯托克斯方程利用牛顿第二运动定律F=ma来描述流体的运动。

They write, The Navier-Stokes equations use Newton's second law of motion, F ma, to describe how fluids move.

重要的是，它们把流体当作连续介质来处理，而不是追踪单个分子。

Importantly, they treat a fluid as a continuous medium rather than tracking individual molecules.

这些方程用于飞机设计、天气预报和血液流动的研究。

These equations are used for aircraft design, weather forecasting, and the study of blood flow.

这些动力学方程的一个基本未解问题是：流体的连续近似是否可能失效。

A fundamental open question for these dynamical equations has been whether the continuum approximation of the fluid can break down.

具体来说，对于密度恒定的三维不可压缩流体，即使运动一开始很平滑，纳维-斯托克斯方程是否会产生奇点。

Specifically, can the Navier-Stokes equations for a 3-dimensional incompressible fluid with constant density develop a singularity even when the motion starts smoothly.

这里，奇点意味着动力学导致流体中的速度在有限时间内变得无限大。

Here, a singularity means the dynamics lead to speeds in the fluid growing without bound within a finite amount of time.

尽管存在趋于平滑运动的黏性，但奇点的出现仍会加深。

The development of a singularity would have to deepen despite the presence of viscosity, which tends to smooth out motion.

因为真实流体不可能无限快地运动，所以这标志着方程在模拟流体时的一个失效点。

Because a real fluid cannot move infinitely fast, this would mark a breakdown in how the equations model the fluid.

要继续模拟这个系统，就需要追踪每个粒子的行为。

To continue modeling the system, one would then need to track the behavior of each particle individually.

所以这就是他们正在解决的问题，我觉得再提供一下背景很重要：这比去年引起轰动的Erdős问题又进了一大步。

So that's the problem they're addressing here, and I think again for context, The important note is that this represents a huge step up from something like the Erdős problems that made news last year.

OpenAI声称已经解决了这个问题，并且他们使用的是内部模型，该模型的能力明显强于GPT-6 Astra。

OpenAI claims to have solved this problem, and they did so using an internal model that is significantly more capable than GPT-6 Astra.

Noam Brown说，得出这个结果花费了几百万美元，而且似乎花了一两周的时间。

Noam Brown said that the result cost several million dollars to find, and it seems to have taken a week or two.

不过，争议的焦点在于OpenAI究竟是如何得出这个结果的。

However, the big controversy surrounded exactly how OpenAI had arrived at this result.

结果公布后不久，纽约大学教授Tristan Buckmaster发表了他对这件事的描述。

Shortly after the result was published, New York University professor Tristan Buckmaster published his version of the events.

据Buckmaster说，他和一位Anthropic的员工Levent Dalpaci已经一起研究纳维-斯托克斯问题一年多了。

According to Buckmaster, he and an Anthropic employee named Levent Dalpaci had been working on the Navier-Stokes problem together for more than a year.

这是Levent的业余项目，他们两人使用了多种不同的AI模型，包括Codex工具中的GPT-5.6 SOL。

This was an outside project for Levent, and the pair had used a range of different AI models, including GPT-5.6 SOL in the Codex harness.

关键的是，Levent 和 Buckmaster 并没有解决纳维-斯托克斯方程，但他们确实在一些相关问题上找到了新颖的解法，这些解法可以被看作是迈向千禧年大奖问题的一个垫脚石。

Crucially, Levent and Buckmaster did not find a solution to Navier-Stokes, but they did find novel solutions to related problems that could be viewed as a stepping stone to the Millennium Prize problem.

而且他们使用了极其新颖的方法论，这在数学界很少有人采用。

They were also using extremely novel methodology that few in the mathematics world were pursuing.

最近几周，他们工作的传闻在人工智能圈子里流传，有人错误地声称 Anthropic 已经解决了一个千禧年大奖问题。

Rumors of their work spread through AI circles in recent weeks, incorrectly claiming that Anthropic had solved a Millennium Prize problem.

Buckmaster 说他上周联系了 OpenAI，想澄清一下情况。

Buckmaster says he reached out to OpenAI last week to clarify the situation.

据 Buckmaster 描述，OpenAI 的 Sebastian Bubeck 在周日告诉他，他们已经解决了纳维-斯托克斯方程，并且想讨论发表事宜。

According to Buckmaster's telling, Sebastian Bubeck from OpenAI informed him on Sunday that they had solved Navier-Stokes and wanted to discuss publication.

我问这个模型是否训练过或者能够访问我们在 Codex 中的会话记录——我们整个项目期间的所有草稿都放在里面。

I asked whether the model had been trained on or had access to our sessions in Codex, into which we had been putting all of our drafts for the whole of the project.

他们告诉我，模型没有查阅用户数据。

I was told the model did not look up user data.

我又问了一下训练数据的问题，但没有得到答复。

I asked again about training and I did not get an answer.

他说他收到了两个方案：要么 OpenAI 单独发表，要么 Buckmaster 加入发表，但前提是他同意把 Levent 从作者名单中移除，因为 Levent 与 Anthropic 有关系。

He said he was offered 2 proposals: for either OpenAI to publish separately, or Buckmaster to join the publication if he agreed to remove Levent from the authorship because of his ties to Anthropic.

Buckmaster 两个都拒绝了，并威胁说如果 OpenAI 发表，他就公开这件事。

Buckmaster declined both options and threatened to go public if OpenAI published.

Buckmaster 继续讲述道，对方回复说，你为什么要毁了自己的职业生涯？

Continuing his account, Buckmaster wrote, the reply was, why would you ruin your career?

我回复说我是一名学者，并且问他为什么认为公开这件事会毁了我的职业生涯。

I replied that I am an academic and asked why he thought going public would ruin my career.

对方回复说，如果你不想让我客气，那我就不必客气了。

The reply was, if you don't want me to be nice, then I don't have to be nice.

OpenAI 的领导们随后发表了一系列声明。

OpenAI leaders responded with a series of statements.

Sebastian Bubeck 称这些指控是虚假且煽动性的，然后公布了他的一部分短信记录，据称这与 Buckmaster 的说法相矛盾。

Sebastian Bubeck called the allegations false and inflammatory, then revealed part of his text message chain, which he claims contradicts Buckmaster's version of events.

Sam Altman 对事情如何变得这么糟糕做了一系列解释，但声称 OpenAI 的做法与 Buckmaster 和 Levent 的不同。

Sam Altman gave a series of explanations for how this went sideways, but claimed that OpenAI's approach was different than that of Buckmaster and Levent.

OpenAI 方面的说法是，我们、研究人员以及智能体，在他们公开发布之前，没有通过任何方式看到他们的任何工作。

The OpenAI account claimed, we, the researchers and the agents, did not see any of their work through any means until they released it publicly.

具体来说，没有为了解决这个问题而访问任何特定的用户数据。

In particular, no specific user data was accessed in order to solve this problem.

虽然不太可能，但我们不能排除，从我们产品使用中获得的去标识化数据帮助改进了我们的模型。

While unlikely, we cannot rule out that de-identified data derived from the usage of our products helped improve our models.

