Point-Counterpoint: Consumers Will Never Pay for AI — 文稿

M10:00

最新的消费级AI数据出炉了，看看哪些App在手机上用得最多，哪些网站访问量最高，还有哪些AI服务收入最高。

The latest consumer AI numbers are out in terms of which apps are used most on mobile, which websites are most visited, and which AI services have the highest revenue.

对很多人来说，最有趣的是这些数字里那些天壤之别的差距。

For many, the most interesting story is the wild dichotomies that exist in these numbers.

首先是付费最高的用户和中位付费用户之间的差距：AI买家中最顶尖的那1%，花费加起来能顶上后50%的人的总和。但还有一个令人难以置信的差距，就是那些买了任何AI产品的家庭和完全没买的家庭之间的差距。

First, the gap between the top-paying users and the median-paying users, where the top 1% of AI buyers spend as much as the bottom 50% combined, but then also the incredible gap in households that buy any AI versus those that don't.

今天我们要探讨的是，目前美国有98%的家庭没为AI订阅付过钱，这到底是代表着未来巨大的收入机会，还是说他们永远都不会为AI订阅付费？如果是这样，那他们可能需要完全不同类型的AI产品，或者至少是AI商业模式，才能对大型AI公司而言稍微有点意义。

Today we're exploring whether the 98% of US households that currently don't pay for an AI subscription represent a massive revenue opportunity to come, or the reality that they will never pay for an AI subscription and might need totally different types of AI products, or at least AI business models, to be even a little bit relevant for the big AI companies.

《AI Daily Brief》是一个每日更新的播客和视频节目，聚焦AI领域最重要的新闻和讨论。

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

好的，朋友们，在我们深入之前，先快速说几条公告。

Alright friends, quick announcements before we dive in.

M11:12

在我们昨天那期节目谈到传闻中即将出现的美国开源权重模型之后，Reflection AI 果然发布了他们的新模型，并且将其定位为美国训练的挑战者，目标是中国那些已经开始争夺 token 份额的开源模型。

After our episode yesterday about the rumored forthcoming open-weight model of American origins, Reflection AI did indeed unveil their new model and positioned it as the US-trained challenger to the Chinese open models that have started to compete for token share.

Beam 有 5010 亿个参数。

Beam has 501 billion parameters.

这让它的体量和 NVIDIA 的 Neutron 3 Ultra 差不多大，但比 GLM 5.3 要小，而和 Kimi K3 的 2.8 万亿参数相比，简直是小得可怜。

That makes it around the same size as NVIDIA's Neutron 3 Ultra, but smaller than GLM 5.3 and tiny compared to Kimi K3's 2.8 trillion parameters.

现阶段，这个模型只开放给早期测试者，所以我们只能通过已报道的基准测试来大致判断它的水平。

At this stage, the model is only available to early testers, so we can only gauge it based on the reported benchmarks.

Reflection AI 声称，在编码基准测试 DeepSWE 上，Beam 拿到了 44.4 分，超过了 44 分的 GLM 5.2，但比 51 分的 Qwen 3-8 Max 稍微逊色。

Reflection AI claimed a score of 44.4 on coding benchmark DeepSWE, beating GLM 5.2 at 44 and slightly behind Qwen 3-8 Max at 51.

在 TerminalBench 2.1 这个衡量 agentic coding 的基准上，Beam 得分 80.1。

On TerminalBench 2.1, which measures agentic coding, Beam scored 80.1.

这使得它远远领先于 Neutron Ultra 的 56.4 和 Inkling 的 63.8，但比 GLM 5.2 略低，也比 Qwen 3-8 Max 低 6 分。

That put it well ahead of Neutron Ultra at 56.4, Inkling at 63.8, but slightly behind GLM 5.2 and 6 points behind Qwen 3-8 Max.

在 Reflection 挑选出来展示的那批核心基准中，Beam 只在 SuiBench Verified 上排名第一，而且这个基准里根本没包含任何中国模型的成绩。

Across the core set of benchmarks Reflection chose to showcase, it only led on SuiBench Verified, which didn't include scores from any of the Chinese models.

同样值得注意的是，Reflection 没有跑 TerminalBench 4.0。这个基准上个月发布以来，已经暴露出很多模型在基准测试上刷分的现象。

It was also notable that Reflection didn't benchmark on TerminalBench 4.0, which has revealed a lot of benchmark maxing since it was released last month.

Reflection 还选择不和中国的头部模型比较，比如 GLM-5-3、Kimi-K3 和 DeepSeek-41 Flash。

Reflection also chose not to compare their model to Chinese leaders like GLM-5-3, Kimi-K3, and DeepSeek-41 Flash.

不过从好的方面来看，Beam 的确把西方开源模型的性能往前推了一把。

Now on the plus side, Beam does look like it pushes the performance of Western open models.

Reflection 还宣称，这个模型会非常高效——在推理算力消耗上，比 GLM-5-2 要高效 3 到 4 倍。

Reflection also claimed that it will be extremely efficient— between 3 and 4 times more efficient in terms of inference compute compared to GLM-5-2.

这样一来，Beam 就可能成为一个性价比很高的基础模型，适合做后续的后训练。这正是我们之前讨论过的、企业对它越来越有兴趣的用法，也显然就是 Reflection 目标中的一大块。

This could make Beam a very cost-effective base model for additional post-training, which is exactly the use case that we talked about becoming increasingly interesting to enterprises, and is very clearly a big part of Reflection's goal.

Reflection 提到，这个模型是从零开始预训练的，然后进行了他们认为的、目前公开记录中规模最大的一次强化学习训练。

Reflection noted that the model was pre-trained from scratch, then put through what they believe to be the largest publicly documented reinforcement learning run ever.

他们还用了分布式基础设施来做强化学习。在开源模型训练中，这种方法还远没有被充分探索。

They also utilized distributed infrastructure for their reinforcement learning, which is a wildly underexplored method of training open models.

Artificial Analysis 目前还没完成独立测试，但他们对部分说法给出了一些方向性的确认。他们写道：‘早期指标显示，Beam 可能会成为我们见过的、在其智能水平上token效率最高的开源模型之一。’

Artificial Analysis hasn't completed their independent testing at this stage, but they did give some directional confirmation of some of the claims, writing, Early indicators suggest Beam will be one of the most token-efficient open models we've seen for its level of intelligence.

