Claude 形态的科学

Claude-shaped science

Anthropic · 2026-10-01 · Anthropic Research ↗

打开互动全文版(逐段中英对照 + 针对本篇提问)→

摘要 · Abstract

本文描述了作者如何从把大语言模型视为潜在的人类科学家,转向设计适合其实际优势的问题,他将这一方法称为寻找“Claude 形态的问题”。核心论点是:科学家的期望与当前大语言模型最擅长的事情之间存在阻抗失配,因为这些模型虽然才华横溢,却并非自主科学家,需要大量人工引导。为弥合这一差距,作者构建了 BootLoops——一个用于精确定量计算的开源工具框架,并用它通过半数值 S-矩阵 bootstrap 计算了此前未解决的椭圆费曼积分。由于类似的数学结构在不同领域反复出现,Claude 识别出了与生态学、群体遗传学、系统发育学及其他领域的联系,但作者发现这些结果往往在技术上正确,却在科学上平平无奇,直到领域专家将工作引导到其领域真正关心的问题上。结论是:研究者应停止与模型对抗,转而让问题匹配其能力,同时依靠专家指导来判断科学意义。

This article describes how the author moved from treating large language models as would-be human scientists to designing problems suited to their actual strengths, an approach he calls finding Claude-shaped problems. The core argument is that an impedance mismatch exists between what scientists want and what current LLMs do best, since the models are brilliant but not autonomous scientists and require substantial hand-holding. To close this gap, the author built BootLoops, an open-source harness for exact quantitative calculations, and used it to compute previously unsolved elliptic Feynman integrals through the semi-numerical S-matrix bootstrap. Because similar mathematical structures recur across disparate fields, Claude identified connections to ecology, population genetics, phylogenetics, and other areas, though the author found these results were often technically correct yet scientifically unremarkable until domain experts steered the work toward questions their fields care about. The conclusion is that researchers should stop fighting the models and instead match problems to their capabilities, while relying on expert guidance to judge scientific significance.

核心贡献 · Key contributions

局限 · Limitations

章节 · Sections(共 5)

全文 · Full text(逐段中英对照)

01 Claude,掌舵! 01 Claude, take the wheel!

_在这篇客座文章中,Matthew Schwartz 教授再次撰文,介绍一种加速科学发现的人工智能新方法。在《Vibe Physics》一文中,Schwartz 讨论了 Claude 与物理学研究生在能力上的相似之处。本文中,他描述了当他不再与 Claude 对抗,而是让 Claude 去寻找“Claude 形状”的问题——即最适合当前一代大型语言模型工具能力的问题——时所发生的情况。这促使他构建了 BootLoops,一个用于定量科学中精确计算的工具包。由于类似的计算常常出现在截然不同的科学领域中,Claude 发现了与生态学、群体遗传学以及其他十几个领域的联系。这些联系在技术上往往是正确的,但起初在科学上并不引人注目,因此 Schwartz 与领域专家合作,引导 BootLoops 转向那些领域所关心的问题。下面,我们将分享更多关于这些项目以及 BootLoops 是如何诞生的内容。_

_In this guest post, Prof. Matthew Schwartz returns to describe a new approach to AI-accelerated science. In Vibe Physics, Schwartz discussed similarities in capability between Claude and a physics graduate student. Here, he describes what happened when he stopped fighting Claude and allowed Claude to find “Claude-shaped” problems: ones best suited to the capabilities of the current generation of LLM tools. This led him to build BootLoops, a toolkit for exact calculations in quantitative science. Because similar calculations often turn up across very disparate areas of science, Claude found connections to ecology, population genetics, and a dozen other fields. These connections were often technically correct but scientifically unremarkable at first, so Schwartz worked with domain experts to steer BootLoops toward questions those fields care about. Below, we share more about these projects and how BootLoops came about._

智能体式 AI 正在迅速进步。每个人都注意到模型似乎更聪明了:它们知道得更多,犯的错误更少,想法也更好。如果你关注趋势线,很容易自信地谈论 AI 彻底改变科学的潜力。然而,如今试图在自己的工作中使用这些模型的学术科学家常常感到一种脱节。这些模型或许正在解决具有挑战性的长期问题,但到目前为止,这些大多只是现有技术的范围明确的应用。许多头条新闻似乎都出现在数学领域,这是科学中唯一可以完整陈述问题并绝对检验答案的部分。但大多数科学并非如此。对于我们许多研究人员和学生来说,那些头条新闻与我们使用这些模型时实际发生的情况之间的差距,会让我们感到沮丧和焦虑。

Agentic AI is improving rapidly. Everyone notices the models seem smarter: they know more, make fewer mistakes, and have better ideas. If you follow the trend lines, it is easy to speak confidently about the potential of AI to revolutionize science. However, academic scientists trying to use the models today in their own work often feel a disconnect. The models may be solving challenging and longstanding problems, but so far these have mostly been well-scoped applications of existing techniques. Many of the headlines seem to be in mathematics, the one part of science where a problem can be stated completely and an answer checked absolutely. But most of science is not like that. And for many of us researchers and students, the distance between those headlines and what happens when we use these models can leave us frustrated and anxious.