当 Levent 将这份声明形容为“某种程度上是在坦白”之后，OpenAI 的首席研究官 Mark Chen 回应说，有两件事需要区分清楚。

After Levent characterized this statement as quote-unquote coming clean, OpenAI Chief Research Officer Mark Chen responded, 2 things to distinguish.

是否有任何人或智能体在纳维-斯托克斯问题的研究中查看了用户数据？

Did any human or agent look at user data as part of the Navier-Stokes effort?

没有。

No.

我们是否使用用户反馈和去标识化数据来全面改进 ChatGPT 和 Codex？

Do we use user feedback and de-identified data to improve ChatGPT and Codex in a holistic way?

是的，每家大型语言模型公司都这么干。

Yes, and so does every LLM company.

现在，关于这场争议，有两种不同的讨论方向。

Now, as for the controversy, there are 2 distinct strains of conversation.

首先，学术界人士对于他们认为的不道德行为感到非常愤怒。

Firstly, academics are up in arms over what they see as unethical behavior.

伊利诺伊大学的助理教授 Talia Ringer 写道，听说别人有了成果就赶紧抢着发表，这太离谱了。

Assistant Professor Talia Ringer of Illinois University wrote, Rushing to get a result after you hear someone else has a result is messed up.

这就是人工智能界的“抢先文化”，完全违背了数学等合理领域的所有学术规范。

That is AI scooping culture and goes against every academic norm that exists in reasonable fields like mathematics.

这种人工智能文化就是这么腐蚀整个领域的。

This is how AI culture rots entire fields.

Hugging Face 的联合创始人 Thomas Wolfe 认为，这可能只是加速版人工智能科学的一个预演，他评论道，希望这不是科学研究未来的样子——这些占主导地位的玩家玩着营销游戏，伤害了真正的科学界。

Thomas Wolfe, the co-founder of Hugging Face, suggested this might just be a preview of accelerated AI science, commenting, hope this is not a glimpse of the future we'll get in science research, with these dominating players playing marketing games hurtful for the real scientific community.

第二个批评，至少对于AI Daily Brief的听众来说可能更相关，是对OpenAI的信任问题。

The second, and likely far more relevant criticism at least for the AI Daily Brief audience, was questions of trust in OpenAI.

根据Buckmaster的说法，我们可以推测他们用的是某个消费者版本的Codex，但不确定他们是否同意共享数据来改进OpenAI的模型。

From Buckmaster's account, we can assume they were using some consumer version of Codex, but it's unclear whether they agreed to share data to improve OpenAI's models.

对一些人来说，这是对所有用AI处理专有任务的人敲响的警钟。

For some, it's a wake-up call for anyone who is using AI to work on proprietary tasks.

前DeepMind员工Susan Zhang写道：大家都被个人drama吸引了，却忽略了一个更有趣的未解问题——这些实验室能看到你所有的工作，并在利益足够大时抢先一步吗？

Former DeepMind employee Susan Zhang wrote, Everyone getting sniped by the personal drama but missed the more interesting unanswered question: can these labs see all your work and scoop you when the stakes are high enough?

为了向OpenAI的领导层寻求澄清，数学家Tryon Zylorus问道：重要的问题，如果我选择退出训练，然后通过我的付费订阅粘贴一个商业机密，你们会去识别我的个人详细信息但保留商业机密，并可能把它添加到你们的训练数据中吗？

Seeking clarification from OpenAI leaders, mathematician Tryon Zylorus asked, important question, if I opt out from training, then paste a trade secret using my paid subscription, do you de-identify my personal details but keep the trade secret and may add it to your training data?

截止录制时，这个问题还没有得到回复。

At the time of recording, that has not received a response.

所以退一步看，这次事件引起如此共鸣有几个原因。

So taking a step back, there are a few reasons that this whole episode is having such resonance.

首先，老实说，是其中的窥视癖。

First is honestly the voyeurism of it.

人们喜欢看戏。

People love drama.

与之抗争就像试图对抗潮汐。

Fighting against that is like trying to fight the tides.

但围绕先进AI的伦理问题以及这些公司能拿数据做什么，这个问题自LLM诞生之初就存在，最近却愈发受到关注，尤其是当模型领导地位开始集中在少数几家公司手中时。

But the question of ethics around advanced AI and what these companies can do with data is a question that, while it has been present basically since the beginning of LLMs, has gotten a lot louder in consideration more recently, especially as model leadership starts to consolidate around a couple of companies.

它给人们带来了许多令人不适的问题。

It brings up a lot of uncomfortable questions for people.

其中一些问题涉及经济激励以及AI公司最终会走向何方。

Now, some of those questions go to economic incentives and where the AI companies ultimately land.

现在人们热议的一件事是，这些公司最终是否不仅把自己看作创新投入的销售者，还是创新产出的销售者。

One of the things that there is a lot more chatter about right now is questions of whether these companies will ultimately not view themselves just as selling the inputs to innovation, But also is selling the outputs of innovation.

换句话说，OpenAI和Anthropic是把进行新型药物发现的能力卖给现有的科学家、实验室和公司更合理吗？

In other words, does it make more sense for OpenAI and Anthropic to sell existing scientists and labs and companies the ability to do novel drug discovery?

还是自己去做药物发现，然后从专利的另一端赚钱更合理？

Or does it make more sense to do that drug discovery yourself and get the money on the other side of the patents?

考虑到我们之前热议的Dario是否说过Anthropic将在某个时间点成为世界上唯一的公司，这些问题现在似乎比过去更贴切了。

Given all the chatter that we had around whether Dario had actually said that Anthropic was going to be the only company in the world at some point, those questions feel a little bit more pertinent now than they might have in the past.

最后，这个故事揭示出OpenAI内部已经在使用一个更强大的模型，这当然也引起了一些关注。

And finally, the fact that the story reveals that OpenAI has a much more powerful model that they're already using internally certainly captured some notice as well.

不幸的是，和通常情况一样，这件事中没有谁看起来特别光彩。

Unfortunately, as is so often the case, no one looks particularly good coming out of this.

《纽约时报》的Mike Isaac写道：我并没有忽略一个事实——两个要求公众信任它们是负责任的AI领导者的实验室，在Twitter上因为一道数学题的功劳而互相撕扯。

The New York Times' Mike Isaac wrote, not lost on me that the 2 labs asking the public to trust them as stewards of responsible AI leadership at existential stakes are having a slap fight on Twitter about who gets credit over a math problem.

就像我说的，这本来可以单独做一整集。

Like I said, this could have been a whole main episode.

但由于时间有限，我们快速过几个其他故事，然后再进入今天的正题——那些我们还没机会聊的各种新模型。

But since we're a little compressed on time, let's quickly rip through a couple of other stories before we get to our main, which is all about all sorts of new models that we haven't had a chance to talk about yet.

首先，说到Anthropic，有一项新的集体诉讼代表Claude Max订阅用户提起，指控Anthropic使用欺骗性营销和模糊的细则来少提供服务。

First up, speaking about Anthropic, there is a new class action lawsuit against the company filed on behalf of Claude Max subscribers with the claim that Anthropic used deceptive marketing and opaque fine print to underserve customers.

