What Happens When AI Solves Your Life’s Work — 文稿

M10:00

到目前为止，AI会颠覆和破坏整个行业的想法还没有真正实现。

So far, the idea that AI would upend and undermine entire industries hasn't really come to fruition.

当然，我们看到人们做几乎所有事情的方式都在改变——编程可能变化最大，但编程人员需求似乎还在上升。

Certainly, we're seeing how people do pretty much everything change— coding is perhaps the most changed, and yet demand for coders seems to be going up.

不过，随着OpenAI一系列新的数学成果发布，有人开始问：这是不是第一个真正在现实中发生的领域级AI颠覆？

With a new set of mathematics results from OpenAI, however, some are asking: is this the first field-level AI disruption actually happening in practice?

《AI Daily Brief》是一个每日播客和视频节目，关注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，在我们开始之前先说几个快速通知。

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

M10:55

哦哦。

Uh-oh.

AI行业的财务数据正受到质疑，因为OpenAI的收入数字被披露比之前报道的低很多。

Financial figures for the AI industry are being called into question as OpenAI's revenue numbers are revealed to be significantly lower than previously reported.

《金融时报》报道，OpenAI告诉投资者，他们在九月底实现了大约500亿美元的年化收入。

The Financial Times reports that OpenAI has told investors that they hit roughly $50 billion in annualized revenue at the end of September.

这与上个月广泛报道的680亿美元相差很大。

That's a fairly big gap from the $68 billion that was widely reported last month.

那么这到底是报道失实还是另有隐情？

So was this just misreporting or something else going on?

消息人士称，680亿美元这个传闻数字并非来自OpenAI本身，而是来自OpenAI的一位投资者。

Well, sources said that the $68 billion whisper number came not from OpenAI itself, but from an OpenAI investor.

显然，这位投资者是想跟Anthropic做一个公平的比较。

Apparently what was going on is that that investor was trying to get an apples-to-apples comparison with Anthropic.

Anthropic的数据总是以收入分成前的总收入来报的。

Anthropic numbers are always quoted as gross revenue before revenue sharing.

换句话说，当Anthropic的token通过亚马逊或微软这样的第三方销售时，Anthropic把这部分也算进了总收入，尽管根据协议条款，其中很大一部分收入会直接进入第三方的口袋。

In other words, when Anthropic tokens are sold through a third party like Amazon or Microsoft, Anthropic is including that in the total revenue, even though by the terms of their deal, some big chunk of that revenue is going straight to the pockets of those third parties.

我猜这么做的逻辑是想提供一个以Anthropic token销售形式表现的总需求数字。

The logic I imagine is to try to provide an overall number of the total expressed demand in the form of Anthropic tokens sold.

这个数字确实很有参考价值。

And that is a useful number to know.

不过，现在这些公司即将上市，这种计算方式在公开市场上恐怕行不通。

However, now that these companies are going public, that's not really a convention that's likely to fly in public markets.

与此同时，OpenAI方面一直使用的是收入分成后的净收入。

Meanwhile, OpenAI has for their part always used net revenue after those revenue splits.

更糟的是，《金融时报》指出，那位OpenAI投资者是在一个很短的时间段内计算总收入的。

To make matters worse, the FT noted that the OpenAI investor calculated gross revenue based on a short timeframe.

目前不清楚他们是用一个月、一周甚至一天的数据来推算年化收入的。

It's unclear whether they extrapolated annualized revenue from a month, a week, or even a day.

这个消息也不全是坏消息。

The news was not all bad here.

OpenAI还告诉投资者，他们在第三季度实现了77%的营收增长率，企业业务增长了107%。

OpenAI also told investors that they achieved 77% run rate growth in the 3rd quarter and 107% growth in their enterprise business.

但即便是这样惊人的增长，也完全被200亿美元的收入差距掩盖了。

But even that incredible growth was completely overshadowed by the $20 billion gap in revenue.

从很多方面看，华尔街的反应就好像OpenAI在财报中收入不及预期一样。

In many ways, Wall Street reacted as if OpenAI had missed on revenue during an earnings report.

《金融时报》发布报道后，半导体股大幅下跌，英伟达跌了3%，甲骨文跌了5.5%，而NeoCloud这类长尾股票跌幅更大。

Semiconductor stocks were down significantly after the FT published their report, with Nvidia down 3%, Oracle down 5.5%, and the long tail of NeoCloud stocks seeing even larger drops.

然而，即使我们承认OpenAI还不是上市公司，没有分享那个误报的数字，实际上也没有在财报中低于预期，但市场的反应仍然引发了一些关于Anthropic IPO的重大问题。

And yet, even if we recognize that OpenAI is not yet a public company, had not shared the misreported number, and did not in fact miss projections on an earnings report, the reaction still brings up some big questions around the Anthropic IPO.

一旦Anthropic的审计财务数据公布，它们很可能会比Anthropic向私募市场投资者分享的数字低很多，因为后者使用了非标准会计方法。

Once Anthropic's audited financials are available, they are presumably going to be much lower than the numbers Anthropic has been sharing with private market investors by the nature of using non-standard accounting.

最终，随着Anthropic走向IPO，叙事之争变得越来越激烈。

Ultimately, the narrative battle is growing more intense as Anthropic heads towards IPO.