在我看来,核心冲突在于:尽管这些模型非常出色,但像人类科学家那样工作并不是当前 LLM 最擅长的事情。Claude 和 GPT 擅长科学,但它们不是 _科学家_:是的,它们很聪明,但可能需要大量的手把手指导才能让它们产生任何具有科学价值的东西。物理学家会称之为“阻抗失配”:两个系统各自运行良好,但匹配得很差,因此一个系统输入的大部分内容永远无法传递到另一个系统。在这里,失配存在于科学家想要的东西和 AI 擅长的事情之间。

The core conflict, as I see it, is that although these models are brilliant, working like a human scientist is not what current LLMs do best. Claude and GPT are good at science, but they are not _scientists_: yes, they are smart, but it can take a lot of hand-holding to get them to produce anything of scientific value. Physicists would call this an “impedance mismatch”: two systems that each work fine but are poorly matched, so most of what one puts in never gets through to the other. Here, the mismatch is between what scientists want and what AI does well.

我开始寻找阻抗失配不那么严重的例子。我首先让 Claude 构建一套易于使用的数学物理工具。不久之后,这些工具就在其他问题上找到了用途。这个迭代过程产生了一套软件和科学协议,我称之为 BootLoops。BootLoops 充当 LLM 的一种 harness,就像 Claude Code 或 Claude Science 是 Claude 的 harness,或者 Codex 是 GPT 的 harness 一样。我发现 BootLoops 特别适合科学中的一类定量问题。它也是开源的,因此可以与任何你喜欢的模型一起使用。

I started looking for examples where the impedance mismatches are less acute. I began by having Claude build an accessible suite of tools for mathematical physics. Before long, the tools found uses for other problems. This iterative process generated a set of software and scientific protocols, which I call BootLoops. BootLoops functions as a kind of harness for the LLM, much like Claude Code or Claude Science is a harness for Claude, or Codex is a harness for GPT. I’ve found BootLoops especially well-suited for a class of quantitative problems in science. It is also open-source, so it can be used with whatever model you like.

一旦 Claude 有了 BootLoops,它就不停地注意到同样的模式:许多领域都有一些问题,如果任何人知道数学、物理或计算机科学中的某项技术存在,就能直接解决。我开始称这些为“Claude 形状”的问题,并跟随它们走出高能理论物理——我的老本行——进入地质学、生物学、经济学和语言学。在这些其他领域,我无法依靠自己的专业知识来判断 Claude 发现的东西是否有趣。所以我找到了一些专家并请教他们。在他们的指导下,BootLoops 能够在许多研究领域取得实质性进展。

Once Claude had BootLoops, it kept noticing the same pattern: many fields have problems that a technique from mathematics, physics, or computer science would solve outright if anyone knew it existed. I started calling these “Claude-shaped” problems, and followed them outside high-energy theoretical physics—my home turf—into geology, biology, economics, and linguistics. In these other areas, I could not rely on my own expertise to know whether what Claude found was interesting. So I found some experts and asked. With their guidance, BootLoops was able to make substantive advances in many research areas.

下面,我将分享 BootLoops 的由来,描述我在应用它的领域中所取得的一些早期发现,并分享我对当今如何解决人类科学家与 AI 科学家之间阻抗失配问题的一些思考。

Below, I share how BootLoops came about, describe some early findings in areas where I have been applying it, and share some of my thinking on how to resolve the impedance mismatch between human and AI scientists today.

Claude,接管方向盘! Claude, take the wheel!

去年十二月,我尝试把 Claude 当作研究助理来用,发现 Claude Opus 4.5 的表现就像一位很强的研究生,速度却是其 20 倍。尽管 Claude 在实验结束时产出了一篇高质量的论文,但走到那一步却十分艰难。它写的每一句话我都得修改,把它从无关的思路上引开,再把它从死胡同里拉回来。

Last December, I tried using Claude as a research assistant, and found that Claude Opus 4.5 performed like a strong graduate student at 20 times the speed. Despite Claude producing a high-quality paper at the end of the experiment, it was a slog to get there. I had to correct every sentence it wrote, steer it away from irrelevant threads, and pull it back from dead ends.

今年夏天,我尝试了另一种做法:不再把 Claude 当作我希望它成为的合作者,而是开始把它当作它实际上的合作者。这要求我去寻找适合其长处的问题。目前,Claude 还无法在深层的概念性问题上帮助我——但它在所有领域都拥有几乎无限广博的知识、惊人的编程能力、数学与统计学的前沿知识,以及以机器速度解析论文、附录和数据的能力。

This summer, I tried doing something different: instead of treating Claude like the collaborator I wanted it to be, I started to treat it like the collaborator it actually is. This required looking for problems suited to its strengths. Right now, Claude is just not able to help me with deep conceptual questions—but it does have a virtually unlimited breadth of knowledge across all domains, incredible coding skills, leading-edge knowledge of mathematics and statistics, and the ability to parse papers, appendices, and data at machine speed.

寻找“Claude 型”问题的一个自然起点,是那些编程能帮上忙的领域。当 Anthropic 于 2026 年夏天发布 Claude Fable 5 时,我想看看它的网络能力能否迁移到科学计算中。于是,我试图通过让它移植、编写并改进我几篇论文及散射振幅相关文献中的各种方法来测试它。

A natural place to start looking for Claude-shaped problems was in areas where coding could help. When Anthropic released Claude Fable 5 in Summer 2026, I wanted to see whether its cyber capabilities would translate to scientific computing. So I sought to test it by having it port, code up, and improve various methods from a handful of my papers and the adjacent literature on scattering amplitudes.