具体来说，诉讼声称每月100美元的5倍计划和每月200美元的20倍计划，实际上并没有提供每月20美元Pro计划的5倍和20倍使用量。

In particular, the lawsuit claims that the $100 a month 5x plan and the $200 a month 20x plan don't actually deliver 5 and 20 times the usage of a $20 a month Pro plan.

原告指控，5小时和周使用限制的计算方式意味着实际使用量远低于广告宣传的倍数。

Plaintiffs allege the way 5-hour and weekly usage limits are calculated mean the actual usage is far lower than the advertised multiples.

通常这类案件我不会太感兴趣，但我认为它代表了随着Anthropic和OpenAI越来越融入日常商业运作，我们会看到更多类似的事情。

Now, usually this type of case wouldn't be all that interesting to me, but I think it's sort of representative of the type of thing that we're gonna see a lot more of as Anthropic and OpenAI become increasingly interwoven with just the normal way of doing business.

负责此案的律师们自己指出，让他们觉得这个案子值得关注的原因是顶级AI订阅已经变得如此普遍和必要。

The lawyers running the case themselves noted that what made it compelling to them was how ubiquitous and necessary a top-tier AI subscription has become.

他们说经常听到员工们反馈，觉得需要支付高价订阅才能在就业市场上保持竞争力，但感觉并没有得到自己付钱应得的东西。

They said that they were frequently hearing from workers who felt they needed to pay high-priced subscription costs to remain relevant in the job market, but felt they weren't getting what they were paid for.

我不太确定这个诉讼会有什么结果，但它确实代表了顶级AI实验室未来将面临的审查级别。

I'm not particularly sure I think this goes anywhere, but it is certainly representative of the level of scrutiny that the top AI labs are going to face going forward.

在市场上，看起来不只是OpenAI和Anthropic在考虑IPO。

Over in markets, it looks like it's not just OpenAI and Anthropic that are thinking about IPO.

ElevenLabs 还聘请了一位首席财务官，帮助这家初创公司走向上市。

ElevenLabs has also hired a chief financial officer to help the startup head for a public listing.

周二，ElevenLabs 宣布 Ethan Tandowski 加入了高管团队，他此前担任荷兰金融科技公司 Adyen 的 CFO，这家公司于 2018 年上市。

On Tuesday, ElevenLabs announced that Ethan Tandowski had joined the executive team, having most recently served as the CFO of Adyen, a Dutch fintech firm that went public in 2018.

ElevenLabs 联合创始人 Mati Staniszewski 在新闻稿中表示，Ethan 带来了在高速增长环境中拓展财务运营的良好记录，以及作为上市公司 CFO 的宝贵经验。

ElevenLabs co-founder Mati Staniszewski said in a press release, Ethan brings a strong track record of scaling financial operations in high-growth environments and valuable experience as CFO of a public company.

这似乎也是 ElevenLabs 的雄心所在。

And that appears to be ElevenLabs' ambition as well.

The Information 报道称，他们开始探索可能的 IPO。

The Information reports that they are beginning to explore a possible IPO.

该公司表示，他们有望在今年年底达到 6 亿美元的年化收入，而去年底是 3.5 亿美元；消息人士称，该公司已实现盈利，目前超过一半的收入来自大型企业客户。

The company said that they are on track to reach $600 million in annualized revenue by the end of the year, up from $350 million at the end of last year, and sources said that the company has reached profitability and is now generating more than half of their revenue from large enterprise customers.

Cognition 最近传闻的融资轮已经完成。

Cognition's recently rumored fundraising round has completed.

该公司筹集了 20 亿美元的新资金，使其估值飙升至 480 亿美元。

The company raised $2 billion in new funds, catapulting them to a $48 billion valuation.

Cognition 上次融资是在五月，当时估值为 260 亿美元，这意味着他们在三个月内估值几乎翻了一番。

Cognition last raised funds in May at $26 billion, meaning they've almost doubled their valuation in 3 months.

在这段时间内，Cognition 的年化收入从 4.92 亿美元增长到了目前的近 9 亿美元。

In that time period, Cognition has gone from a $492 million revenue run rate to almost $900 million at present.

除了数字之外，这轮融资表明 Cognition 将继续作为独立的 agent 实验室运营。

Beyond the numbers, the raise suggests that Cognition will continue to operate as an independent agent lab.

继 SpaceX 以 600 亿美元收购 Cursor 之后，有传闻称他们也会收购 Cognition。

Following SpaceX's acquisition of Cursor for $60 billion, there were rumors that they would pursue Cognition as well.

CEO Scott Wu 当时坚决否认了这些传闻，而这轮融资无疑为 Cognition 提供了更多的空间，继续构建他们的编码 agent Devin 并践行他们的理念。

CEO Scott Wu strongly denied the chatter at the time, and this fundraising round certainly gives Cognition more runway to continue building their coding agent Devin and pursuing their thesis.

在一篇公告中，他们写道：“我们仍处于自动驾驶软件时代的黎明。”

In an announcement post, they wrote, we're still at the dawn of the self-driving software era.

在下一阶段，agent 将默认变得主动，软件将自我改进，甚至资源分配也将变得智能，计算预算会自动分配给影响最大的用例。

In this next chapter, agents will become proactive by default, software will improve itself, and even resource allocation will become intelligent as compute budgets self-allocate towards the highest-impact use cases.

人类工程师将越来越多地扮演架构师的角色，设定目标和优先级，而 agent 则承担更多工作来实现这些目标。

Human engineers will increasingly act as architects, setting goals and priorities while agents take on more of the work to achieve them.

Cognition 写道，独立性是这一战略的核心，他们表示：“我们可以选择并组合最适合工作的模型，包括我们自己的模型，而不是将客户绑定到一家供应商。”

Cognition wrote that independence is core to this strategy, saying, we can choose and combine the models best suited to the work, including our own, rather than tie customers to one provider.

你得想想，在看到 OpenAI 因为 SpaceX 收购而切断对 Cursor 客户的访问之后，Cognition 更加深刻地认识到了保持独立的价值。

You gotta think that after watching OpenAI cut off access to Cursor customers because of SpaceX's acquisition, Cognition sees the value of staying independent even more acutely.

不过，今天 AI Daily Brief 的头条新闻就到这里。

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

接下来是主要节目。

Next up, the main episode.

F116:17

欢迎回到 AI Daily Brief。

Welcome back to the AI Daily Brief.

我得说，朋友们，完全回到“返校、复工、告别夏天”的模式真是太好了。

I gotta say, friends, it is so nice to be fully back in the back-to-school, back-to-work, out-of-summer mode.

相比其他行业，AI 行业在夏季的放缓程度肯定要小一些，但当九月来临时，你仍然能感觉到那种变化，伙计。

Relative to other industries, AI certainly has less of a summer slowdown, but you can still feel the difference, man, when September hits.

仅仅在过去 9 天里，我们就迎来了 Llama 3.1、GPT-4o，以及一个包含三个模型的 agent 产品，我们将在今天的节目中介绍它。

In the last 9 days alone, we have gotten Llama 3.1, GPT-4o, and the 3 models in one agent product that we're going to cover in today's episode.