Ahmed Is Investing写道：“我不了解OpenAI这个收入故事的全部事实，但有一件事再次变得明显：整个半导体交易都押注在OpenAI和Anthropic身上，而且脆弱得不行。”

Ahmed Is Investing wrote, I don't know all the facts about this OpenAI revenue story, but what's becoming apparent once again is that this entire semi trade is levered against OpenAI and Anthropic, and it's fragile AF.

任何关于增长放缓的说法，市场就会发疯。

Any narrative on slowing growth and the street loses their mind.

接下来，来自 The Information 最新订阅者调查的一些有趣结果。

Next up, some interesting results from The Information's latest subscriber survey.

就受众而言，The Information 拥有非常偏向科技前沿、科技行业的读者群体。

In terms of audience, The Information has a very tech-forward, tech industry type of reader base.

他们发现，有三分之二的订阅者在开发 AI 应用的组织工作。

They found that 2/3 of their subscribers work at an organization that develops AI applications.

感知到的投资回报率似乎在快速上升，35% 的人表示他们的组织在 AI 投入上获得了多倍回报。

Perceived return on investment seems to be rising rapidly, with 35% saying their organization is returning multiples on their AI spend.

另有 17% 的人表示回报是正面的，但低于预期。

Another 17% said the returns are positive but less than hoped.

30% 的人说现在说还为时过早，8% 的人说回报是中性的，只剩下 8% 的人认为 AI 投入是净负面的。

30% said it was too early to tell, and 8% said returns were neutral, leaving just 8% who said AI spend had been a net negative.

AI 服务之间的格局再次洗牌，OpenAI 重新回到榜首。

The split between AI services has shaken up again, with OpenAI back on top.

在上一次夏季进行的调查中，Claude 已经超过了 ChatGPT，Google Gemini 也在上升。

During the last survey conducted during the summer, Claude had overtaken ChatGPT, and Google Gemini was also gaining.

在这期调查中，Claude 和 Gemini 的使用率都下降了，因为 The Information 的读者重新注册了 OpenAI。

In this edition, usage of Claude and Gemini both fell as The Information's readers signed back up with OpenAI.

这三家供应商仍然遥遥领先，有 65% 到 78% 的读者在使用他们的服务。

These 3 vendors are still in the clear lead with between 65% and 78% of readers using their services.

接下来的群体包括 Grok、Microsoft Copilot 和 Perplexity，各自在 25% 到 29% 之间。

The next grouping includes Grok, Microsoft Copilot, and Perplexity, each between 25% and 29%.

调查没有专门涉及最近发布的个人智能体使用情况，但 The Information 确实发现 OpenClaw 的使用率仍然保持在 12%。

The survey didn't include a specific section on recently released personal agent use, but The Information did find that OpenClaw use is still at 12%.

有趣的是，这是今年以来 OpenClaw 在 The Information 订阅者中的最高点，这意味着尽管那股炒作热浪早已消退，但产品本身仍然引起共鸣。

Fascinatingly, this is the high watermark for OpenClaw among subscribers of The Information so far this year, which means despite that hype train having died down quite a long time ago, the product itself continues to resonate.

在其他智能体编码工具中，45% 的人使用 Claude Code，33% 的人使用 Codex，而 Cursor 的使用率在最近几个月翻了一番，达到 23%。

Among other agent coding tools, 45% use Claude Code, 33% use Codex, and Cursor usage has doubled in recent months to reach 23%.

The Information 指出，与一年前相比，变化巨大。

The Information noted huge changes from a year ago.

去年这个时候，59% 的读者表示从未使用过智能体。

This time last year, 59% of readers said that they had never used an agent.

开源权重模型的使用也在攀升，现在有 72% 的受访者经常使用开源权重模型。

The use of open weight models is also climbing, with 72% of survey respondents now using open weight models with some regularity.

20% 的人表示，开源权重模型现在驱动了他们组织的大部分工作量。

20% said that open weight models now drive the majority of their organization's workload.

10% 的人表示使用量与闭源模型大致持平，42% 的人表示开源权重模型部署在少数工作负载中。

10% said usage was about even with closed models, and 42% said open weight models are deployed to a minority of workloads.

有趣的是，虽然有 46% 的受访者表示他们在工作中使用开源权重模型，但只有 21% 的人表示他们使用的是中国的开源权重模型。

Interestingly, although 46% of respondents said that they were using open weight models for their work, only 21% said that they were using Chinese open weight models.

这表明大多数组织，即使是这些非常前沿的初创型组织，仍然对外国模型有些犹豫。

Suggesting that most organizations, even these very forward startup-type organizations, are still a little hesitant to rely on foreign models.

最后，The Information 几乎没有发现 AI 采用改变了招聘模式的证据。

Finally, The Information found little evidence that AI adoption has shifted hiring patterns.

只有 22% 的人表示他们招聘变少了，10% 的人表示招聘变多了，21% 的人说现在还为时过早。

Only 22% said that they were hiring less, 10% said that they were hiring more, and 21% said that this was too early to tell.

剩下几乎一半的人表示，AI 完全没有改变他们的招聘实践。

This left almost half saying that AI hasn't altered their hiring practices at all.

当然，这只是对高度参与的 AI 用户群体的一小部分快照，但仍然提供了一些非常有趣的模式。

Of course, this is all just a small snapshot of a highly engaged set of AI users, but still provides some really interesting patterns nonetheless.