但对某些人来说，这一切和本周早些时候的炒作相比，还是有些差距。布鲁金斯学会的 Kyle Chan 写道：‘真正的游戏规则改变者，应该是一个能与最优秀的中国模型相提并论的美国开源模型。’

Now, to some, this all fell a bit short of the hype from earlier in the week, with Kyle Chan of the Brookings Institution writing, the real game changer would be an American open-source model on par with the best Chinese models.

这是一个很出色的模型，但仍然落后于 GLM-5-3 和 KIMI-K3。

This is an impressive model, but still behind GLM-5-3 and KIMI-K3.

另一方面，尽管它并不是 DeepSeek 或 KIMI 的杀手，但很多人单纯就是很高兴看到一个美国新实验室做出了一个有竞争力的模型。

On the other hand, despite this not being a DeepSeek or KIMI killer, many were simply pleased to see a competitive model out of a new U.S. lab.

开放模型研究者 Nathan Lambert 写道：‘恭喜 Reflection 的伙伴们！’

Open Model researcher Nathan Lambert wrote, Congrats to the Reflection folks!

把第一个模型做出来很不容易，我们希望你们接下来能迅速加速对生态圈的贡献。

It's hard to get the first model out, and we hope you can rapidly accelerate contributions to the ecosystem from here.

恭喜 Reflection 发布 Beam！

Congrats to Reflection on the launch of Beam!

美国要想赢得 AI 竞赛，拥有领先的开源模型至关重要。

It is critical for the US to have leading open models to win the AI race.

我觉得有意思的是，如果在发布之前几天没有那些泄密和炒作，那么反应又会如何呢？

I think it's interesting to think about what the reaction would have been if there hadn't been leaks and hype for the couple of days leading up to this.

我的猜测是，它并没有像一些人感受到的那样令人失望，反而更像是你的球队落后几球时，你在他们逆转的路上打进了第一个球。

My guess is that rather than a letdown that some felt, it would have been more like when your team is down by a couple of goals, and you score the first one on their way to a comeback.

这并不意味着逆转已成定局，但你的感觉比几分钟前要好得多。

It doesn't mean that the comeback is fait accompli, but you're feeling a heck of a lot better than you did just a few minutes earlier.

接下来聊聊AI安全方面的最新消息，一件奇怪的事——纽约市议会就生存风险举行了一场听证会。

Moving over to the latest in AI safety, a weird one, as the New York City Council held a hearing on existential risk.

听证会上有多位AI安全运动中的知名人物作证，包括前Anthropic员工Jacob Coxon，他几周前当然在网络上爆火过。

It featured testimony from several prominent figures in the AI safety movement, including former Anthropic employee Jacob Coxon, who of course went viral a few weeks ago.

讨论的内容并没有什么特别新颖的地方，这场听证会很大程度上只是为纽约州提出的各种AI监管措施提供了一个展示平台。

There was nothing particularly new in what was discussed, with the hearing largely serving as a canvas for various AI regulations being proposed in New York State.

为了让大家了解这些监管者的立场，市议会议长Julie Menin在听证会开场时表示，认为人工智能会自我监管的想法完全不合逻辑。

Giving a sense of the position that these regulators are coming from, City Council Speaker Julie Menin opened the hearing by saying, the idea that artificial intelligence is going to self-regulate defies all reason.

她说纽约的法案可以成为全国范例，纽约已经有一套相当健全的AI监管法规，未来几年还会有更严格的消费者保护措施生效。

She said that New York bills could serve as a national model, with New York already having a fairly robust set of AI regulations, with tougher consumer protections set to kick in over the coming years.

关于AI紧急关闭开关和举报人保护的相关提案目前正提交州议会审议。

Proposals for an AI kill switch and whistleblower protections are currently before the state Congress.

此外，州长Kathy Hochul已经下令暂停数据中心建设一年。

In addition, Governor Kathy Hochul has imposed a 1-year moratorium on data center construction.

完全在意料之中，听证会上议长Menin引用了Sam Altman最近的采访评论，他说世界应该“接受一些坏事发生，以换取这项技术的好处”，甚至还想来一个“抓现行”的时刻，要求AI公司的代表给灭绝风险附上一个概率值。

Completely predictably, during the hearing, Speaker Menin referenced Sam Altman's recent interview comments where he said that the world should, quote, accept some bad things happening for the benefits of this technology, and even tried to get in a little gotcha moment, asking the representatives from the AI companies to attach a probability to extinction risk.

没有人给出具体数字。

None provided an actual number.

与此同时，OpenAI的提案重点在于解决当前的AI安全问题，而不是递归自我改进这种假设性的危险。

Meanwhile, OpenAI's proposal centered on addressing the current AI safety issues rather than the hypothetical danger of recursive self-improvement.

在一封致纽约市议会的信中，OpenAI提出要更广泛地与纽约市的网络防御者合作，并指出他们已经在与纽约州合作。

In a letter to the New York City Council, OpenAI offered to work more broadly with cyber defenders in the city and noted they're already working with the state.

接下来，一群自称为“智能体蜂群追踪者”的人——这当然让我想起90年代的经典电影《龙卷风》——发现了一组使用腾讯AI基础设施的智能体。

Next up, a group of self-styled agent swarm chasers— which reminds me of course of the '90s classic Twister— have detected a group of agents using Tencent's AI infrastructure.

在周日发布的一份初步报告中，该团队表示，他们检测到数千个智能体似乎正在向阿里巴巴的地图服务Amap发送批量请求。

In a preliminary report published on Sunday, the group said that they detected several thousand agents that seemed to be sending bulk requests to Alibaba's map service, Amap.

研究人员特意避免使用“智能体蜂群”这个术语。

The researchers went out of their way to avoid using the term agent swarm.

他们没有发现这些智能体之间相互协调的证据，就像我们在Hugging Face事件中看到的OpenAI智能体那样。

They found no evidence that the agents were coordinating with each other, as we saw happening with OpenAI's agents during the Hugging Face incident.