我希望你和我一样兴奋，因为有很多新东西值得一看。

I hope you are as excited as I am because there is a lot of new stuff to check out.

首先，上周，正当我要去度假时，当然，我们收到了 Google 的新模型。

First up, last week, right as I was leaving for vacation, of course, we got a new model from Google.

它仍然不是 Pro 系列模型，但值得注意的是，Google 在其较小的 Flash 系列模型上迭代的速度非常快。

It still is not a Pro series model, but it is notable how quickly Google is iterating on their smaller Flash series models.

新的 Gemini 1.5 Flash 在 1.5 发布仅几周后就推出了。

The new Gemini 1.5 Flash comes just a few weeks after 1.5.

Google 围绕新 Flash 模型的核心主张是，它将比上一代更努力地工作。

The central claim from Google around the new Flash model is that it will work harder than the previous version.

它经过训练，可以迭代调用工具，并在复杂任务上执行更多的推理步骤，从而产生更好的结果。

It's trained to call tools iteratively and perform more reasoning steps on complex tasks, yielding much better results.

在基准测试中，这个模型看起来挺扎实的，就是有点波动。

On the benchmarks, the model looks solid if a little spiky.

它在编码基准测试 SWE-bench 上得了73.7%，只比 Claude 3 Opus 的74%差一点点，还比 GPT-5.6 SOL 高了1%。

It scored 73.7% on coding benchmark SWE-bench, just a hair shy of Claude 3 Opus's score of 74%, and outperforming GPT-5.6 SOL by 1%.

TerminalBench 就是另一回事了。

TerminalBench was another story.

这个模型在 2.1 版本上得了89.4%，跟 Opus 和 SOL 差不多。

The model scored 89.4% on version 2.1, in line with Opus and SOL.

不过，在 4.0 版本上分数暴跌，降到了19.1%，而 Opus 的分数是51.8%。

However, the scores tanked on version 4.0, falling to 19.1%, compared to, for example, Opus's 51.8%.

在 GPQA 上，分数很一般，1545 Elo 分，比 Opus 少了差不多300分，更接近 Claude Sonnet 和 GPT-5.6 Terra。

On GPQA, the scores were very middle of the road at 1545 Elo points, around 300 points shy of Opus and closer to Claude Sonnet and GPT-5.6 Terra.

Artificial Analysis 发现这个模型在他们的基准测试中表现相当扎实，得了59分，发布时排在第七位，仅仅落后于 GLM-4，只差几分就到前沿水平。

Artificial Analysis found the model was pretty solid on their benchmark run, scoring 59, slotting it in just behind GLM-4 in 7th place at the time of release, and only a few points off the frontier.

不过，这是旧版 Intelligence Index，那个给 GPT-4o 打了相当低的分数的版本，导致 Artificial Analysis 赶紧推出了他们新版指数。

However, this was the old formation of the Intelligence Index, the one that gave GPT-4o a fairly low score, prompting Artificial Analysis to rush forward their new version of the index.

一旦 AA 更新了公式，3.8 Flash 就从第7位滑到了第12位，落后于 GPT-5 6T 和 GLM-5.3 Flash。

And once AA updated their formula, 3.8 Flash slipped from 7th to 12th place behind GPT-5 6T and GLM-5.3 Flash.

这次 Artificial Analysis Intelligence Index 更新的想法是重新调整权重、重新排序优先级，并加入一些新的测试，更好地反映计算机使用和更广泛的智能体范式，而不是仅仅做那些现在基本算是入门级且饱和了的通用知识测试。

Now, the idea of this Artificial Analysis Intelligence Index update was to reweight, reprioritize, and add some new tests that better reflected the computer use and broader agentic paradigm, as opposed to just general knowledge tests, which are now pretty much table stakes and saturated.

Gemini Flash 在速度上仍然是无争议的领先者，每秒输出的 token 数比亚军 MuseSpark 1.3 多大约20%——如果你想知道这是什么东西，我们马上就会讲到——而且比 GLM-5.3 Flash 快将近4倍。

Gemini Flash remains the undisputed leader in speed, outputting around 20% more tokens per second than runner-up MuseSpark 1.3, which if you're wondering what that is, we will get to in just a moment, and almost 4 times faster than GLM-5.3 Flash.

问题当然是哪些使用场景需要那么快的速度，而要在性能上有所取舍。

The question is of course which use cases require that much speed at the cost of tradeoffs in performance.

也许他们报告里最大的亮点是成本，Artificial Analysis 写道，3.8 Flash 是我们在这个智能水平上测到的最便宜的模型。

Maybe the biggest bright spot from their writeup was cost, with Artificial Analysis writing that 3.8 Flash was the cheapest we've measured at this level of intelligence.

他们接着说，这比 Gemini 3.7 Flash 提升了40%，尽管每 token 的定价没有变。

They continued, this is up 40% from Gemini 3.7 Flash despite unchanged per-token pricing.

这得益于每个任务输出 token 数增加了30%，以及在智能体评估中更多的轮次。

Driven by a 30% increase in output tokens per task and more turns on agentic evaluations.

对谷歌来说不幸的是，正如我们接下来会看到的，随着第二天 MuseSpark 1.3 的发布，谷歌很快就失去了他们在帕累托前沿上的位置。

Unfortunately for Google, as we will see, with the release of MuseSpark 1.3 the following day, Google would very quickly lose their place on the Pareto frontier.

当然，成本和效率是谷歌宣传的重点。

Certainly cost and efficiency is a big part of the pitch from Google.

在宣布新模型时，谷歌 AI 账号写道，虽然解决模糊和高摩擦的任务非常有用，但也可能很贵。

Announcing the new model, the Google AI account wrote, while solving ambiguous and high-friction tasks is immensely helpful, it can also be expensive.

幸运的是，3.8 Flash 具有通常的投入控制功能，确保完成当前任务所需的思考量跟 token 消耗成正比。

Fortunately, 3.8 Flash features the usual effort controls, ensuring that the amount of thinking required to accomplish the task at hand is proportional to the token spend.

谷歌的 Logan Kilpatrick 强调了谷歌推出这些新版本的速度，指出这只是他们在六周内推出的第三个更新的 Flash 模型。

Logan Kilpatrick from Google emphasized the speed at which Google is putting out these new versions, pointing out that it's just their third updated Flash model in six weeks.

现在，到了用户测试的时候，大家证实它确实很快，但发现很多性能不足。

Now, when it came to user testing, people validated that it was really fast, but found a lot of performance lacking.

Intelligence AI 的 Aditya 写道，用 Kimi K3 和 3.8 Flash 构建同一个粘球游戏。

Building the same sticky ball game in Kimi K3 versus 3.8 Flash, Aditya from Intelligence AI wrote.

Flash 快得离谱，用的 token 也少得多，但实际游戏差远了。

Flash was insanely fast and used way fewer tokens, but the actual game was nowhere close.

K3 的机制、移动和整体游戏设计好得多。

K3 had much better mechanics, movement, and overall game design.

Flash 显然在速度和效率方面做得很好，但它实际能构建的东西差距还很大。

Flash clearly has the speed and efficiency part down, but the gap in what it can actually build is pretty big.