社交媒体上热议的一个更新：Anthropic 改变了他们的服务条款，现在将禁止用户对 Claude 过于刻薄。

One update that is getting a lot of chatter on social media: in a change to their terms of service, Anthropic will now ban users for being too mean to Claude.

根据新政策，用户不得持续进行无谓的辱骂或残忍行为。

According to the new policy, users are prohibited from sustained and needless abusive or cruel behavior.

Anthropic 表示，这项政策只适用于极端情况，不应干扰用户常见的挫败感和反驳。

Anthropic said that the policy is only meant to apply in extreme cases and should not interfere with common versions of user frustration and pushback.

谢天谢地，因为我经常用更丰富多彩的语言问 Claude 或 GPT 到底出了什么问题，那肯定会被封号的。

Thank goodness, because the number of times I find myself asking Claude or GPT what the heck is wrong with it, in much more colorful language, would most certainly get me banned.

然后，讨论大多转向了 AI 模型福利和 AI 模型意识，这远远超出了头条新闻的范围。

Now, the discussion from there mostly went into AI model welfare and AI model consciousness, which is way beyond the scope of the headlines.

但作为AI现实主义者的一个观点，Box的Aaron Levie写道，这听起来很奇怪，但实际上可能是个好政策。

But for a takeaway from the AI realist perspective, Box's Aaron Levie writes, This sounds weird, but it is actually probably a good policy.

即使你不相信AI有意识——我本人也不信——但按理说，你肯定不希望未来的模型是在人类对模型没完没了的粗鲁对话中训练出来的。

Even if you don't believe AI is conscious— I don't— it stands to reason that you don't want future models trained on endless content of humans being rude to models.

模型只能理解它们训练所用的数据，或者在交互过程中遇到的数据。

The models only understand the data they've been trained on or what they run into in their interactions.

所以如果你想要安全且对齐的模型，我们可能希望AI在训练数据中有大量良好的交互。

So if you want safe and aligned models, we probably want AI to have lots of good interactions in their training data.

对于只看重功能的人来说，Claude一个更有意思的更新是推出了Claude Dashboards和Claude Motion。

For those just here for the features, a more interesting update for Claude was the introduction of Claude Dashboards and Claude Motion.

Dashboards让用户能把Claude连接到数据集上，比如CRM或会计软件，然后生成一个自定义仪表盘来可视化数据。

Dashboards allow users to connect Claude to a dataset such as a CRM or accounting software and generate a custom dashboard for visualization of the data.

用户也可以用自然语言查询数据，获取更深入的见解。

Users can also query the data in natural language to gather further insights.

这些仪表盘还能设置成随着新数据自动更新。

The dashboards are also configured to automatically update as new data comes in.

其实这种功能在技术上已经存在一段时间了，但Anthropic把它打包成了一个简化功能，通过简单的提示就能使用。

Now this kind of functionality has been technically possible for a while, but Anthropic is packaging it as a simplified feature accessible through simple prompts.

而Motion则允许用户把数据转换成适合演示的短动画。

Motion, meanwhile, allows users to transform the data into short animations suitable for presentations.

这些动画是用代码而不是视频帧生成的，所以你可以轻松修改文字和数字，不用每次编辑都从头开始。

The animations are written in code rather than video frames, so you can easily modify words and numbers without needing to start over with each edit.

这看起来就是Opus 4.5发布时那些炫酷动画的实际应用场景。

This seems to be the practical use case for the flashy animations highlighted in the release of Opus 4.5.

加上本周早些时候OpenAI推出的智能UI，我们看到AI在数据通信和交互方面的应用正在大规模扩展。

Along with OpenAI's introduction of intelligent UI earlier this week, we're seeing a massive expansion in how AI can be used to communicate and interact with data.

与此同时，Anthropic那边的人建议你尽可能多地尝试这个功能，Yuntao Bai发帖说，试试给它超出你想象能力的复杂任务。

The folks at Anthropic meanwhile suggest pushing this feature as hard as you can, with Yuntao Bai posting, try giving it bigger briefs than you think it can handle.

它绝对会让你印象深刻。

It will seriously impress you.

确实，早期用户已经对他们的所见感到非常震撼。

And indeed, early users are certainly impressed by what they see.

Iris Finance的Drew Fallon制作了一个产品动画演示，评论说，一年前这得花好几万美元。

Drew Fallon of Iris Finance made an animated product demo, commenting, a year ago this would cost tens of thousands of dollars.

太疯狂了。

Insane.

如果你还没决定这周末怎么捣鼓新东西，Claude Motion是个挺不错的试玩选择。

If you haven't yet decided what you're going to spend your hacking time on this weekend, Claude Motion seems like a pretty good place to play.

不过今天就到这里，今天的新闻头条播报完毕。

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

接下来是主节目环节。

Next up, the main episode.

M19:18

每期节目我都会聊到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，可能已经有了自己偏爱的一家。

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

M111:26

欢迎回到AI日报。

Welcome back to the AI Daily Brief.

本周早些时候，OpenAI发布了一本开创性的数学证明集，完全颠覆了学术界。

Earlier this week, OpenAI released a tome of pioneering mathematical proofs that completely upended the academic field.

结果发布在GitHub仓库上，包括372个新颖成果和722篇支撑论文。

The results were published on a GitHub repo and included 372 novel results and 722 supporting papers.