散射振幅是我们解读大型强子对撞机数据的方式:以 13 万亿电子伏特对撞两个质子,振幅便是碎片与碰撞所产生的任何新粒子(希格斯玻色子或某种未知粒子)之间的理论桥梁。其核心是费曼图,即某种特定形式的多维积分。我们目前正在艰难处理的那些积分,每一个都可能相当于一篇博士论文,或让一个研究组耗费数年。

Scattering amplitudes are how we interpret data from the Large Hadron Collider: smash two protons at 13 trillion electron volts, and the amplitude is the theoretical bridge between the debris and whatever new particle, a Higgs boson or something unknown, the collision produced. At their core are Feynman diagrams, multidimensional integrals of a particular form. The ones we are struggling with now can each be a PhD thesis, or occupy a group for years.

过去 20 年里,一种替代方案发展了起来:S-矩阵 bootstrap。在 bootstrap 方法中,你不是硬算出积分,而是施加物理约束,直到只剩下一个可能的答案。知道振幅在何处为无穷(即其“奇点”)也许能把它缩小到 20,000 个选项;一个对称性再把它削减到 500 个;如此一路缩减到一个。传统的 bootstrap 是纯解析的,并在对称性最高的理论中走得最远——在那里约束能一路推进到唯一选项(N=4 超对称杨–米尔斯理论中的九圈振幅便是一例)。而在更接近真实世界的情形中,你往往在到达终点之前就用尽了约束。一个更新的转向——半数值 bootstrap——在你还能以荒谬的精度(有时达 1,000 位数字)计算振幅在少数几个点上的取值时,便填补了这一缺口:如果剩下的选项足够少,这些数值就能精确地确定其余系数。

Over the past 20 years an alternative has grown up: the S-matrix bootstrap. In the bootstrap approach, instead of grinding out the integral, you impose physical constraints until only one answer is possible. Knowing where the amplitude is infinite (its “singularities”) might narrow it to 20,000 options; a symmetry cuts that to 500; and so on down to one. The traditional bootstrap is purely analytic and has gone furthest in the most symmetric theories, where the constraints reach all the way to a single option (the nine-loop amplitude in N=4 super-Yang–Mills is an example). Closer to the real world you often run out of constraints before the end. A newer pivot, the semi-numerical bootstrap, closes the gap when you can also compute the amplitude at a handful of points to absurd precision (sometimes 1,000 digits): if few enough options remain, those numbers pin down the remaining coefficients exactly.

半数值 bootstrap 似乎非常适合智能体式 AI。它涉及数学、物理和计算机科学,而这些领域没有任何一个人能完全掌握;它需要大量的编码和算法开发;而且它是可检验的,因为同样的数值方法让任何人——无论是否为专家——都能通过运行两个脚本,将最终答案与积分对照验证到任意精度。该领域的专业知识分布也很不均衡:好的想法散落在 Wolfram Language、C++、Python 或 Julia 中,还有更多想法存在于完全没有代码的论文里。因此,我给 Fable 5 的第一个任务就是将所有内容移植到一个通用框架中,并补写论文从未提供的代码。

The semi-numerical bootstrap seemed ideal for agentic AI. It draws on mathematics, physics, and computer science that no one person has mastered; it needs a great deal of coding and algorithm development; and it is checkable, since the same numerics let anyone, expert or not, verify the final answer against the integral to as many digits as they like by running two scripts. The community's expertise is also unevenly distributed: good ideas sit in Wolfram Language, C++, Python, or Julia, and many more sit in papers with no code at all. So my first assignment for Fable 5 was to port all of it to a common framework, and to write the code the papers never provided.

Claude 毫不费力地完成了这件事。令我惊讶的是,它在大约 20 分钟内就复现了我论文中的结果,而我自己为此编写的代码花了好几周。不过,当它告诉我我的做法效率极低、并且存在一种我不知道的更好算法时,我并不感到意外。

Claude did this effortlessly. I was surprised when it reproduced the results from my paper in around 20 minutes, while the code I wrote to do it took me weeks. However, I was not surprised when it informed me that I was doing something very inefficiently and that there was a better algorithm I was unaware of.

接着,我让 Claude 去寻找它可以计算的未解决振幅。事实证明,对 S-矩阵 bootstrap 来说足够简单的问题,对人类来说也足够简单——而且事实上,大多数问题都已经被解决了。尽管如此,它还是找到了几个未解决的问题。在与模型讨论之后,我清楚地发现 Claude 把自己限制在具有最简单函数族的振幅上:对数。于是我问道,它能否对次简单的函数族——椭圆函数——做同样的事情?

Then, I asked Claude to search for unsolved amplitudes it could compute. It turns out that problems that are simple enough for the S-matrix bootstrap are also simple enough for humans to do—and, indeed, most have been done. It nevertheless found a few unsolved problems. After some discussion with the model, it became clear that Claude was limiting itself to amplitudes with the simplest family of functions: logarithms. So I asked, could it do the same thing for the next simplest family, elliptic functions?