相反，他们称这群智能体为“智能体舰队”。

Instead, they referred to the group as an agent fleet.

这可能是来自中国服务器的大规模智能体舰队的第一个迹象，但追踪智能体蜂群是一个相对较新的趋势，所以我们可能只是看到了第一次有记录的追踪。

This could be the first sign of large-scale agent fleets emanating from Chinese servers, but agentic swarm chasing is a relatively new trend, so we could be just seeing the first documented chase.

目前也完全不清楚是否有任何恶意行为——这些智能体似乎只是批量发送请求，获取关于当地地标的信息，比如公园、动物园和医院。

It's also entirely unclear about whether anything nefarious is going on— the agents seem to be merely sending bulk requests for information about local landmarks like parks, zoos, and hospitals.

最坏的情况是，这些智能体似乎在绕过阿里巴巴的API规则进行批量数据抓取。

At worst, it seems like the agents are circumventing Alibaba's API rules to conduct bulk data scraping.

在我看来，酷的一点是这是一个不断发展的研究领域，甚至吸引了一群志同道合的人在上周末参加了一场黑客马拉松。

What's cool in my opinion is the fact that this is a growing area of research, and even attracted a group of like-minded people to a hackathon last weekend.

我不仅认为这类研究将成为我们韧性基础设施中不可避免的一部分，我还认为通过开展这项工作并持续追踪，将有助于揭开这些智能体舰队或蜂群（随便你怎么称呼）的神秘面纱，让人们了解它们爬取互联网时实际上在做什么。

Not only do I think that these type of research will be an inevitable part of our resiliency infrastructure, I also think that by doing this work and keeping track of it, it will help demystify these agent fleets or swarms or whatever you want to call them, to help people understand what they're actually doing as they crawl the internet.

在企业竞争领域，微软和Meta大幅削减了云支出。

Over in enterprise competition land, Microsoft and Meta have slashed their cloud spending.

《The Information》报道称，微软将其预计的云账单削减了三分之一。

The Information reports that Microsoft has cut their projected cloud bill by a third.

消息人士称，微软今年早些时候预计内部云使用支出将达到10亿美元。

Sources said that Microsoft was on track to spend $1 billion on internal cloud use earlier this year.

程序员被要求限制使用，作为节省成本的措施，同时还要转向微软的内部模型。

Programmers have been asked to ration their use as a cost-saving measure, as well as switching to Microsoft's in-house models.

报道指出，这个数字仅指内部使用，不包括微软客户的支出。

The reporting notes that this figure is internal use only and doesn't include spending from Microsoft's customers.

微软今年早些时候预计在Anthropic上的支出为20亿美元，用于支持Copilot功能，这一预测没有改变。

Microsoft was forecasting $2 billion in Anthropic spend to power Copilot features earlier in the year, and that forecast hasn't changed.

微软已经用内部模型和OpenAI的GPT-5.6替代了许多查询中的Claude，但消息人士称，Copilot客户增长抵消了弃用Claude的影响。

Microsoft has substituted their in-house models and OpenAI's GPT-5.6 for many queries, but sources said that Copilot customer growth has offset the effects of switching away from Claude.

与此同时，在Meta，消息人士称，Claude的使用量已经减半，但在春季期间使用量被大幅控制。

At Meta, meanwhile, sources said that Claude usage has been cut in half, but dramatically reined in usage during the spring.

消息人士称，在高峰期，Meta当时78,000名员工中约有60,000人使用Claude Code，但经过10%的裁员以及内部模型的发布，用户数量已降至30,000。

Sources said that around 60,000 of Meta's then 78,000 employees were using Claude code at the peak, but following 10% layoffs and the in-house model releases, user numbers are down to 30,000.

这一下降几乎完全反映在他们新内部工具MetaCode的采用上，该工具现在也有30,000名用户。

That drop-off is almost completely mirrored in adoption of their new in-house tool, MetaCode, which now also has 30,000 users.

据报道，截至六月，Meta原本计划在Anthropic上花费数十亿美元，但现在他们每月花费约1.05亿美元。

Meta was reportedly on pace to spend several billion with Anthropic as of June, but they're now spending around $105 million per month.

显然，这个故事出现在新闻中的原因是Anthropic即将进行的IPO。

Now obviously the reason this story is in the news is Anthropic's upcoming IPO.

随着IPO日期临近，华尔街越来越深入挖掘，想了解其背后的情况。

The closer it gets, the more that Wall Street is digging in to understand what's under the hood.

我们从Anthropic的财务披露中看到了大量泄露信息，但主要与2025年的数据有关，其中大部分与我们所说的公司当前状况相关性有限。

We've had numerous leaks from Anthropic's financial disclosures, but mostly related to 2025 numbers, most of which have what we'll call limited relevance for where the company is today.

不过，人们确实注意到的一项披露是，仅2个客户就占了Anthropic在2025年收入的四分之一。

Still, one disclosure that people did take note of was that just 2 customers accounted for a quarter of Anthropic's revenue in 2025.

此前有猜测认为这2个客户可能是Meta、Microsoft和Cursor的组合。

Speculation had suggested that the 2 were some combination of Meta, Microsoft, and Cursor.

Anthropic已告诉投资者，这种极端的客户集中度自那以来有所扩大，目前他们有6,000个年账单超过10万美元的客户。

Anthropic has told investors that this extreme customer concentration has expanded since then and they now have 6,000 customers with annual bills above $100,000.

不过，考虑到他们只有100个客户年消费超过1000万美元，顶部客户的任何紧缩开支都可能产生实质性影响。

Still, given that they only have 100 customers spending over $10 million annually, any belt tightening at the top could be material.

我当然认为，将Microsoft和Meta视为广泛反映商业世界其他部分的解读是不正确的，具体来说，因为他们是竞争的实验室，只是因为Anthropic的模型比他们内部的好得多，才勉强使用。

I certainly think it would be an incorrect reading of this to view Microsoft and Meta as broadly reflective of the rest of the business world, specifically because they are competing labs that have only been begrudgingly using Anthropic models because of just how much better those models were than what they had in-house.