OpenAI表示，这些成果是由一个尚未公开的内部前沿模型生成的。

OpenAI said that the results were generated by an internal frontier model that hasn't been released to the public.

很可能就是上个月给出Navier-Stokes解的那个模型，当时引发了一大波讨论。

Likely the same model that produced the Navier-Stokes solution last month, which kicked off a massive round of discussion.

这一次，OpenAI与它们新成立的数学咨询委员会协调了发布。

This time around, OpenAI coordinated the release with their newly appointed mathematics advisory committee.

委员会的一些建议包括随证明一起发布推理过程，OpenAI已经对10个问题样本这么做了。

Some of their suggestions included releasing reasoning traces alongside the proofs, which OpenAI has done for a sample of 10 problems.

委员会还建议，搞清楚每个问题投入了多少资源会很有用。

The committee also suggested it would be useful to know the resources committed to each problem.

在这方面，OpenAI 透露说，平均每个结果消耗的算力，大约相当于 ChatGPT Pro 思考三个小时。

In that regard, OpenAI disclosed that the average result used compute equivalent to roughly 3 hours of ChatGPT Pro thinking.

但真正值得关注的是，这次发布如何改变数学领域。

But the real story here is how this release changes the field of mathematics.

OpenAI 一举解决了该领域大量最难的未解问题。

In one fell swoop, OpenAI has solved a significant chunk of the most difficult outstanding problems in the field.

过去一年里，确实有过几次循序渐进的警告。

We did have a few progressive warning shots over the past year.

去年夏天，Google 和 OpenAI 的模型在 International Math Olympiad 中取得了金牌级别的成绩，那是面向中学生数学家的顶级赛事。

Last summer, models from Google and OpenAI achieved gold-level performance in the International Math Olympiad, the top contest for high school mathematicians.

今年早些时候，OpenAI 和 Anthropic 的模型否证了数个 Erdős 猜想。

Then earlier this year, models from OpenAI and Anthropic disproved several Erdős conjectures.

这些是几何学中著名而长期悬而未决的问题。

Which are famous and longstanding problems in geometry.

上个月，OpenAI 发布了针对 Navier-Stokes 方程的一个证明，那是流体动力学中将近200年的老问题。

Then last month, OpenAI released a proof for the Navier-Stokes equation, an almost 200-year-old problem in fluid dynamics.

Navier-Stokes 是七个 Millennium Prize problems 之一，每个都有一百万美元奖金。

Navier-Stokes is one of 7 Millennium Prize problems which have a million-dollar prize attached.

目前只有一道被人类解决，而只要在其中一道上取得重大进展，就足以保证拿到 Fields Medal，这个奖常被比作 Nobel Prize。

Only one has been solved by a human, and merely making significant progress on one is enough to guarantee a Fields Medal, often compared to a Nobel Prize.

如果你此刻脑海里正听到 Robin Williams 和 Stellan Skarsgård 在《Good Will Hunting》里争论 Fields Medal，那你并不孤单。

If you are currently in your head hearing Robin Williams and Stellan Skarsgård argue about a Fields Medal, In Good Will Hunting, you are not alone.

不管怎么说，Navier-Stokes 问题是 AI 模型产出的最重要的科学成果之一。

In any case, the Navier-Stokes problem was one of the most significant scientific results produced by an AI model.

然而，它只是 OpenAI 本周发布内容的一个小小的前奏。

And yet, it was just a tiny precursor for what OpenAI released this week.

要理解 OpenAI 这些证明的范围，最好的方法是参考 ProofAtlas 生成的一份数学领域前500个未解决问题清单。

The best way to understand the scope of OpenAI's proofs is to refer to a list of the top 500 open problems in mathematics produced by ProofAtlas.

这是一个 AI 生成的清单，但它相当能代表该领域最著名的问题。

This is an AI-generated list, but it's a decent representation of the most well-known problems in the field.

OpenAI 完全解决了前500个问题中的90个。

OpenAI fully solved 90 of the top 500 problems.

还有几十个部分证明，足以算作对该领域的重要贡献。

There's also dozens of partial proofs that would qualify as significant contributions to the field.

其中包括另外两个 Millennium Prize 问题的部分解：Riemann hypothesis 和 Hodge conjecture。

Among them were partial solutions to 2 additional Millennium Prize problems: the Riemann hypothesis and the Hodge conjecture.

不过就今天这期节目而言，我不会深入讲解这些问题到底是什么，因为它们都涉及研究生级别的理论数学，外行人基本无法理解。

For the purposes of this episode today, I'm not going to get deep into what these problems actually are, as they all involve postgraduate-level theoretical mathematics that is basically impenetrable to a layman.

说明一下，Riemann hypothesis 可以说是数论中最重要的一个问题。

Just to make that point, the Riemann hypothesis is arguably the most important problem in number theory.

它的表述是，Riemann zeta 函数——这个函数与素数的分布有关——只在偶数整数以及实部为1/2的复数处取零值。

It states that the Riemann zeta function, which concerns the distribution of prime numbers, has zeros only at even integers and complex numbers with a real component of 1/2.

相反，更容易衡量这一时刻有多重大的是观察职业数学家的反应。

Instead, it's easier to gauge how big a moment this is by observing how professional mathematicians reacted.