它是一个非常好的 Flash 模型，但并不等同于前沿模型。

It is a very good Flash model, but not equivalent to a Frontier model.

其他人对那种速度可能代表的意义更为乐观。

Others are more optimistic about what that speed could represent.

在另一场直接对比中，Noclip Pepe 写道：Opus 5 赢了，但 Gemini 3.8 Flash 快了39倍。

In another head-to-head, Noclip Pepe wrote: Opus 5 won, but Gemini 3.8 Flash was 39x faster.

Opus 5 花了24分钟，Gemini 3.8 Flash 花了37秒。

Opus 5 took 24 minutes, Gemini 3.8 Flash took 37 seconds.

Opus 显然更细致更精致，这点没争议。

Opus is clearly more detailed and polished, no debate there.

但能在37秒内拿到这么好的结果，就完全改变了权衡的格局。

But getting a result this good in 37 seconds completely changes the trade-off.

Opus跑完一轮的时间里，Flash理论上能跑大概39轮。

In the time Opus finished one run, Flash could theoretically finish around 39.

在什么时候，速度会比最后一点点质量更重要？

At what point does speed matter more than the last bit of quality?

而且，在座的各位肯定不意外，我仍然认为，看待这些新模型的正确方式，不是看它能不能取代你日常用的模型，而是要看有没有特定的使用场景，让它这套特有的权衡正好合适。

And it will come, of course, as no surprise to anyone here that, as always, I think that the right way to look at these new models is not whether it's going to replace your daily driver, but instead whether there are specific use cases for which its particular set of trade-offs are the right fit.

你现在做的事情里，有没有哪件是连续重复39次去迭代，会比让Opus一次性跑完更好？

Is there something you're doing right now where being able to do it 39 times in a row to iterate is likely to be better than just letting something like Opus do it once?

另一个更大的模型发布是Meta的MuseSpark 1.3，Meta首席AI科学家Yann LeCun毫不掩饰地推广了这个进步。

Now, the bigger other model release was Meta's MuseSpark 1.3, and Meta Chief AI Officer Yann LeCun was not shy about promoting the progress that's been made.

他写道，这是我们迄今为止最强大的模型。

He wrote, this is our most capable model yet.

前沿性能，几乎便宜到不用计费。

Frontier performance almost too cheap to meter.

在代理编码方面更强，可用性也更好。

Much stronger at agentic encoding with better usability.

我们认为用户会明显感受到这个飞跃。

We think users will really notice the jump.

Meta选择将他们的新模型与GPT-5.6 Sol和Opus 5进行对比，在那组比较中，它相当有竞争力。

Meta chose to compare their new model to GPT-5.6 Sol and Opus 5, and in that grouping, it was pretty competitive.

总体上，它在代理基准测试上稍微落后一点，但在编码上稍微领先。

Generally, it lagged a little behind on agentic benchmarks, but was a little ahead on coding.

Spark 1.3在DeepSWE上得了75.4分，而5.6 SOL是73%，Opus-5是74%。

Spark 1.3 scored a 75.4 on DeepSWE, Compared to 73% for 5.6 SOL and 74% for Opus-5.

在TerminalBench上，它得了88.8%，和5.6 SOL持平，比Opus-5高出几个百分点。

On TerminalBench, it scored 88.8%, which tied it with 5.6 SOL and beat Opus-5 by a couple of points.

Wang声称，相比之前的Spark 1.2，这个模型减少了20%的工具调用和25%的token消耗，同时在基准测试上也有明显提升。

Wang claimed that the model used 20% fewer tool calls and 25% fewer tokens compared to their previous Spark 1.2, while also producing a very clear jump on the benchmarks.

首次发布几天后，Meta还增加了一个Max Effort设置，进一步提升了性能。

A couple of days after the initial release, Meta also added a max effort setting that boosted performance even more.

这次发布中最让人注意的部分，是Artificial Analysis发布他们的基准测试结果的时候。

Now, the part of the release that really made everyone sit up and pay attention came when Artificial Analysis released their benchmark run.

在Max设置下，Spark 1.3在编码代理指数上得了68分，和Opus 5并列第一。

On max settings, Spark 1.3 scored a 68 on the Coding Agent Index, making it tied for first place with Opus 5.

值得注意的是，当时对Fable 5.1和GPT-6 Astra的测试还没完成，但结果仍然相当令人印象深刻。

Now worth noting that at the time testing for Fable 5.1 and GPT-6 Astra hadn't been completed, but still a pretty impressive result.

在整体智能指数上，Spark 1.3在Max设置下得了62分，排在Fable 5.1和Opus 5之后，位列第三，和Fable-5并列，比GPT-5.6 SOL高一分。

On the overall Intelligence Index, Spark 1.3 on max settings scored 62, placing it in third place behind Fable 5.1 and Opus 5, tied with Fable-5 and a point ahead of GPT-5.6 SOL.

很多人猜测，这个模型绝对是为了在基准测试上拿最高分而做了最大优化。

Now, the assumption for many is that this model was absolutely benchmark maxed to achieve the maximum possible score.

当然，这也是SemiAnalysis的观点，他们写道，Gemini 3.8 Flash和MuseSpark 1.3是我们迄今见过的最明显的两个为基准测试优化的模型。

Certainly, that's what SemiAnalysis argued, writing, Gemini 3.8 Flash and MuseSpark 1.3 are 2 of the most clearly benchmaxed models we've seen yet.

尽管它们在TerminalBench 2.1上和GPT-6以及Fable-5.1不相上下，但在TerminalBench 4.0上的表现就差得多了。

Despite being comparable to both GPT-6 and Fable-5.1 on TerminalBench 2.1, their TerminalBench 4.0 performance is markedly worse.

这怎么可能？

How is this possible?

TerminalBench 2.1里的所有任务都是完全公开的。

All of the tasks in TerminalBench 2.1 are fully public.

尽管Meta和Google不会直接在这些任务上训练，但他们绝对会从RL环境初创公司那里购买数据，这些数据被设计成尽可能模仿TB 2.1的任务。

Though Meta and Google would never train on the tasks directly, they absolutely will buy data from RL environment startups that's designed to mimic TB 2.1 tasks as closely as possible.

最终效果是一样的。

The net effect is the same.

通常你会期望TB 2.1性能的提升能泛化到其他代理任务上，但Gemini和Muse甚至连TB 4.0都泛化不了。

You'd typically expect improved TB 2.1 performance to generalize to other agentic tasks, but Gemini and Muse don't even generalize to TB 4.0.

最终，SemiAnalysis总结道，这就是所有好的公开基准测试的命运。

Ultimately, SemiAnalysis concludes, this is the fate of all good public benchmarks.

TB 4.0也不例外。

TB 4.0 is no exception.

现在它算是有用的信号，因为它两周前才发布。

It's only useful signal now because it was released 2 weeks ago.

由于所有任务都是公开的，不久就会被那些所谓的“前沿实验室”给爬坡超越。

Since all the tasks are similarly public, it won't be long until it's hillclimbed by all the aspiring quote-unquote frontier labs.