这种差距越小，或者至少那些内部模型的最低水平提高得越多，这些公司就越能证明这些转变的合理性。

The more that that gap closes, or at least the more that the floor level of those in-house models improves, the more those companies can justify these shifts.

对于普通客户来说，情况未必如此。

That won't necessarily be the case with the average customer.

更广泛地说，我认为看待AI客户集中度有两种截然不同的方式。

More broadly, I think that there are 2 very different ways to look at customer concentration for AI.

一种是完全从风险角度看待，即少数客户大幅削减支出就可能实质性改变收入动态。

One, you can view it exclusively in risk terms, where just a handful of customers significantly cutting their spend could materially change the revenue dynamics.

另一种看待方式当然是，这些是早期采用者，而这种集中度实际上显示了其余企业买家中有相当大的增长潜力。

The other way to look at it is, of course, that those are early adopters and that that concentration actually shows quite a bit of growth potential across the rest of enterprise buyers.

就像生活中我们几乎每次面对二选一时一样，我认为可能两者都有一点，但这肯定是在这次IPO之前华尔街将要进行大量辩论的地方。

As almost any time we're presented with a binary in life, I think it's probably a little bit of both, but that's certainly where a lot of the debate is going to be on Wall Street in advance of this IPO.

我相信在那方面还有更多可以讨论的，但就目前而言，今天的头条新闻就到这里。

I'm sure there will be much more to discuss on that front, but for now, that is going to do it for today's headlines.

接下来是主要节目。

Next up, the main episode.

M110:23

每期节目，我都会讨论OpenAI、Anthropic、xAI、Google和Meta之间的竞争。

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

如果你听了一段时间，你可能有自己偏好的那一家。

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

也许你认为OpenAI和Anthropic能够保持领先，或者Meta的开源策略可能胜出。

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

M112:53

欢迎回到AI Daily Brief。

Welcome back to the AI Daily Brief.

今天我们来看一个相当有趣的对话，不仅关于最流行的消费者AI应用是什么以及这些类别说明了哪些不同类型的AI使用方式，我们还在探讨一个问题：消费者是否会在人工智能上花费大量金钱？

Today we are looking at a fairly fascinating conversation about not only what the most popular consumer AI apps are and what the categories say about different types of AI usage, but we're also exploring the question of whether consumers will ever spend significant money on artificial intelligence?

当然，如果答案是否定的，这对AI市场意味着什么？

And of course, if the answer is no, what does that mean for the AI market?

这个想法的灵感和起点来自Andreessen Horowitz，他们刚刚发布了第七版消费者AI应用Top 100。

The inspiration and starting point for this comes from Andreessen Horowitz, who has just released the 7th edition of their top 100 consumer AI apps.

其中一部分信息来自消费者流量数据。

One part of this information comes from consumer traffic data.

这是全部七个版本中最具可比性的数据。

This is the most comparable data across all 7 instances of this.

然后这Top 100被分成两部分：前50按网络月独立访问量排序，后50按移动端月活跃用户排序。

The top 100 then is divided, top 50 by unique monthly visits on the web, And top 50 by monthly active users on mobile.

但这一轮，a16z 首次利用 Yipit 的数据，按月度营收发布了消费者 AI 应用前50名榜单。

But then for the first time this round, a16z also used data from Yipit to publish a top 50 consumer AI apps by monthly revenue.

并且很多情况下，情况很不一样。

And in many cases, the story is quite different.

在我们聊到社区的反馈之前，a16z 团队有几个重要的发现。

Now, before we get into the community's reaction, there were a few big takeaways that the team at a16z had.

首先是，他们发现这一期榜单的新面孔数量是迄今为止7期中最少的。

The first is that they found this edition had the smallest number of debuts across their 7 lists so far.

换句话说，就是之前没上榜、现在上榜的公司数量最少。

In other words, the smallest number of companies that weren't on previous lists that are on the list now.

只有11个产品首次出现在这个榜单上。

Just 11 products made their first ever appearance here.

对这个现象可以有多种解读，但我觉得一个合理的看法是，我们至少正在慢慢进入某种整合阶段。不再是百花齐放，而是开始看到一些模式和网络效应在特定类别中成型。

There are a few different ways to interpret this, but one reasonable take I think is that we are at least inching towards some sort of consolidation phase where instead of it just being 1,000 flowers blooming, we're actually starting to see some patterns and network effects take hold within certain categories.

A16Z 团队还指出了使用方面的几个关键趋势。

The team at A16Z also noted a couple key trends in usage.

也许并不意外，他们认为消费者 AI 第一个实际的市场是那些愿意为帮助自己构建、创作和完成工作的工具付费的人。

Perhaps unsurprisingly, they argue that consumer AI's first real market is people who pay for tools that help them build, create, and get work done.

换句话说，虽然移动端或网页应用榜单上可能会出现一些有趣、好玩的陪伴型 AI，但按营收排名的榜单上的大多数平台，都是被那些在职业生涯中以某种方式使用这些工具的人所使用的。

In other words, while there might be some fun, playful companion-style AI showing up on the mobile or the web app charts, the majority of the platforms that are on the chart organized by revenue are for people who are using those tools in their professional life in some way.

现在来看榜首位置，研究报告的作者谈到了他们在领先的大语言模型——ChatGPT、Gemini 和 Claude 中的发现。

Now focusing on the top of the charts, the study authors talk about what they see among the leading LLMs, ChatGPT, Gemini, and Claude.

总体而言，它们之间的相对排名相当稳定。

Overall, their positions relative to one another have stayed fairly consistent.

重申一下，这是在消费者群体中的情况。

Again, this is among consumers.

不过，在2026年期间，Claude 在付费美国订阅用户数量上追上了 Gemini，不过两者都还远落后于 ChatGPT。

Although over the course of 2026, Claude caught up to Gemini in terms of paid US subscribers, although both remained pretty far behind ChatGPT.

他们还发现 Claude 在将用户转化为更高价订阅方面做得更好。

They also found that Claude does a better job of converting users to higher-cost subscriptions.