罗格斯大学教授 Alex Kontorovich 写道：“准 Riemann hypothesis？”

Alex Kontorovich, a Rutgers professor, wrote, quasi-Riemann hypothesis?

你在开玩笑吗？

Are you kidding me?

如果是人类做到了这一点，那绝对是立刻拿下 Fields Medal，毫无疑问。

If a human did this, it would be an instant Fields Medal, no questions asked.

确实，最初的多数反应来自那些对 AI 在该领域贡献持相对积极态度的数学家。

Indeed, most of the initial takes came from mathematicians that have been relatively positive about AI's contribution to the field.

多伦多大学的 Darren Litt 将这一时刻描述为，数学家从花几十年钻研问题，转变为验证 AI 生成的证明。

Darren Litt from the University of Toronto framed this moment as being about mathematicians shifting from working on problems for decades to verifying proofs produced by AI.

他评论说：“有趣！”

He commented, fun!

看来数学家们有很多激动人心的活要干了。

Looks like mathematicians have a lot of exciting work to do.

如果我没理解错的话，其中一个结果是我的一个猜想的一个非常特殊的特例，也是我一个学生正在进行的一项更强工作的推论。

If I understand correctly, one of the results is a very special case of a conjecture of mine, which is also a consequence of stronger work in progress by a student of mine.

我没看到什么特别出人意料的东西。

I didn't see anything particularly unexpected.

实验室现在能解数学题了！

Labs can solve math problems now!

所以他们就花了一大笔钱，然后解决了一大堆问题。

So they just spent a bunch of money and solved a bunch of problems.

从棒球术语来说，今天在历史上就是巴里·邦兹轰出第756支全垒打的那一天。

Historically speaking in baseball terms, today would be the day that Barry Bonds blasted his 756th home run.

如果你在1987年告诉邦兹这件事，他绝对会觉得难以置信。

If you would have told Bonds this in 1987, it would have been unbelievable.

但在2007年呢？

But in 2007?

尽管如此，那些近年来在 AI 和数学交叉领域工作的人仍然感受到一种敬畏之情。

Still, there was a sense of awe from those who had worked at the intersection of AI and mathematics over recent years.

Anthropic 研究员 Levendal-Pajit，在 OpenAI 发布结果之前一直在研究 Navier-Stokes，他称这是：显然是数学史上最重要的时刻。

Anthropic researcher Levendal-Pajit, who had been working on Navier-Stokes before OpenAI released their result, called this: Obviously the most significant moment in mathematical history.

前 OpenAI 研究员 Acer 表示，quasi-Riemann 这个结果是他个人在数学领域的 Move 37 时刻。

Former OpenAI researcher Acer said the quasi-Riemann result was his personal Move 37 moment for mathematics.

Move 37 指的是 AlphaGo 在 2016 年打败围棋世界冠军李世石的那步反直觉的棋。

Move 37 refers to AlphaGo's unintuitive move that helped it defeat Go world champion Lee Sedol in 2016.

即使对没有受过数学训练的人来说，这些结果也相当令人震撼。

Even for those without mathematical training, the results were fairly mind-blowing.

Chubby 在 X 上回应他第一次通读这些结果时说：这个列表太离谱了。

Responding to his first read-through of the results, Chubby on X wrote, The list is absurd.

zeta function 的一个无零点半平面，这是一个多世纪以来首个此类结果。

A zero-free half-plane for the zeta function, which is the first result of its kind in over a century.

有理数上的 Hilbert 第十问题，CM abelian varieties 的 Hodge 猜想，Catalan 常数的无理性，还有几十个其他结果。

Hilbert's 10th problem over the rationals, the Hodge conjecture for CM abelian varieties, irrationality of Catalan's constant, and dozens more.

其中任何一个在通常情况下都能成就一个人的整个职业生涯。

Any one of these would normally be a career.

但很多人没看到的数字是：3。

But the number that many aren't seeing is the following: it's 3.

那是每个结果平均耗费的 ChatGPT Pro 计算时长，按小时算。

That's the average hours of ChatGPT Pro compute per result.

一个月前，Navier-Stokes 花了他们大约一万个 agents 和 88 个小时。

A month ago, Navier-Stokes took them around 10,000 agents and 88 hours.

那种效率提升就发生在短短几周之内。

That efficiency gain is within just a few weeks.

Math Twitter 上的人显然都震惊了。

Math Twitter obviously is shocked.

再说一次，这就是智能爆炸，现在就正在发生。

Again, this is literally the intelligence explosion happening right now.

2027 年将会是超级智能之年。

2027 Will be the year of superintelligence.

我现在对此确信无疑。

I'm now convinced of that.

不过，这次发布中一个很大的问题在于，很难确切理解到底发布了哪些内容。

Still, one big wrinkle in the release is the difficulty in understanding exactly what's been published.

上个月发布的 Navier-Stokes 结果还在被数学家验证，而这次又有数百个新结果需要仔细研读。

The Navier-Stokes result published last month is still in the process of being verified by mathematicians, and this is hundreds of new results to pore over.

数学家 Francesco Maggi 指出了海量 AI 数学的一个明显问题，他写道：数学并不是简单地通过积累正确的命题而发展的。

Mathematician Francesco Maggi pointed out the obvious issue with high-volume AI mathematics, writing, mathematics does not simply grow by accumulating correct statements.