椭圆积分真的很难,即使对我这样花大量时间计算积分的人来说也是如此。迄今为止,只有少数几个椭圆费曼积分被计算出来,而且据我所知(或 Claude 所知),没有一个是通过 bootstrap 完全计算出来的。问题不在于方法行不通,而在于没有人尝试过,因为所需的专业知识分散在许多人类身上。相比之下,Claude 轻松地将其为对数情形移植和构建的所有机制推广到了这些其他积分类别。这一次,大部分软件是它自己编写的,或者是从数学而非物理学中借用的。随着工具包的扩展,它开始一个接一个地攻克积分。很快,我们就有了 30 个通过 BootLoop 端到端完成的积分,其中包括用这种新方法复现的 15 个已知结果和 15 个此前从未被计算过的结果。

Elliptic integrals are really hard, even for people like me who spend a lot of their time computing integrals. Only a handful of elliptic Feynman integrals have ever been computed, and none completely by the bootstrap, at least to my or Claude's knowledge. The issue is not that the methods wouldn't work, but rather that nobody had tried, since the expertise needed to do so is distributed among many humans. Claude, by contrast, easily generalized all of the machinery it had ported and built for the logarithmic case to these other integral classes. This time it wrote most of the software itself, or borrowed it from mathematics rather than physics. As the toolkit grew, it started to land one integral after another. Soon we had 30 integrals BootLooped from end to end, comprising 15 reproductions of known results by this new method and 15 that had never before been computed.

这一切仅仅发生在几周之后。起初,我以为只要能就此写一篇报告就会心满意足,但我实在太想看看我们(也就是我和使用 BootLoops 的 Claude)还能做些什么。

That was all after only a few weeks. Initially, I thought I would be satisfied just to do a write-up on that, but I was too tempted to see what else we (that is, me and Claude with BootLoops) could do.

“我懂功夫” “I know Kung Fu”

科学的一个偶然特征是,相同的方程常常在不同背景下反复出现。例如,在物理学中,扩散方程、Fokker-Planck 方程和 Schrödinger 方程都具有相同的数学形式,因此如果你开发出一种方法来解决一个问题,你通常可以将其应用于许多其他问题。我知道 BootLoops 擅长的计算在其他领域也有相关性:例如在宇宙学和弦理论中。我不知道但 Claude 乐于告诉我的是,这些积分也可以映射到群体遗传学中的贝叶斯证据积分,或者用于 Feynman 积分约化的有限域方法也可以应用于进化生物学中的问题。

A serendipitous feature of science is that the same equations often appear over and over again in different contexts. In physics, for example, the diffusion equation, Fokker-Planck equation, and Schrödinger equation all have the same mathematical form, so if you develop a method to solve one problem, you can often apply it to many others. I knew that computations BootLoops was good at were relevant elsewhere: in cosmology and string theory, for example. What I didn’t know, but Claude was happy to tell me, was that these integrals could also map onto Bayesian evidence integrals in population genetics, or that the finite-field methods used for Feynman integral reduction could also apply to problems in evolutionary biology.

这开启了一个特别丰硕的时期,寻找并解决 Claude 形状的(更狭义地说,BootLoops 形状的)问题。有些想法立即显现为伟大的应用。例如,系统发育学研究如何利用 DNA 将物种组织成家谱树;相关的计算是贝叶斯证据积分,其输出是一个单一数字,表示候选树对观察到的 DNA 的解释程度,这与 BootLoops 最初构建时所做的积分类型相同。随着想法的出现,我坚持要求 Claude 既使用旧工具又构建新工具,以便工具集能够增长。每个工具,即使来自失败的项目,也打开了新的大门。Claude 变得越来越有能力。就像《黑客帝国》中的 Neo 在从功夫下载中醒来后一样。

This kicked off a particularly fertile period of searching for and solving Claude-shaped (and, more narrowly, BootLoops-shaped) problems. Some ideas were immediately obvious as great applications. For example, the science of phylogenetics studies how to organize species into family trees using DNA; the relevant computation is a Bayesian evidence integral, whose output is a single number saying how well a candidate tree accounts for the observed DNA, an integral of the same kind BootLoops was originally built to do. As the ideas came, I insisted that Claude both use old tools _and_ build new ones, so that the harness would grow. Each tool, even from a failed project, opened up new doors. Claude was getting more and more capable. It was like Neo in _The Matrix_ after waking up from the Kung Fu download.

然而,随着项目偏离我的专业舒适区,我开始担心。当 Claude 声称它在我领域所做的某件事是极好的时,我可以判断那是真是假(通常不是)。但当它声称它在另一个领域所做的某件事是极好的时,我发现自己会同意。我的怀疑加剧了,我知道我需要请一些专家来确认。确实,我发现几乎在所有情况下,Claude 在技术上是正确的,但结果并不那么有趣,直到专家帮助指导我们。

As the projects drifted from my professional comfort zone, however, I started to worry. When Claude claims something it did in my field is fantastic, I can judge whether that’s true or not (it often isn’t). But when it claims something it did in another field is fantastic, I find myself agreeing. My suspicion heightened, I knew I needed to bring in some experts to be sure. Indeed, I found that in almost all cases, Claude was technically correct, but the result was not all that interesting until the expert helped steer us.