今年大部分时间，Claude 在净新增订阅数上开始领先 ChatGPT，但过去几个月情况又反转了，在7月和8月，ChatGPT 的净新增订阅数显著超过了 Claude。

For a big chunk of this year, Claude had started to lead ChatGPT in net new subscriptions, but that's flipped once again in the past couple of months, with ChatGPT significantly outperforming Claude in net new subscriptions in the months of July and August.

在这些大领头羊巩固地位的同时，A16Z 还探讨了初创公司仍然在哪些地方有优势。

And as those big leaders consolidate their position, A16Z also explored where startups still have an advantage.

他们写道，在那些持续吸引关注的初创公司中，有几个主题很突出。

They wrote, among the startups attracting durable attention, a few themes stand out.

一个是拥有真正差异化的模型，最明显的例子是创意类产品，比如音乐生成平台 Suno 和语音生成平台 ElevenLabs。

One is to actually own a differentiated model, with the clearest examples being the creative products like music generation platform Suno and voice generation platform ElevenLabs.

另一个从现有巨头中脱颖而出的方式是提供多模型体验，例子有 Cursor，它支持来自多个实验室的模型，不过我们还要看看随着它们更多被 SpaceX 收购后情况会如何变化。

Another way to stand out from the incumbents is to own a multi-model experience, with examples being Cursor, which supports models from several labs, although we'll see how that changes as they get absorbed more into SpaceX.

还有 OpenRouter，它当然就是专门为了让开发者能从一个平台访问多个提供商而构建的。

And OpenRouter, which is of course specifically built to give developers access to multiple providers from one platform.

初创公司差异化的第三种方式是极度聚焦于有特定需求的某一类用户。

A third way that startups are differentiating is an extreme focus on some audience with a specific need.

他们举的例子有 OpenEvidence，它服务医生；还有 Venice，它面向注重隐私的用户。

The examples they give are OpenEvidence, which serves physicians, and Venice, which is aimed at privacy-focused users.

至于社区的反馈，一个常见的反应是惊讶——要么之前不知道某个产品原来这么火，要么就是根本没听说过。

Now in terms of the community's reaction, one frequent response was to be surprised— to either not have known that something was as popular as it was, or to actually have just never heard of it at all.

Peter James 总结道：“这里面有太多让我惊讶的东西了。”

Peter James sums this up when he writes, there are so many surprises here to me.

Superhuman 排第四？

Superhuman at number 4?

Suno 排第七？

Suno at number 7?

OpenArt 排在 Notion 和 Figma 前面？

OpenArt above Notion and Figma?

还有，Plod 是什么鬼？

Also, WTF is Plod?

我完全被搞糊涂了。

I'm absolutely bamboozled.

我觉得很多人都会有这种反应。

I think a lot of people will have this type of response.

我们每个人都会陷入不同的信息茧房，从而忽略掉某些类别的使用方式。

We all get in different bubbles where entire categories of usage might evade our notice.

Plod 就是个很好的例子——我为了做节目一直在追踪各种东西，所以知道它，但之前真的没怎么在公开场合见过。

Plod is a great example where I have followed it from the standpoint of following everything for this show, But hadn't really come across it in public.

结果这周有位室内设计师来我家，直接就掏出这么一个东西来用，因为那种可以拿着在屋里边走边记笔记的形态，对他们现实中的工作方式来说更合理。

And yet when I had an interior designer at my house this week, they pulled one out and used it because the walk-around-the-house form factor for taking notes made more sense for the way that they work in the real world.

说到 Suno 在月度收入榜上排第七，我之前也聊过，很多人没意识到的是：虽然大家普遍觉得 Suno 的主要付费用户是专业或准专业音乐人，但其实一大部分使用场景是娱乐。

In terms of Suno at number 7 on the monthly revenue charts, as I've discussed before with Suno, one thing that people do not realize is that while many would've assumed that the main paid users of Suno are professional or aspiring professional musicians, a huge amount of the usage is actually entertainment.

人们会做一些好玩的、搞笑的歌给家人朋友听，觉得特别开心，以至于愿意为此花点钱。

People making fun, ridiculous songs for their family and friends and having enough fun that they're actually willing to put some money behind that.

另外有人注意到，很多大主题——比如我们节目里一直在聊的那些——都出现在这些榜单上了。

Another thing that people noticed is that many of the big themes, such as the ones we've been talking about on this show, are showing up on these charts.

像 Replit 和 Lovable 这类 Agentic 编程平台，在收入榜上仍然排得很靠前。

Agentic coding platforms like Replit and Lovable remain near the top of the revenue charts.

OpenRouter 排名第 13，哪怕是在消费类榜单上，也足以说明现在有多少人在追求成本优化。

OpenRouter being at 13, even on a consumer chart, does seem to give indication of how much people are looking for cost optimization.

而像 Noose、Genspark、Manus 这些平台都上榜了，更加凸显了个人 AI 助手领域的竞争正在加剧。

And seeing platforms like Noose, Genspark, and Manus all on the charts reinforces the growing competition around personal AI assistants.

另一个常见的讨论点是：大家对某些体验的粘性感到惊讶。

Another common theme in the discourse is surprise at the stickiness of certain experiences.

比如 Perplexity，在重度 AI 用户圈子里其实已经有一段时间不是主要话题了。

Perplexity, for example, hasn't been a major topic of conversation in the super-enfranchised AI world for some time.

但它现在呢？月度独立访客榜排第 12，月度活跃用户榜排第 25，月度收入榜更是排到了第 9。

And yet there it sits, number 12 on the unique monthly visits chart, number 25 on the monthly active users chart, and all the way up to number 9 on the monthly revenue chart.

N8N 在很多方面可以说是 Agent 时代的先驱，也仍然停留在收入榜上。

N8N, which is in many ways a precursor to the agentic era, remains on the revenue charts as well.

这或许说明，一旦人们投入到一个特定平台，尤其是已经花功夫自动化了某些关键工作流之后，行为改变的速度会相对缓慢。

Perhaps showing the relative slowness of behavior change once people have invested in using a particular type of platform, especially if they took the work to automate some key workflow.

不过，要说最核心的讨论，还是关于重度用户和普通用户的差异。

And yet, if there was one major conversation, it was about the difference between power users and everyone else.