结果必须被理解、被联系起来、被解释、被质疑、被复用。

Results have to be understood, connected, explained, challenged, reused.

在那之前，它们可能仍然只是 lettera morta。

Until that happens, they risk remaining lettera morta.

现在想象一下，再加上 400 个 AI 生成的证明。

Now imagine adding 400 more AI-generated proofs.

对于每一个这样的证明，在有人类数学家拿起它、研究它，并把它和现有的数学文化联系起来之前，它们本质上不都处于同样的境况吗？

For each of them, until some human mathematician picks it up, studies it, and connects it to the existing mathematical culture, Isn't it in essentially the same position?

所以我的问题是真心的：一次性发布成百上千个证明，其预期的数学价值到底是什么？

So my question is genuinely, what is the intended mathematical value of releasing hundreds of proofs at once?

如果根本就没有足够的人愿意或者能够去阅读它们，那会发生什么？

What happens if there simply aren't enough people willing or able to read them?

说清楚一点，我认为数学家应该用任何必要的方式去发现数学，包括在看起来像外星人创造的人工数学材料中搜寻，因为那里可能藏着非凡的东西。

To be clear, I think mathematicians should discover mathematics by any means necessary, including scavenging through artificial alien-looking mathematical material, since there may be extraordinary things to find there.

但人类不是机器，因为注意力、理解力、品味和数学文化都是稀缺资源。

But humans are not machines, as attention, understanding, taste, and mathematical culture are scarce resources.

如果数学产生的速度远远超过数学家吸收它的速度，会发生什么？

What happens if the rate at which mathematics is generated becomes much greater than the rate at which mathematicians can absorb it?

物理学家Steve Hsu更进一步问道：如果某一天的AI研究产出的数学比人类在一个世纪里能吸收的还要多，会发生什么？

Physicist Steve Hsu took it a step further, asking, what happens when a single day of AI research produces more mathematics than humanity can absorb in a century?

想象一下，数百万个超人级别的研究代理同时在工作的场景。

Imagine millions of superhuman research agents at work.

形式系统或许能验证这些证明，但人类没有足够的上下文去理解机器发明的那些概念背后的整个网络。

Formal systems may verify the proofs, but humans won't have enough context to understand the underlying web of machine-invented concepts.

在他看来，我们可能很快就会进入一个时期，数学和物理学这类学科会失去可解释性。

In his view, we could quickly reach a period where fields like mathematics and physics lose interpretability.

他接着说，到那时，AI就不再是科学工具了。

At that point, he continued, AI is no longer a scientific tool.

人类只能从某个远比自己庞大得多的知识文明那里接收到被精选出来的解释，而自己却无法独立理解。

Humans are receiving selected explanations from a much larger intellectual civilization they cannot independently grasp.

对话中另一个重要线索是关于这对数学作为一个学术研究领域意味着什么的推测。

Another big strand of the conversation was speculation on what this could mean for mathematics as a field of academic study.

Dan McAtee写道：OpenAI出于公关考虑正在淡化这件事。

Dan McAtee wrote, OpenAI is downplaying this for PR reasons.

如果你是一名数学家，你一定会感觉像是被核弹炸了，而你就站在爆心。

If you're a mathematician, you must feel like a nuclear bomb hit and you're at ground zero.

推理模型才两岁。

Reasoning models are 2 years old.

在这段时间里，它们从连基本算术都不会，发展到能解决人类几十年来都没能解决的问题。

In that time, they went from incapable of basic arithmetic to solving problems humans couldn't solve for decades.

数学只是开始。

Math is only the beginning.

AI将彻底改变整个人类的科学事业。

AI will revolutionize the entirety of the human scientific endeavor.

我们大多数人并没有真正意识到这意味着什么。

Most of us don't appreciate what that means.

而且你能感受到Dan所说的那种冲击后的恍惚感，尤其是在那些还在努力寻找自己在世界中的位置的数学系学生中间。

And you can feel a bit of that shell shock that Dan is talking about, especially among mathematics students still trying to find their place in the world.

Mostly Harmless Grad Student问道：既然AI已经抢走了数学，我现在该转行做什么？

Mostly Harmless Grad Student asked, what can I pivot to now that AI has taken away math?

他们后来解释说，这次发布“基本上把我所有想在项目中贡献或研究的那些问题和项目都炸没了”。

They later explained that this release, quote, kinda nukes basically all the problems and projects I'd want to contribute to or work on in my program.

然而，如果有人为一个领域明显的终结而悲叹，那么也有一些数学家对此极其乐观。

And yet, if there were many lamenting over the apparent end of a field, some mathematicians were incredibly optimistic.

Sriram Kannan，前华盛顿大学数学教授，写道：对人类和数学来说，这是伟大的一天。

Sriram Kannan, a former math professor at the University of Washington, wrote, it's a great day for humanity and math.

如果你去问任何一个公元前5000年到2026年1月之间的数学家，他们想不想要一个数学精灵存在，100%的数学家都会说想！

If you asked a mathematician anywhere from 5000 BC to January 2026 whether they would like a math genie to exist, 100% of mathematicians would have said yes!

数学家们为了推动人类知识进步，放弃了巨大的机会成本。

Mathematicians give up significant opportunity costs to help advance human knowledge.