其中一个项目是关于生态学中的中性生物多样性理论。在任何生态系统中,一些物种繁荣而另一些死亡。中性理论询问这种生命史中有多少是由于随机机会。在早期工作的基础上,2001 年,生态学家 Stephen Hubbell 挑衅性地提出,也许这一切都是随机的。2005 年,生态学家 Rampal Etienne 为 Hubbell 的理论提出了一个方程,使其(至少在原则上)能够以精确的方式被检验。不幸的是,20 年来没有人能够大规模求解该方程。Claude 认识到 Etienne 方程是 BootLoops 形状的,并解决了它。将计算应用于数据,我们确定在地球上生态学研究最多的森林——巴拿马运河的 Barro Colorado 岛——树种混合变化的速度比中性理论允许的快 4.5 倍。

One such project was on neutral biodiversity theory in ecology. In any ecosystem, some species thrive while others die off. Neutral theory asks how much of that life history is due to random chance. Building on earlier work, in 2001, the ecologist Stephen Hubbell suggested, provocatively, that perhaps it was all random. In 2005, the ecologist Rampal Etienne put an equation to Hubbell’s theory allowing it (at least in principle) to be tested in a precise way. Unfortunately, for 20 years nobody could solve the equation at scale. Claude recognized Etienne’s equation as BootLoops-shaped, and solved it. Applying the computation to data, we established that in ecology’s most studied forest on earth, Barro Colorado Island in the Panama Canal, the mix of tree species changes 4.5 times faster than neutral theory allows.

对这个结果感到兴奋,我将其带给 James O’Dwyer,一位植物生物学教授和中性理论专家。James 对我很有耐心。尽管他对这一技术壮举印象深刻,但他说结果可能会“被许多生态学家耸耸肩”。生态学家已经更定性地观察到,中性理论无法跟上真实的森林。但 James 有一个更好的主意:减去中性预测并研究剩余部分。这将使我们更好地了解哪些变化可归因于自然选择、竞争以及物种之间的差异。

Excited by this result, I brought it to James O’Dwyer, a professor in plant biology and an expert in neutral theory. James was patient with me. Though he was impressed by the technical feat, he said the results would likely “be met with a shrug by many ecologists.” Ecologists had already observed, more qualitatively, that neutral theory can’t keep up with real forests. But James had a better idea: subtract the neutral prediction and study the remainder. This would give us a better sense of what changes were attributable to natural selection, competition, and the differences between species.

James 和我接着与 Claude 紧密合作,打造一个后继模型。James 有物理学背景,因此学科间的语言障碍更容易克服,但像许多科学家一样,他尚未真正体会到智能体式 AI 的力量。随着合作深入,我成了某种 Claude 处理者,将 Claude 语翻译给 James,并让模型保持正轨,而 James 则推动模型产出生态学家可能重视的东西。最终成果是我们三人都引以为豪的:一个与数据高度吻合的生命史最小预测模型。我们目前正在将模型从巴拿马扩展到全球其他森林样地,使用 Claude 帮助整理的数据集。

James and I proceeded to work intensively with Claude to craft a successor model. James has a physics background, so the disciplinary language barrier was more easily surmountable, but, like many scientists, he had not yet really appreciated the power of agentic AI. As the collaboration progressed, I became a sort of Claude handler, translating Claude-speak to James and keeping the model on track, while James pushed the model to produce something ecologists might value. The final result is something all three of us are proud of: a minimal predictive model of life histories in excellent agreement with data. We are currently extending the model from Panama to other global forest plots, using datasets that Claude has helped curate.

另一个项目涉及群体遗传学中的精确计算,该领域研究基因在群体内如何变异及其原因。Claude 首先使用从数学物理引入的方法,解决了一个 30 年之久的积分表达式,该表达式描述自然选择如何塑造稀有突变。我们将其应用于 gnomAD,这是最大的公开人类遗传变异目录。Claude 对此结果非常兴奋,但我并不那么确定。我不得不写信给三位不同的生物学家寻求验证,才有一位回复,但最终我成功招募到同事 Michael Desai,他从事该领域工作,尽管不是这个具体问题。与 James 一样,Michael 对技术结果印象深刻,但并未被其科学意义说服。

Another project involved looking at exact calculations in population genetics, a field that studies how genes vary within a population and why. Claude first used methods imported from mathematical physics to solve a 30-year-old integral expression for how natural selection shapes rare mutations. We applied it to gnomAD, the largest public catalog of human genetic variation. Claude was very excited about this result, but I wasn’t so sure. I had to write to three different biologists for validation before one responded, but at last I was able to recruit my colleague Michael Desai, who works in the field, though not on this exact problem. As with James, Michael was impressed by the technical result but not compelled by the science.

然而,Michael 指出,更有影响力的结果可能来自研究单条染色体上突变对之间的相关性,使用相同或类似的方法。在此范围内,Claude 发现并构建了新工具,比如用于数学计算机辅助证明的不等式证书,并将其加入 BootLoops 工具包。然后,我们分析了 1000 Genomes 项目中基因组里 57 亿对邻近突变,并发现了名为基因转换的机制的证据。这是一个重要发现,因为几乎所有使用连锁遗传变异的分析,从群体历史到疾病定位,都忽略了基因转换。更大的教训是,AI 可以帮助发现隐藏在巨大基因组数据集中的生物学,公共档案中有成千上万个这样的数据集。

Michael noted, however, that more impactful results might come from studying correlations between pairs of mutations on a single chromosome, using the same or similar methodology. Within this scope, Claude found and built new tools, like the inequality certificates used for computer-assisted proofs in mathematics, and added them to the BootLoops kit. We then analyzed 5.7 billion pairs of nearby mutations in genomes from the 1000 Genomes Project, and found evidence for a mechanism called gene conversion. This is an important finding because nearly every analysis that uses linked genetic variation, from population history to disease mapping, ignores gene conversion. The bigger lesson is that AI can help find biology hidden in enormous genomic datasets, of which there are thousands sitting in public archives.