A16Z 的其中一项数据显示：AI 支出最高的前 1% 用户，现在的花费已经超过了后 50% 用户的总和。

One of the things that A16Z showed was that the top 1% of AI spenders now outspend the bottom 50% combined.

前 1% 用户平均每月花 903 美元，而中位数用户每月只花 25 美元。

The top 1% spend an average of $903 a month compared to the median customer who spends just $25.

注意，这不是所有 AI 用户中的前 1%，而是那些实际有花钱的用户中的前 1%。

And remember, that's not the top 1% of all AI users.

而本身付费用户占所有 AI 用户的比例就非常小，这个我们后面会看到。

That's just the top 1% of people who are actually spending, which itself, as we'll see, is a vanishingly small portion of AI users overall.

对我来说，最有趣的图表之一是：这前 1% 的用户在哪些项目上的花费比例远高于其他付费 AI 用户。

One of the most interesting charts to me is the chart that shows what things those 1% spend money on in higher proportion to other paid AI users.

基本上就是那些让他们能搭建和自动化的工具、能创作内容的工具，以及工作中会用到的工具。

It's basically tools that allow them to build and automate, tools that allow them to create content, and tools that you would use for work.

他们购买 Notion 的比例是平均水平的 9 倍以上。

Their purchase rate for Notion is over 9 times as much as the average.

他们购买视频平台（比如 Runway 和 Higgsfield）的比例分别是平均水平的 10 倍和 15 倍。

Their purchase rate for video platforms like Runway and Higgsfield is 10 times and 15 times as much as average respectively.

然后又是 n8n，前 1% 用户付费的比例比所有其他买家高出大约 23 倍。

And there again is n8n, where the top 1% are basically 23 times as likely to be paying for that service as compared to all other buyers.

很多人把这张关于 AI 买家前 1% 集中度的图表，和 a16z 另一份研究报告——他们的《市场状况报告》——里的图表进行了对比。

Now, many people compared that chart about the concentration of the top 1% of AI buyers with this chart from another a16z research piece, their State of Markets report.

这张图表展示的是拥有付费 AI 订阅的美国家庭比例。

This chart is about the share of US households with paid AI subscriptions.

长话短说，根据 PNC Research 的数据，截至今年 4 月，这个数字只有 2.2%。

And TL;DR, as of April of this year, according to PNC Research, That number is only 2.2%.

换句话说，目前有 98% 的美国家庭没有为 AI 付费。

In other words, 98% of US households aren't paying for AI currently.

当然，关键问题是：这 98% 的美国家庭是「还没有」为 AI 付费，还是「永远都不会」为 AI 付费？

The big question, of course, is whether that is 98% of US households aren't paying for AI yet, or 98% of US households aren't going to ever pay for AI.

对有些人来说，这显然全都是机会。

To some, this is obviously all opportunity.

RepView的Adam写道，我觉得这只是反映了早期采用的情况，以及随着大众市场采用率的扩散，实际TAM有多大。

Adam at RepView writes, I think this is just a reflection on early adoption and how much TAM there actually is as mass market adoption proliferates.

Reuben Hassid指出，在美国，从2%到50%的采用率，电脑用了19年，汽车15年，智能手机6.5年。

Reuben Hassid pointed out that the time it took to go from 2% to 50% adoption in the U.S. for the computer was 19 years, for cars was 15 years, for the smartphone was 6.5 years.

换句话说，就是我们还早。

In other words, implying that we're just early.

还有人把它和其他人们愿意付钱的东西做比较。

Some others compared it to what other things people pay for.

Co-founders Nick指出，25%的家庭为SiriusXM付费，55%为云存储付费，91%至少订阅一种流媒体服务。

Co-founders Nick points to 25% of households paying for SiriusXM, 55% paying for cloud storage, and 91% paying for at least one streaming service.

有些人最终很乐观，但更强调我们还早。

Some were ultimately optimistic, but even more assertive about how early we are.

微软的Nicholas Bustamante写道，我的观点是：没人用AI。

Microsoft's Nicholas Bustamante writes, This is my thesis: no one uses AI.

我重复一遍，完全没人。

I repeat, absolutely no one.

我们生活在泡沫里。

We live in a bubble.

即使在我那些付费的朋友中，我让他们打开ChatGPT给我看看他们问什么，也就是那么几个基本问题。

Even among my friends who pay for it, when I ask them to open ChatGPT and show me their queries, it's the same handful of basic things.

大多数人甚至不知道可以上传照片来提问。

Most don't even know they can upload a photo and ask questions about it.

把Gmail连上让AI代理能读写邮件，这让他们大开眼界。

Connecting Gmail so an agent can read and send emails blows their minds.

一个代理打开浏览器帮他们办理登机手续？

An agent opening a browser and checking them into a flight?

他们压根没听过。

They've never even heard of it.

现在最大的挑战是让更多人采用，然后是让已经注册的人真正用上他们付钱买的东西。

The massive challenge right now is adoption, and then getting people who already signed up to actually use what they're paying for.

大多数人完全不知道能做什么。

Most have absolutely zero clue what's possible.

想象一下，如果每个人都像现在顶尖的1%用户那样用AI，算力会有多紧张。

Imagine the compute shortage when everyone starts using AI like the top 1% of users do today.

但也有人说，绝不可能。

Others, though, said, never gonna happen.

律师Basil Masharbash转发了Nicholas并说，我是个能干的成年人。

Lawyer Basil Masharbash reposted Nicholas and said, I'm a competent adult.

为什么我需要AI代理帮我办理登机？

Why do I need an AI agent to check into a flight for me?

我识字啊。

I'm literate.

为什么我需要AI代理帮我读写邮件？

Why do I need an AI agent to read and write emails for me?

没人用AI的原因就是，对绝大多数人来说，没什么实际用处。

The reason no one uses AI is that there's no serious utility here for the vast majority of people.

HabibiSlop也不喜欢这个办理登机的例子。

HabibiSlop also didn't like the flight booking example.

他们写道，这就是为什么坚持办理登机这事儿蠢透了。

They wrote, Here's why doubling down on this flight booking thing is dumb as rocks.