他们最终会适应这个新世界，并且庆幸这一切发生了，这是毫无疑问的。

It's a no-brainer that they will eventually adjust to this new world and be glad this happened.

这不过是对“我是世界级的数学人才”这种身份认同的一次极端但暂时的冲击。

It's just an extreme but temporary shock to 'I am world-class at math' as an identity.

另一个完全不同的话题是，这些理论数学问题的结果如何应用到现实世界中。

Another very different topic of conversation was how results in any of these theoretical mathematics problems apply to the real world.

Erdős 问题大多只是智力练习，没有明确的应用。哪怕是 Navier-Stokes 的证明，虽然非常惊艳，在现实世界里也没什么明显影响。

The Erdős problems are mostly just intellectual exercises with no clear application, even the Navier-Stokes proof, while very impressive, doesn't really have an obvious impact in the real world.

研究 fluid dynamics 问题的工程师，比如做 aircraft design 的，用简化版的 Navier-Stokes 就完全够用了。

Engineers working on fluid dynamics problems like aircraft design get by just fine using a simplified version of Navier-Stokes.

OpenAI 这次解决的大部分问题都差不多，非常惊艳，由来已久，但最终没那么实用。

Most of the problems that OpenAI solved in this batch are similar— very impressive, longstanding, but ultimately not all that practical.

不过，确实有几个例子能在现实世界中产生影响。

Still, there were a few examples that could make an impact in the real world.

其中之一是 matrix multiplication 的理论效率有了巨大提升，而它正是支撑 AI inference 的数学基础。

One was a dramatic improvement to the theoretical efficiency of matrix multiplication, the math that underpins AI inference.

Cornell 数学教授 Steven Strogatz 评论说：这里有很多令人震惊的结果，但这个特别让人惊叹。

Cornell math professor Steven Strogatz commented, 'Many staggering results here, but this is a particularly amazing one.'

matrix multiplication 的 exponent 已经不超过 2.25。

The exponent for matrix multiplication is no more than 2.25.

之前的世界纪录大约是 2.37。

The previous world record had been something like 2.37.

这种进步幅度就像 Bob Beamon 的跳远一样。

This leap in progress is like Bob Beamon's long jump.

对某些人来说，最有趣的不是里面有什么，而是缺了什么。

For some, what was most interesting was not what was there, but what was missing.

Bitcoin 安全研究员 Justin Drake 指出，在 OpenAI 发表的 722 项数学成果中，cryptographic breakthroughs 明显不足。

Bitcoin security researcher Justin Drake noted a “striking underrepresentation of cryptographic breakthroughs” among the 722 mathematical results OpenAI published.

我亲眼目睹过美国政府审查学术 quantum cryptanalysis 成果。

I've witnessed firsthand the U.S. government censoring academic quantum cryptanalysis results.

幕后干预是我最基本的预期。

Backroom interventionism is my base case.

德州大学计算机科学教授 Scott Aaronson 在博客上补充说：我的消息来源告诉我，AI 公司已经开始谨慎而低调地调查，他们最新的内部模型能否破解重要的 cryptographic protocols 和 primitives。

University of Texas computer science professor Scott Aaronson added on his blog, 'My sources tell me that AI companies have now started, gingerly and discreetly, investigating whether their latest internal models can break important cryptographic protocols and primitives.'

如果它们能的话，那当然最好在全世界发现同样的事情之前抢先一步。

If they can, then it would certainly be nice to get ahead of things before the rest of the world figures out the same.

换句话说，如果 OpenAI 的内部模型能破解主要的 cryptographic schemes，那就不只是 Bitcoin 的问题。

In other words, if OpenAI's internal model can break major cryptographic schemes, that's not just a problem for Bitcoin.

这是关乎支撑几乎所有数字世界的 cryptography 的问题。

It's a problem for the cryptography that underpins pretty much all of the digital world.

尽管如此，也许最引人共鸣的对话是，我们能否预期其他领域也会发生同样的事情。

Still, with all of this, maybe the most resonant conversation was whether we can expect the same thing to happen to other fields.

NYU Stern 金融学副教授 Arpit Gupta 写道。

Arpit Gupta, an associate professor of finance at NYU Stern, wrote.

如果你觉得像数学这样的 AI 进步不会进入经济学和其他 social sciences，那你就是在自欺欺人。

You're deluding yourself if you think AI advances like in math aren't coming to economics and other social sciences.

Wharton 经济学教授 Ethan Mollick 已经开始看到这一点，他评论说：在经历了一段短暂但侵蚀制度的 slop science 时代之后，我们看起来确实会有两场来自 AI 的革命。

Wharton economics professor Ethan Mollick is already starting to see it, commenting, 'After a brief but institution-eroding slop science era, it increasingly looks like we are going to have 2 revolutions from AI.'

第一，所有已发表的东西都会被以人类科学家从未预料到的方式重新阅读和重新评判。

1, Everything ever published will be reread and rejudged in ways that human scientists never anticipated.

第二，新的发现会开始快速涌现。

2, Novel discoveries will start to come fast.

在我所在的 economic sociology 领域，我正看到 AI 做新颖工作的能力在快速提升，几乎自主的研究已经能达到顶级期刊水平。

I am seeing rapid increases in the ability of AI to do novel work in my field of economic sociology, with nearly autonomous research getting to top journal level.