Alexander Wang 对此回应说，我们并没有宣称 MuseSpark 1.3 和 Astra 或 Fable 5.1 一样强，但它明显更具成本效益。

Alexander Wang actually responded to that one saying, we don't claim MuseSpark 1.3 is as strong as Astra or Fable 5.1, but it is significantly more cost-effective.

我们未来的模型会更直接地与那些模型竞争。

Our future models will compete more directly with those models.

同样值得一提的是，在 Artificial Analysis 修订了他们的指数之后——无疑部分原因是原先的公式让 Spark 1.3 排名超过了 GPT-6 Astra。

It is worth also noting that after Artificial Analysis revised their index, no doubt in part because the original formula had Spark 1.3 outranking GPT-6 Astra.

修订之后，Spark 1.3 以48分仍排在第五位，略高于 GPT-5.6 Sol，低于 Fable 5、Opus 5 和 Astra。

After the revision, Spark 1.3 remained in 5th place with a score of 48, which was slightly ahead of GPT-5.6 Sol and behind Fable 5, Opus 5, and Astra.

尽管在基准测试上大幅跃升，但该模型仍然极其便宜。

And despite the big jump on the benchmarks, the model is still extremely cheap.

AA 发现该模型每个任务花费0.55美元，比 Gemini 3.8 Flash 略便宜，比 GLM 5.3 便宜20%，大约只有 Opus 5 成本的四分之一。

AA found that the model spent $0.55 per task, which made it slightly cheaper than Gemini 3.8 Flash, 20% cheaper than GLM 5.3, and about a quarter of the cost of Opus 5.

总结一下，Artificial Analysis 写道，MuseSpark 1.3 Extra High 是其智能水平上最具成本效益的模型。

Summing up, Artificial Analysis wrote, MuseSpark 1.3 Extra High is the most cost-efficient model at its intelligence level.

没有得分59分及以上的模型每个任务成本更低。

No model scoring 59 or above costs less per task.

现在，对这个模型的第一印象相当积极。

Now, the first impressions on this one were pretty positive.

我一直在测试 MuseSpark 1.3 Max，老实说，它好得离谱而且效率惊人。

I've been testing MuseSpark 1.3 Max, and honestly, it's insanely good and surprisingly efficient.

MuseSpark 1.3 作为 OpenCode 上的免费小模型，有点离谱。

MuseSpark 1.3 is kind of ridiculous for a free small model on OpenCode.

它还避免了一些明显的 AI 设计陷阱，比如紫色渐变之类的。

It also avoids some of the obvious AI design traps like purple gradients and all that.

接下来我想把它推向更棘手的边缘情况，但就这个价格而言，这东西已经非常好了。

I want to push it into nastier edge cases next, but for the price, this thing is already very good.

现在回到我们在头条中讨论的关于对实验室信任的话题，对一些人来说，Meta 是个棘手的问题。

Now on the topics that we were discussing in the headlines about trust in the labs, for some Meta is a tough one.

在写了 MuseSpark 1.3 很好、基本上就是便宜版 Opus 5 之后，X 上的 ZWin 补充说，不过有个问题。

After writing about MuseSpark 1.3 being good and basically Opus 5 for cheaper, ZWin on X adds, there's a catch though.

Meta 可能会使用你的输入和输出进行训练，这正是 OpenCode 上的那个层级被称为“贡献者”的原因，也是它免费的全部原因。

Meta may use your inputs and outputs for training, which is the whole reason the tier on open code is called contributor and the whole reason it's free.

当然，很多人都很兴奋地去尝试打折的 OpenCode 版本。

Certainly many people were excited to try the discounted open code version.

Open Code 的 Dax 写道，Meta MuseSpark 已经击败 DeepSeek，成为当日使用量最大的模型。

With Dax from Open Code writing, Meta MuseSpark has dethroned DeepSeek as the most used model of the day.

这是美国模型首次登顶这个榜单。

First time an American model tops this list.

然而，如果说 MuseSpark 1.3 让一些人开始思考：“Meta 回来了吗？”

And yet, if MuseSpark 1.3 made some start to wonder, is Meta back?

真正获得更多关注的是他们新推出的个人 AI 助手 Muse。

It was the launch of their new personal AI assistant Muse that garnered even more attention.

9月8日星期二，Meta 发布了他们承诺已久的个人代理，名为 Muse。

On Tuesday, September 8th, Meta announced their long-promised personal agent called Muse.

该产品几个月来一直以代号 Hatch 在开发中，我们之前听到的基本定位是面向普通用户的 OpenCLAW，附带一堆可用性改进。

The product has been rumored to be in the works for months under the codename Hatch, with the basic pitch that we had heard being OpenCLAW for normal people with a bunch of usability improvements.

在介绍这个代理时，Alexander Wang 写道，今天我们推出 Muse，我们的新个人 AI 助手。

Presenting the agent, Alexander Wang wrote, today we're rolling out Muse, our new personal AI assistant.

Muse 始终在线，速度极快，可以使用浏览器，连接你的应用，并且设计为安全可靠。

Muse is always on, wicked fast, can use a browser, connect to your apps, and is designed to be secure.

Meta 声称 Muse 可以做我们对个人代理所期待的一切。

Meta claims that Muse can do everything we've come to expect from personal agents.

它可以整理你的收件箱、安排日历、进行预订或为你购物。

It can triage your inbox, organize your calendar, make bookings, or shop for you.

它还具备最近几个月推出的一些更令人印象深刻的功能，比如操作一个独立的虚拟计算机，这和 GrokBot 的工作原理相同。

It also has some of the more impressive features introduced in recent months, such as operating a separate virtual computer, which is the same way that GrokBot works.

Meta 似乎也认真尝试解决了早期个人助手版本带来的安全噩梦。

Meta also made a solid attempt, it seems, at dealing with the security nightmare associated with the earliest versions of personal agents.

Wang 还写道，我们的一大重点就是确保让 Muse 访问你的收件箱、日历和财务信息是安全的。

Wang again wrote, A big focus for us here was making sure it was safe to give Muse access to your inbox, calendar, and finances.

每个 Muse 都运行在它自己的安全虚拟机上，这是一个专属于你的独立计算机。

Each Muse runs in its own secure VM, an isolated computer dedicated to you.

另一个独立的系统，叫 Sentinel，会在任何内容离开虚拟机之前检查每一个操作。

A separate system, the Sentinel, checks every action before anything leaves the VM.

你的 Muse 永远不会看到你的实际密码或卡号。

Your Muse never sees your actual passwords or card numbers.

当然，Meta 内部的评价很高，CTO Andrew Bosworth，也就是 Boz，写道：今天 Muse 的发布让我非常兴奋！

Now certainly reviews from inside Meta were glowing, with CTO Andrew Bosworth, aka Boz, writing, Very excited for the launch of Muse today!

我已经在内部使用好几个月了，很难想到还有哪个产品能让我在这么短的时间内如此依赖。

I've been using it internally for months and I am hard-pressed to think of any product that I've come to rely on more in such a short period of time.

我已经把它关联到我的邮箱、日历和信用卡了。

I have it linked to my email, calendar, and credit cards.

我用它来帮我规划旅行、打包行李、研究并购买东西，还有整理我孩子学校发来的各种通知。

I use it to help me plan travel, pack for trips, research and make purchases, and sort through all the communication I get from my kids' school.