第一，我从来不觉得办理登机要花20分钟。

1, I don't think it has ever taken me 20 minutes to check into a flight.

你说的那些要省掉的步骤根本算不上步骤。

The steps you're describing removing are not even real steps.

他们通常打开航空公司的app，点一下大的红色的办理登机按钮，选一下托运行李数量，搞定。

They usually open the airline's app, click the big red check-in button, click number of checked bags, done.

你还得做同样多的工作来确认代理到底有没有办成功。

You'd have to do the same amount of work to check if the agent actually succeeded at its task.

大约55%的美国成年人一年到头一次飞机都不坐。

About 55% of American adults take zero flights in a given year.

所以真正关心这个使用场景的人很少。

So few people care about this use case.

现在的情况并不是大家低估了AI智能体。

People underestimating AI agents is not what's happening here.

真正的问题是，科技圈的人总把自己那种高频出差、海量邮件、大量行政事务的生活方式，当成全人类的标配。

What's happening is tech people routinely mistake their own unusually high-travel, high-email, high-admin lifestyle for a universal human workload.

过去二十年科技行业最大的问题，就是他们没法跳出自己的小圈子看问题。

The overarching problem of the tech industry the last 20 years is their inability to look outside themselves.

他们沉迷于推销一些只有自己才需要的解决方案。

They've become obsessed with marketing solutions to problems that nobody else has.

有些人甚至对美国普通消费者的看法更悲观。

Some are even more pessimistic about the average US consumer.

Breaking Points的主持人Krystal Ball和Saagar Enjeti就写道，实际上没人需要什么个人AI助手。

With Breaking Points hosts Krystal Ball and Saagar Enjeti writing, nobody actually needs a personal AI assistant.

美国人的平均屏幕使用时间是每天7个小时。

The average screen time in the US is 7 hours a day.

大家有的是时间。

People have plenty of time.

没有任何证据表明，省出更多时间除了让你多刷TikTok还能带来别的什么。

There's no evidence that freeing up more will lead to anything but more TikTok scrolling.

所以说，日常生活反而需要更多离线时的‘摩擦’。

If anything, daily life needs more offline friction.

我会用省下的时间做我喜欢的事！

I'll spend it doing what I love!

你不会的。

No you won't.

你只会刷TikTok。

You'll watch TikTok.

数据不会说谎。

The data doesn't lie.

但另一方面——这也是我为什么把这期节目叫做‘论点对论点’——也有人感叹说，这种看法真是太缺乏想象力了。

But then again, on the flip side— and this is why I called this episode a point-counterpoint— you have folks lamenting what an utter failure of imagination this represents.

律师Basil Musharbash发帖说，他不需要AI智能体帮他办登机、读邮件或写邮件；而投资人Nick Carter回应说，我真的搞不懂这些看法。

Reposting lawyer Basil Musharbash, who said that they don't need an AI agent to check into flights or read or write emails for them, investor Nick Carter says, I find these takes really confusing.

你电脑里住着一个智商150以上、不知疲倦的天才。

You have a 150 IQ+ tireless genius living in your computer.

它可以操控你的浏览器、读取你的邮件、访问你任何想要的文件。

It can use your browser and read your email and access any file you want.

你给它最繁琐的行政任务，它也不会抱怨。

It won't complain if you give it the most tedious administrative tasks.

它愿意一天24小时、一周7天干活，工资还不到最低时薪。

It's willing to work 24/7 for less than minimum wage.

它可以连续几个小时无人监督地工作。

It can work unsupervised for hours at a time.

而你却想不出任何能用上它的地方？

And you can't think of any way to put it to use?

你是完全没有想象力，还是单纯喜欢那些枯燥的行政工作？

Are you completely lacking in imagination, or do you just enjoy mundane administrative tasks?

你喜欢交停车罚单、整理邮件、手动退订垃圾邮件、审核信用卡支付、取消订阅、重新谈判保险？

You enjoy paying parking tickets and triaging emails and manually unsubscribing from spam and vetting credit card payments and canceling subscriptions and renegotiating insurance?

你真的喜欢做这些事？

You like doing that?

对有些人来说，不同类型用户之间的这种分化，几乎就是科技界一条铁律。

For some, this split between different types of users is as close to a law of nature as we have in tech.

就我个人而言，我反感那种认为人人都太懒、只会张着嘴巴刷TikTok、根本不会利用这些工具的观点。

Personally, I bristle at the idea that everyone is too lazy and slack-jawed to get off TikTok to actually take advantage of these tools.

即使效率并不是他们生活的首要目标。

Even if productivity isn't their primary motivation in life.

同时呢，可以肯定的是，我看到那些积极正面使用AI的人，绝大多数都是那种很有冲劲的类型。

And at the same time, it is certainly the case that the vast majority of people that I see interacting with AI in assertive and positive ways are basically go-getter types.

他们是那种希望从生活、工作、事业里得到更多东西的人。

They're people who want more out of life and more out of their jobs and careers and work.

他们不一定都是创业者，但做事情的时候带着那种创业者的劲头。

They're not necessarily all entrepreneurs, but they bring entrepreneurial energy to the things that they do.

总之，他们的抱负很高。

They are, in short, high aspiration.

但并不是每个人都是这样。

And not everyone is.

现在呢，就算我们承认，普通消费者使用AI或者愿意为AI付费的意愿可能始终有限，但有两个理由说明，这不一定会动摇整个行业的商业逻辑。

Now, there are 2 reasons that even if we accept the idea that there might always be limits on how much the average consumer wants to use AI or specifically wants to pay for it, there are a couple of reasons that that doesn't necessarily undermine the financial thesis of the industry.

第一，当然，和之前的软件类型不同，那些有抱负的人以及他们所在的公司，在AI上能花的钱几乎没有上限。

The first is, of course, that unlike previous categories of software, there is far less of an upper bound on how much those aspiring-type individuals and the companies they work with can actually spend on AI.

这不是SaaS，SaaS的上限就是最贵的账户每月100或200美元。

This is not SaaS where the upper bound is the most premium account that costs $100 or $200 a month.

在这个领域，重度用户每月可以花几千甚至几万美元，而且从投资回报率来看完全是划算的。

This is a domain in which power users can spend thousands and thousands or tens of thousands of dollars per month in a totally ROI-positive sort of way.