其他项目也有类似的弧线:Claude 提出一个我觉得有说服力的初步发现。我将其带给专家,专家不为所动但看到了潜力,然后我们一起将其塑造成有意义的科学。有趣的是,其中一些从 BootLoops 式的定量计算转向更一般的 Claude 式工作。例如:

Other projects had a similar arc: Claude comes up with an initial finding I find compelling. I bring it to an expert, who is unmoved but sees the potential, and together we sculpt it into meaningful science. Interestingly, some of these drifted from BootLoops-shaped quantitative calculations to more general Claude-shaped work. For example:

* **经济学**。经济学期刊现在要求作者提供复制包,编辑通常负责检查。这是细致的手工工作。我们与两位经济学家合作,构建了一个 AI 数据编辑器:它将五家领先期刊的 4,452 篇论文的复制包从 MATLAB、Stata 和其他商业工具移植到开源代码(约 30,000 个例程),并对照已发表的表格检查了基本上每一个可验证的数字。结果出现在 NBER 工作论文中。

* **Economics**. Economics journals now ask authors for a replication package, and an editor is often tasked with checking it. That is careful, manual work. Working with two economists, we built an AI data editor: it ported the packages of 4,452 papers in five leading journals from MATLAB, Stata, and other commercial tools to open-source code (some 30,000 routines) and checked essentially every validatable number against the published tables. The results appear in an NBER working paper.

* **语言学**。词重音目录通常由手工汇编,仅涵盖几百种语言,且存在选择效应和人为偏好。Claude 搜索了所有公开可访问的来源,并与三位语言学家合作,我们构建了 AccStack:一个涵盖 6,072 种语言的词重音数据库,几乎每个条目都引用了决定性段落,并附有 160,000 篇音系学著作的参考文献。

* **Linguistics**. Word-stress catalogs are typically assembled by hand, with a few hundred languages, selection effects, and human preferences. Claude scoured every openly accessible source and, working with three linguists, we produced AccStack: a database of word stress covering 6,072 languages, quoting the deciding passage for nearly every entry, plus a bibliography of 160,000 phonology works.

其他一些亮点,每个亮点都是与专家合作完成的,并且每个都在进行进一步的探索和验证,包括:

Some additional highlights, each of which was done in collaboration with experts, and each of which is undergoing further exploration and verification, include:

* **系统发育学**。我们使进化树的贝叶斯证据计算快速且可精确检查,并发现单个基因在竞争树之间的关联差异通常小于标准采样程序的误差。

* **Phylogenetics.** We made the Bayesian evidence for evolutionary trees fast to compute and exactly checkable, and found that single genes are often tied between competing trees by less than the error of standard sampling programs.

* **地球科学**。我们结合了大气科学、地球化学和气候建模的方程和数据,构建了一个定量、预测性的模型,描述大氧化事件如何通过四次冰期展开。

* **Earth science**. We combined equations and data from atmospheric science, geochemistry, and climate modeling to build a quantitative, predictive model of how the Great Oxidation Event unfolded through four glaciations.

* **基因组学**。我们大规模研究了单细胞 RNA 计数的统计特性,以表征爆发速率和偏离教科书式基因表达电报模型的情况。

* **Genomics**. We studied the statistics of single-cell RNA counts at scale to characterize bursting rates and the departure from the textbook telegraph model of gene expression.

* **太阳黑子。** 我们采用人类人口统计学的方法推断了太阳黑子的生命周期,随后将该方法推广到星斑的表征上——星斑是许多系外行星搜寻中的一个混杂因素。

* **Sunspots.** We deduced the lifecycle of sunspots using the methods of human population demographics, then extended the methodology to characterize starspots, a confounding factor in many exoplanet searches.

* **数学物理。** 我们解决了 Watson 的“最终问题”:具有三个不相等跳跃速率的三维随机游走的精确返回概率,这是 George Watson 自 1939 年起研究的一系列晶格积分中的最后一个。

* **Mathematical physics.** We solved Watson's "final problem": the exact return probability of a 3D random walk with three unequal hopping rates, the last of the lattice integrals George Watson began in 1939.

* **宇宙学。** 我们构建了一套完整的理论工具包,用于将宇宙学参数拟合到宇宙大尺度结构的数据上,其中包括完整的双圈功率谱和单圈三谱。

* **Cosmology.** We built a complete theoretical toolkit to fit cosmological parameters to data on the universe's large-scale structure, including the full two-loop power spectrum and the one-loop trispectrum.

* **统计学。** 混合模型以行为不良而著称;我们推导出了对其贝叶斯证据的精确近似,使模型选择变得实用,应用范围从生物统计学延伸到文本分析。

* **Statistics.** Mixture models are notoriously ill-behaved; we derived precise approximations to their Bayesian evidence that make model selection practical, with applications from biostatistics to textual analysis.