这是一个适合所有人的工具。

This is a tool for everyone.

你不需要成为专家，甚至不需要刻意去想 AI 这回事。

You don't need to be an expert or even think about AI.

你只需要在应用里或者 WhatsApp 上跟它说话，就像自己的私人助理一样，只不过它能同时并行处理很多任务。

You just talk to it from the app or from WhatsApp like your own personal assistant, except it can do lots of tasks in parallel at the same time.

当然，很快你也能通过你的 Meta Glasses 跟它对话了。

And of course, you'll soon be able to talk to it from your Meta Glasses too.

Meta 的 Jason Toff 说，今年夏天我搬到加州时，就把我的 Mac Mini 和 Mac Studio 都拔掉了，这两台机器原本都在本地跑着 Claude，然后完全切换到了 Muse。

Jason Toff from Meta said, when I moved to California this summer, I unplugged my Mac Mini and Mac Studio, both running Claude locally, and switched entirely to Muse.

它们到现在还拔着插头呢。

They're still unplugged.

我最喜欢 Muse 的一点就是它用起来非常自然。

My favorite thing about Muse is how natural it feels.

你像跟人一样跟它说话，它就像顶级私人助理一样回应你。

You talk to it like a person and it responds like a top-notch personal assistant.

即使从外部来看，早期的评价也相当正面。

And even from the outside, early reviews are pretty positive.

EAC 的精神导师 Beth Jezos 写道，我提前用到了这个产品。

EAC spiritual guru Beth Jezos writes, got to try this product early.

它非常扎实，功能也很丰富。

It's very solid and quite feature-rich.

既然模型智能不再是实用性的瓶颈，你的生活上下文就成了关键，在安全计算上运行的个人助手才是正确的方向。

As model intelligence is no longer the bottleneck for utility, context on your life is, and personal agents running on secure compute is the way.

Muse 用起来确实是个让人印象深刻的产品。

Muse has been a genuinely impressive product to play with.

它具备目前 iMessage 助手的所有功能，同时还拥有让数亿人接触到的所有要素，包括通过 Instagram 的巨大好友关系图、来自邮箱和其他你连接的服务日益丰富的上下文，更重要的是还有像 Facebook Marketplace 这样的东西。

It has all the functionality of the iMessage agents in flight today, but also has all of the ingredients that will expose it to hundreds of millions including a massive friend graph through Insta, increasingly rich context from email and other services which you connect, and more importantly stuff like Facebook Marketplace.

特别是 Marketplace，它是一个非常棒的推广切入点。

Marketplace in particular is an incredible distribution wedge.

数百万普通用户可能只是因为一个助手帮他们找东西、谈价格、安排取货，就接触到了 Muse。

Millions of normal people could encounter Muse simply because an agent helps them try to find something, negotiate the price, and arrange pickup.

FB 这波执行得相当不错！

Pretty good execution here from FB!

a16z 的 Olivia Moore 说，她很喜欢应用内丰富的连接器库，认为原生连接器比浏览器使用更可靠，还说她喜欢可以设置与数据相关的目标，并附上附件来可视化进度。

Olivia Moore from a16z said that she likes the rich library of connectors that are available in-app, thinking that the native connectors will be more reliable than browser use, and also said that she liked that she can set up goals connected to that data and attach artifacts to them to visualize progress.

她担心界面还是有点杂乱，要做的事情也太多，但总结说这可能是最早一批真正面向主流消费者的个人助手，有望获得广泛采用。

She worried that the UI was still too cluttered and there was a few too many things to do, but concluded this could be one of the first true mainstream consumer agents to get adoption.

不过，在消费者信任方面，Meta 还有一些难关要过。

Still, Meta has some hills to climb when it comes to consumer trust.

同样是那位 Olivia Moore 写道，我在 Muse 上点击连接邮箱按钮时，比我用过的十多个创业公司的助手产品都要犹豫。

That same Olivia Moore wrote, I was more reluctant to press the connect email button on Muse than on 10+ startup agent products I've tried.

在我看来，Meta 的分发优势其实是一把双刃剑。

In my opinion, Meta's distribution advantage cuts both ways here.

我真的愿意把我的个人数据交给一个代理，然后让它在我所有朋友都在的网络上自由行动吗？

Do I really want to give an agent my personal data and then set it loose on networks where all my friends are?

Meta 在整个人工智能竞赛中之所以值得关注，是因为在同等规模的公司里，他们是唯一一家主要聚焦于消费者端而非企业端应用场景的公司。

One of the things that makes Meta interesting and worth paying attention to in the broader AI race is that they are the only company at their scale that is primarily focused on a consumer rather than a business use case.

当然，这个界限有点模糊，尤其是考虑到大量小企业把 Meta 的产品当作主要沟通渠道，但归根结底，我觉得说 Meta 更关心消费者而不是B2B，这一点没什么争议。

Now, obviously this is all a little blurry, especially when you consider the legions of small businesses that use Meta products as their key communication channels, But ultimately, I think it's pretty uncontroversial to say that what Meta cares about is consumers more than B2B.

有一段时间，OpenAI 看起来两头都想抓，名义上他们现在也还是这样。

For a while, OpenAI looked like it was going after both, and nominally they still are.

但当然，Anthropic 带来的压力意味着他们最近不得不在B2B和工作场景上投入更多资源，尤其是越来越多人在质疑普通消费者到底会不会真的在意AI代理。

But of course, the pressure from Anthropic has meant that they have really had to focus a lot more resources on the B2B and work use cases of late, especially as there are more and more questions about whether general consumers will ever really care about AI agents.

Meta 的这类实验意义远远超出他们自身。

These sort of experiments from Meta have significance that goes beyond just them.

代理购物真的能成为现实吗？

Does agentic shopping actually become a thing?

人们真的喜欢有个个人代理助手帮他们处理日常事务，比如订旅行吗？

Do people really like having a personal agent assistant to help them with daily things like booking travel?

我们得等到这些东西都上线了，而且足够好用、能兑现承诺，才能真正知道答案。

We're not really going to know until those things are available and broadly good enough that they actually do what they promise.

而感觉Meta终于到了一个层面，这个承诺有可能成真。

And it feels like Meta is finally playing at the level where that promise might be real.

Box 的 Aaron Levie 写道，个人助理代理将会是一个非常令人兴奋的人工智能类别。

Writes Box's Aaron Levie, personal assistant agents are going to be a very exciting AI category.

这是第一次你能为消费者提供高token量的代理用例，而且说得通。

It's the first time you can have high token volume agentic use cases that make sense for consumers.

现在很多不同的方法在涌现，而且竞争会非常激烈，因为这些代理未来会介入大量消费者的支出。

Lots of different approaches emerging right now, and it's going to be hypercompetitive because these agents will mediate a lot of consumer spend over time.

但这正好契合了 Meta 的优势。

But this certainly plays directly to Meta's strengths.

需要大量算力，可以通过广告和电商变现，软件驱动体验所以能够大规模分发，等等。

Lots of compute required, can monetize with ads and commerce, software-focused experiences so can distribute it at scale, and so on.