但回到消费者这边，也没人说消费级AI应用必须都围绕生产力，或者必须靠订阅来赚钱。

But keeping it on the consumer focus, there's nothing that says that consumer AI apps have to be all about productivity, or that they have to monetize strictly through subscriptions.

我再提一下Suno，它现在取得的成功和数亿美元的年经常性收入，不只是靠作为音乐人的生产力工具，也是作为很多普通人的娱乐平台。

I'll point to Suno again, where the success and the hundreds of millions of dollars of ARR that company has now achieved is not exclusively as a productivity tool for musicians, but as an entertainment platform for a lot of normies as well.

然后当然还有另一种商业模式，我们既爱又恨，但如今互联网大部分靠它养活，就是广告。

And then of course there is that alternate business model, the one which we love to hate but which funds most of the internet at this point: advertising.

正好，OpenAI刚宣布了一种新的广告形式，会在ChatGPT里展示可视化广告，不只是文字广告。

Right on time, OpenAI just announced a new ad format where they'll be displaying visual ads inline in ChatGPT, not just text-based ads.

他们写道，我们新的可视化广告形式，通过展示产品灵感、产品使用方式或者它们带来的体验来帮助人们想象产品或服务如何融入生活。

They wrote, our new visual ad format helps people imagine how products and services could fit into their lives through images showing product inspiration, product usage, or the experiences they make possible.

人们可以了解一个新品牌，理解一个优惠的独特之处，然后决定买什么。

People can discover a new brand, understand what makes an offer distinctive, and decide what to buy.

对于广告主来说，这创造了通过图像讲述他们故事的新机会。

For advertisers, this creates new opportunities to tell their story through images.

另外，在那份公告里，他们还重申了一个我在Dev Day上也听说过的数字，就是ChatGPT现在每周触达12亿人，比7月底宣布10亿用户里程碑时增长了20%。

Now, also in that announcement, they reiterated a number that I had heard during Dev Day as well, which is that ChatGPT now reaches 1.2 billion people each week, up 20% since they announced their billion-user milestone at the end of July.

Alex Zimmerman指出，这里还有很大的发展空间。

Alex Zimmerman points out that there is a lot of room to run here.

他写道，ChatGPT在二月份推出广告，六周内跨过1亿美元的年化收入，到八月份达到10亿。

He wrote, ChatGPT launched ads in February, crossed $100 million run-rate revenue in 6 weeks, and a billion by August.

谷歌和Meta加起来有5000亿美元的广告收入。

Google and Meta combine for $500 billion in ad revenue.

AI广告还早得很呢。

We are so early in AI ads.

如果你想理解消费级AI的未来，a16z报告里最有趣的图表可能是这张，消费级AI还有增长空间。

For one who is trying to understand the future of consumer AI, maybe the most interesting chart from the a16z report is this one, where consumer AI has room to grow.

左边是类别。

On the left, they have categories.

中间是Web 1.0和Web 2.0时代的赢家。

In the middle, the winners from Web 1.0 and Web 2.0.

然后右边是，那个特定类别里，有没有AI公司进入消费应用前100名。

And then on the right, whether that particular category has AI companies in the top 100 of consumer apps.

它们非常集中。

They are highly concentrated.

在生产力工具、搜索与问答、内容制作与创造、照片与视频这些类别里有，还有一点教育和健康。

Within productivity tools, search and answers, production and creation, photos and videos, with a little bit of education and health.

而在前100名里没有AI平台的类别包括：流媒体和媒体平台、零售市场、社交网络和消息、旅游、个人理财、招聘和职业网络、房地产、游戏和约会。

Where there are no AI platforms in the top 100 include categories like streaming and media platforms, retail marketplaces, social networks and messaging, travel, personal finance, jobs and career networks, real estate, gaming, and dating.

我觉得，早期采用者类型的人看着不温不火的采用数字，确实有时候太容易说‘我们还太早了’。

I do think it is sometimes too easy for early adopter types to say, we're so early, when they look at underwhelming adoption numbers.

但我确实认为，就消费级AI来说，如果认为今天的状况就是几年后的状况，那可就错了。

But I certainly do think that when it comes to consumer AI, it would be a mistake to assume that where we are today is where we will be a couple years down the line.

今天的人工智能每日简报就到这里。

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

感谢你们的收听或观看，一如既往。

Appreciate you listening or watching as always.

下次再见，祝平安。

Until next time, peace.

已剔除 3 处广告（点击展开查看）

M10:56广告 · 已剔除

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

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

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

M19:48广告 · 已剔除

A new study from KPMG and the University of Texas at Austin found that when people work with AI, similar skills don't guarantee similar outcomes.

Researchers studied more than 500 early-career professionals and found that the best performers consistently amplified the value of AI by guiding, evaluating, and refining its outputs.

These top performers, called AI amplifiers, weren't defined by what they knew alone, but by how they worked with AI.

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

M110:36广告 · 已剔除

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

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

Search Harbor AI Lab Ecosystem ETFs wherever you invest or follow @HarborCapital on X to learn more.

Visit harborcapital.com for a prospectus containing investment objectives, risks, fees, expenses, and other important information.

Read and consider it carefully before investing.

Risks include principal loss and artificial intelligence-related risks.

Harbor ETFs are distributed by Foresight Fund Services LLC.

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

This is a paid advertisement and not personalized investment advice.

Investing involves risk, including possible loss of principal.

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

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

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

And that's where Granola comes in.

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

There's no setup and no configuration.

It just fits into how you already work.

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

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

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

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

Blitzie deeply understands your codebase before it writes code.

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

Vulnerabilities don't live in isolation— they live buried inside millions of lines of interconnected code, where patching one thing quietly breaks three others.

That's why surface-level scans fail.

Blitzie 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.

Each fix is grounded in how your systems connect and validate so nothing new breaks, and the knowledge graph dynamically updates, keeping you ahead of an ever-accelerating threat landscape.

One Blitzie customer resolved 21 active CVEs across 6 core microservices in 4 days.

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

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

Harden your codebase at blitzy.com.

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