总体而言,这种对 BootLoops 进行迭代改进、寻找 Claude 形态的应用、并与专家一起推进到科学前沿的方法,成效惊人。有关这些项目及其他项目的更多信息,可在 bootloops.ai 上找到。

Overall, this approach of iterative improvement of BootLoops, searching for Claude-shaped applications, and pushing to the scientific frontier with experts has been astonishingly productive. More information about these projects and others can be found on bootloops.ai.

技术细节 The technical details

同时推进多个项目(在约 400 个候选问题中,三个月内于 18 个领域完成 36 篇手稿,涉及 19 位合作者)需要大量协调工作,无论对我还是对 Claude 都是如此。让这一切变得可控的配置相当简单。所有 Claude Code 会话都运行在 Google Cloud 虚拟机上的终端中,并根据合作者的偏好链接到不同的 GitHub 和 Overleaf 仓库。我为每个项目设置独立的会话,另有一个主会话负责协调其他会话、分配算力并验证结果。每个会话都包含在后台运行计算的智能体,中间结果以 markdown 文件形式存储在各项目文件夹中。这些子智能体很有用,因为我经常触发 Fable 5 的分类器;这些拦截不会破坏我的整个会话,而只会关闭单个智能体。我还设有独立会话,分别用于撰写结果;创建 GitHub 仓库、工具手册和 BootLoops 网站;编写并验证所有独立的工具代码;以及以对抗性审稿人的身份反复核查结果。

Running many projects at once (36 manuscripts in 18 fields with 19 coauthors over three months, out of some 400 candidate problems) takes a lot of coordination, both for me and for Claude. The setup that makes this manageable is fairly simple. The Claude Code sessions all run in terminals on Google Cloud virtual machines, linked to various GitHub and Overleaf repos depending on my collaborators' preferences. I have separate sessions for each project, plus a master session that coordinates the others, allocates compute, and validates results. Each session involves agents that run the computations in the background, with intermediate results stored in markdown files in their respective folders. These subagents are helpful because I regularly run into Fable 5's classifiers; instead of these blocks corrupting my whole session, they just shut down a single agent. I also have separate sessions for writing results; creating the GitHub repos, tool manuals, and the BootLoops website; writing and validating all the separate tool codes; and checking and rechecking results as an adversarial referee.

尽管如此,这些会话仍需要大量引导。例如,Claude 没有时间概念,而且喜欢夸张。在一个项目结束时,它告诉我:“Matt——公式已解。两年的攻关,四次深度推进,……”而它实际上只工作了三天。当我询问预计完成时间(ETA)时,估计值总是要么太长要么太短。我试过告诉 Claude 在运行测试确认之前不要给出估计,但效果也不好。随着时间推移,由于无法依赖 Claude 告诉我,我开始自己摸索事情大概需要多长时间。

Still, the sessions require a lot of guidance. For example, Claude has no sense of time, and loves to be dramatic. At the end of one project, it told me "Matt — the formula is solved. Two years of campaign, four deep marches, …". It had only been working for three days. When I ask for ETAs, the estimates are always either way too long or way too short. I tried telling Claude not to commit to an estimate until it ran a test to confirm, but that didn't work well either. Over time, I started to get a feel for how long things would take myself, since I couldn't rely on Claude to tell me.

更大的问题在于,Claude 的默认做法似乎是硬啃一个长达数天的计算,而不是构建一个新工具让同样的计算只需几分钟。我一次又一次不得不告诉它要更聪明地思考,而不是更费力地思考。在这些长期项目中,上下文压缩(compaction)经常触发,导致 Claude 丢失重要上下文。为防止这一点,我让 Claude 定期整理并合并其文件,以便它始终能访问最新版本的计划。我已将协议技能和其他工具写入 BootLoops 框架来解决这个问题。其中一些问题在优秀的商业框架(如 Claude Science)中也已得到解决。但许多问题只是当前一代智能体式 AI 固有的恼人之处,我预计模型最终会克服它们。

A bigger problem was that Claude's default approach seemed to be to try to grind through a long, multiday calculation rather than build a new tool that would make that same calculation take mere minutes. Time and again, I had to tell it to think smarter, not harder. Over the course of these long projects, compaction would often kick in, causing Claude to lose important context. To guard against this, I had Claude periodically organize its files and consolidate them, so it always had access to the latest version of the plan. I've written protocol skills and other tools into the BootLoops harness to address this. Some of these problems have also been solved in good commercial harnesses, like Claude Science. But many are just the sorts of irritations inherent in the current generation of agentic AI that I expect the models will eventually outgrow.

以下是我遇到的另外几种失败模式,以及应对建议:

Here are a few more failure modes I encountered, and tips for dealing with them:

* **Claude 喜欢宣布胜利。**“完成了,只有一个星号”往往意味着“根本没完成”。“完全正确,只有一处改进”通常意味着“不对”。为 Claude 设定清晰且严格的成功标准有助于解决这个问题。在一个项目中,它对自己的证明很自豪,直到“一个未证明的引理”。那个引理就是整个证明!我说不允许有未证明的引理。然后它又说“完成了”,但这次多了一个新公理。

* **Claude loves to declare victory.** "Done, with one asterisk" is often "not done at all." "Exactly that, with one refinement" usually means "no." Giving Claude clear and rigid standards for what success looks like can help with this. On one project, it was proud of its proof, up to "one unproven lemma." That lemma was the whole proof! I said no unproven lemmas. Then "done" again, but now with a new axiom.