Y Combinator 总裁 Garry Tan 总结说，代理之战已经全面打响，Muse 非常令人印象深刻。

Sums up Y Combinator president Garry Tan, harness wars are full on now and Muse is very impressive.

我本人并不参与这场代理之战或模型之战，但我肯定会支持这个产品类别，哪怕仅仅因为人们真正从个人助理代理中获得价值，可能会让他们对人工智能的第一印象不那么抵触。

Now I don't have a horse in the harness wars or the model wars, But I will certainly be rooting for this as a product category, if for no other reason than people actually getting value out of a personal assistant agent might make them just a little less hostile to AI in the first place.

最后，还有一个模型发布要聊。

Lastly, one more model release to talk about.

同样是在周二，OpenAI 发布了 ChatGPT Images 2.5。

Also on Tuesday, OpenAI released ChatGPT Images 2.5.

这是 ChatGPT 内置图像生成功能的最新模型系列，这个功能现在仍然被大量使用。

This is the latest in the series of models that power built-in image generation in ChatGPT, a feature which still gets a ton of use.

OpenAI 说用户每周生成超过30亿张图片，并称新模型将提供更清晰的细节、更精确的编辑，以及更快的生成速度，延迟降低50%。

OpenAI says that users are generating more than 3 billion images a week and write that the new model will provide sharper details, more precise editing, and faster generation with a 50% reduction in latency.

同时，OpenAI 还推出了一个新功能叫 Sketch，顾名思义，你可以在应用内直接画图输入来引导图像生成。

Alongside the model, OpenAI is releasing a new ChatGPT feature called Sketch, which, as the name suggests, allows you to draw an input to help guide your image generation directly in the app.

用户还可以添加文字描述来说明特定风格或提供额外细节来指导模型输出。

Users can add a text prompt to describe a particular style or provide additional details to guide the model output.

该模型有两个变体：Flare，快速版本，适合快速迭代；Sunburst，为专业工作流优化，在编辑时需要更好的控制。

The model comes in 2 variants: Flare, which is the fast version designed for quick iteration, and Sunburst, which is optimized for professional workflows that require better control across edits.

我觉得“控制”这个词真的很关键。

And I think that that word control is really key here.

就像 Nano Banana 的重大创新和更新是更精细地控制编辑过程一样，这似乎也是 OpenAI 这款新模型追求的重点。

In the same way that the big innovation and update of Nano Banana was more fine-grained control over the editing process, that seems to be a big part of what OpenAI is going for with this new model as well.

Higgsfield 的产品主管 Alam Kulov 兴奋地写道，GPT Image 2.5 Flare 最让我们印象深刻的是，它非常清楚哪些地方不该改变。

Exultant Alam Kulov, the head of product at Higgsfield, wrote, what impressed us most about GPT Image 2.5 Flare is how well it understands what not to change.

你可以做有意义的编辑，同时不会丢失原图的角色、构图或视觉特征。

You can make a meaningful edit without losing the character, composition, or visual identity of the original image.

这对于创作者和团队在电影、UGC 和广告中的实际操作来说极其重要。

That's incredibly important for the way creators and teams actually work across film, UGC, and advertising.

当你将那种控制力与速度、质量和成本结合起来时，Image 2.5 Flare 确实脱颖而出。

And when you combine that level of control with the speed, quality, and cost, Image 2.5 Flare really stands out.

一个你经常看到的例子是，新的、更好的控制和图像一致性带来的结果，就是像定格动画这样的全新类型首次变得可行。

One example that you're seeing a lot of, of what the new, better controls and image consistency can lead to is entire new genres like stop-motion animation that become viable for the first time.

我越来越感觉到的一件有趣的事情是，我认为目前总体上，我们低估了图像的价值，不仅是作为消费者差异化的因素，实际上也是作为OpenAI的商业用例差异化的因素。

One of the interesting things that I increasingly feel is I think right now in general, we underappreciate the value of images, not just as a consumer differentiator, but actually as a business use case differentiator for OpenAI.

举个例子，虽然总体上我仍然更喜欢Fable创建的网站的美学，而不是GPT创建的网站，但我可以调用GPT图像来生成UI的某些方面或某些类型的美学，这一事实带来了很大的不同，并使我比原本更频繁地使用Codex中集成的GPT模型和图像生成功能。

As a, for example, while in general I still like the aesthetics of Fable-created websites better than GPT-created websites, the fact that I can call upon GPT image to generate aspects of the UI or certain types of aesthetics makes a pretty big difference and leads me to use the integrated GPT models and image generation in Codex more often than I otherwise would.

关键是，尽管这次更新感觉像是例行公事，但不要忽视它可能有多么重要。

Point is, although this update feels routine, don't sleep on how significant it could be.

那么这就是目前关于新模型的情况。

So that is the new model story for now.

就像我说的，有很多令人兴奋的好东西可以尝试，而且我相信接下来还会有更多。

Like I said, lots of exciting goodies to try out, and I'm sure there is more on the way.

那么，今天的AI Daily Brief就到这里。

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

一如既往地感谢您的收听或观看，下次再见，和平。

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）

F11:10广告 · 已剔除

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

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

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

And finally, if you haven't yet, you can check out our latest free self-directed training program.

It is called the Multiplayer AI Sprint for Teams.

And basically the idea is to shepherd you through a process of figuring out how to build agents that don't just help you, but actually sit at the intersection of work that is shared across your teams.

I'm pretty convinced that this is the next big paradigm for AI inside companies.

And so I wanted to build a sprint that could help you guys fully embrace that.

There's of course a link to that on the aidailybrief.ai website.

But you can also find it at multiplayerai.ai.

F113:17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 Ema, May Habib of Writer, 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.

Here's why most legacy modernization projects fail.

The AI doing the work can't understand codebases at scale.

It sees a small slice of context, examines syntax, and misses years of decisions distributed across the global application ecosystem.

Blitzy solves this the way it solves everything.

Grounded in your code before any migration begins, Blitzy's agents reverse engineer the entire legacy system into a persistent knowledge graph.

Every dependency, every constraint, every piece of tribal knowledge that used to live in one engineer's head.

From that understanding, Blitzy autonomously executes language migrations, framework upgrades, and monolith-to-microservices transformations, all validated end-to-end.

One Blitzy customer modernized a $10 million monolithic insurance stack in 16 weeks against a 137-week baseline with coding agents.

That's 9x compression.

Retire technical debt while accelerating your roadmap.

See how at blitzy.com.

That's blitzy.com.

Here's a harsh truth.

Your company is probably spending thousands or millions of dollars on AI tools that are being massively underutilized.

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

Most employees are still using AI to summarize meeting notes.

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

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

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

The result?

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

Your employees move from meeting summaries to solving actual business problems.

And you can prove the ROI.

Stop guessing if your AI investment is creating value.

Check out Section at sectionai.com.

That's S-E-C-T-I-O-N-A-I dot com.

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

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

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

Marketing agents turn competitor moves into landing pages.

Sales agents enrich leads, draft emails, and updates the CRM.

Ops agent chases the paperwork and tracks the budget.

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

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

Hire yours at HyperAgent.

Get $100 in credits at hyperagent.com/AIDailyBrief.