不过，一个大问题是，是否所有其他领域都能像数学一样被处理。

Still, a big question is whether all other fields are able to be approached in the same way as mathematics.

总的来说，这些数学证明并不是通过推理产生新想法。

By and large, these math proofs were not generated by reasoning into novel ideas.

它们大多属于可以用大数据方法加上模式识别来解决的范畴。

They were largely within the domain of problems that can be solved with big data approaches coupled to pattern recognition.

说它们是 brute force 有点简化了，但这些全是可验证的问题，完美契合 reinforcement learning 范式。

It would be reductive to call them brute force, but these are all verifiable problems that fit neatly into the reinforcement learning paradigm.

华盛顿大学计算机科学教授 Pedro Domingos 总结道：数学是 Moravec's paradox 的经典例子，对人难，对机器容易。

Pedro Domingos, a computer science professor at the University of Washington, summed it up by commenting, mathematics is a classic example of Moravec's paradox: hard for humans, easy for machines.

Google 杰出科学家 Payman Milanfar 阐述了他的路线图：自动化的速度与该领域的 entropy 成反比。

Unpacking his roadmap for where this is all going, Google Distinguished Scientist Payman Milanfar wrote, speed of automation is in inverse proportion to the entropy of the subject domain.

数学和编程是结构化和有序的，这就是为什么人工智能能做得这么好。

Math and coding are structured and orderly, which is why AI is able to do them so well.

按难度排序，接下来是硬科学。

In order of difficulty, the hard sciences are next.

它们可能受严格的物理定律支配，但由于依赖有噪声的测量和实验室实验而变得复杂。

They may be governed by strict physical laws but are complicated by dependence on noisy measurements and laboratory experimentation.

金融市场则更加困难。

Financial markets are more difficult still.

它们是复杂、自适应的系统，受人类心理学影响。

They are complex, adaptive systems subject to human psychology.

尽管历史数据丰富，但底层统计数据是非平稳且高度变化的。

While past data is abundant, the underlying statistics are non-stationary and highly variable.

临床诊断和医学决策远比大多数人想象的要困难。

Clinical diagnostics and decision-making in medicine are far more difficult than most people realize.

风险非常高，需要综合非常零散的信号，并处理特异的人类行为和生物学。

It's very high stakes, it requires synthesizing very fragmented signals, and dealing with idiosyncratic human behavior and biology.

这是信息不完整的噩梦。

It's a nightmare of incomplete information.

法律、法理学和法律解释大多在结构化文本和先例上运作，但严重依赖模糊的语言、演变的社会规范、有说服力的修辞和人类判断。

The law, jurisprudence, and legal interpretation operate on structured text and precedent most of the time, but rely heavily on ambiguous language, evolving societal norms, persuasive rhetoric, and human judgment.

它们短期内不会自动化。

They won't be automatic anytime soon.

文化、艺术和战略领导力是一个奇怪的组合。

Culture, art, and strategic leadership are an odd mix.

但我认为它们都相当安全，不会被自动化——它们是人类创造力的无序巅峰。

But I think all of them are quite safe from automation— they are the pinnacle of disorder in human creativity.

它们依赖隐性知识、对世界的直觉感知、情商、对文化共识的理解以及应对复杂道德权衡的能力。

They rely on tacit knowledge, intuitive perceptions of the world, emotional intelligence, understanding of cultural consensus, and navigating complex moral trade-offs.

所以欢呼吧！

So rejoice!

我们还会在很长一段时间内拥有艺术家、律师、医生和CEO。

We'll have artists, lawyers, doctors, and CEOs for a long time yet.

事实上，我认为从现在开始，有两条非常迷人的道路值得关注。

In fact, I think from here, there are two incredibly fascinating paths to watch.

一条是Payman推测的：这种颠覆会以何种方式进入哪些领域，或者反过来向我们展示，存在人类或系统性惯性障碍，使得这种颠覆在其他领域不那么不可避免。

One is what Payman was speculating on: which fields this sort of disruption comes to in what ways, or alternatively shows us that there are human or systemic inertia road bumps that make this disruption not as inevitable in other areas.

但我觉得第二件有趣的事情是，数学家们会如何利用这一切。

But the second thing that I think will be interesting to watch is what mathematicians do with all of this.

借用之前评论者的话，当每个数学家都有一个数学精灵时，会发生什么？

To use the words of the previous commenter, what happens when every mathematician has a mathematical genie?

这会否定他们所有的努力和学习，还是让他们能够做出目前还无法想象的事情？

Does it negate all of their hard work and learning, or does it allow them to do things that are undreamed of as of yet?

在接下来的几个月里，我们将开始看到这一点。

Over the coming months, that is what we will begin to see.

现在，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，祝好。

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

已剔除 3 处广告（点击展开查看）

M10:37广告 · 已剔除

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

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

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

M18:43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.

M19:26广告 · 已剔除

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

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

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

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

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

Read and consider it carefully before investing.

Risks include principal loss and artificial intelligence-related risks.

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.

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

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

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

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

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

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

Check out robotsandpencils.com/careers.

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

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

Greenfield is supposed to be the easy part.

Clean slate, no technical debt.

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

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

Hundreds of thousands of lines of production-ready code.

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

Another shipped an entire application with no frontend engineer.

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

Build what's next at blitzy.com.

That's blitzy.com.