* **亲自查看一切。** 我总是要求看图表。即使我设置了所有监控,我发现自动检查仍然不可信。要警惕诸如“吻合良好”这类定性说法。

* **Look at everything yourself.** I always ask to see plots. Even with all the monitors I set up, I've found the automated checks still can't be trusted. Beware of qualitative claims like "good agreement."

* **质疑结论。** Claude 擅长执行计算,但它得出的结论可能是错的。让它解释它发现了什么,直到 _你_ 相信为止。

* **Question the conclusions.** Claude is good at performing calculations, but the conclusions it draws can be wrong. Have it explain what it found until _you_ believe it.

* **提供品味。** Claude 能找到 Claude 式的问题,但这类问题有成千上万个,甚至可能数百万个。Claude 似乎偏爱那些被高度引用但早已被遗忘的旧争论。让它关注什么是 _新近可能_ 的,而不是仅仅更快或涉及更多数字的东西,这会有帮助,但你仍然不能信任它对某事物是否有趣的判断。

* **Supply the taste.** Claude can find Claude-shaped problems, but there are thousands, perhaps millions, of them. Claude seems to favor old debates, highly cited but long forgotten. Asking it to focus on what is _newly possible_, not what is merely faster or involves more digits, helps, but you still can't trust its judgment of whether something is interesting.

* **追问。** 一旦 Claude 取得了很好的结果,就要求一个更好的结果。它总能想出点什么,偶尔还会非常出色。

* **Follow up.** Once Claude has a great result, ask for an even greater one. It will always come up with something, and occasionally it's brilliant.

* **当心无休止的消耗。** 老实说,我从未成功让 Claude 很好地估计时间。相反,我培养了一种感觉:一件有趣的事情应该花多长时间;我也学会了感知 Claude 是否在我认为合理的时间尺度上取得了进展。如果你放任不管,模型会永远磨下去。

* **Watch out for the grind.** Honestly, I never succeeded in getting Claude to estimate time well. Instead, I developed a sense of how long something interesting should take, and I learned to sense whether Claude was making what I considered progress on a sensible time scale. The model will grind forever if you let it.

同样值得提及的是,这些项目在算力和 token 上都是密集型的。对我而言,部分开销来自尝试构建能超越任何单一项目、具有更广泛用途的工具:连接不同领域,并组装出一个他人可以直接接手的软件包。我希望 BootLoops 框架能够广泛用于定量科学,而无需每位用户都从头重建一切。智能体式 AI 让连接想法和共享代码库变得更加容易;如果我们学会利用这一优势,整个社区就能比任何个人单打独斗走得更快。

It also bears mentioning that these projects were compute- and token-intensive. Part of the expense, for me, came from trying to build tools that would be useful beyond any one project: connecting fields, and assembling a package that others could pick up. I hope the BootLoops harness can be used broadly across quantitative science without each user having to rebuild everything from scratch. Agentic AI makes it easier to connect ideas and share codebases, and if we learn to use that strength, the community can move faster than any individual could alone.

展望 Outlook

无论模型变得多么聪明,大多数科学进展仍来自现实世界的数据——这些数据必须被获取、理解并核验,每一轮都会为下一个问题提供线索。AI 可以置身于这一循环之中,这确实很有价值,但它并不能把循环压缩成一个点。理解是逐步积累起来的,每一步都建立在前一步之上;癌症的治愈方法或聚变反应堆也将以这种方式到来,而不是来自某一条提示词。

No matter how smart the models become, most scientific progress comes from real-world data that has to be acquired, understood, and checked, with each round informing the next question. AI can sit inside that loop, and that is genuinely valuable, but it does not collapse the loop to a point. Understanding is built up in increments, each one resting on the last; a cure for cancer or a fusion reactor will arrive that way too, not from a single prompt.

对千禧年大奖难题和“大科学”的关注,或许正如 Claude 喜欢说的那样,是一把“脚枪”(footgun):一种让人很容易搬起石头砸自己脚的特性。这里的风险在于,不切实际的期望可能会抑制那些已经可行的、富有成效的用途。真正的进展很可能仍会以一贯的方式到来:夯实科学研究的基础,并在其上构建日益复杂的应用。正如我所描述的工作所示,这一进程将被大幅加速——但我看不到任何证据或必要去重新审视科学方法。

The focus on Millennium Prize Problems and Big Science may be, as Claude likes to say, a "footgun": a feature that makes it easy to shoot yourself in the foot. Here, the risk is that unrealistic expectations may discourage productive uses that are already possible. Real progress will probably come the way it always has: by strengthening the foundations of scientific research and constructing increasingly sophisticated applications upon them. It will be sped up dramatically—as illustrated by the work I've described—but I don't see any evidence or need to revisit the scientific method.

我喜欢的一个术语有时被用来讨论 AI 科学,那就是**凸包(convex hull)**。取一个形状,用直线连接其任意两点;得到的新形状就是包含原形状的最小凸形状:它的凸包。如今科学非常零散:到处是参差不齐的前沿。生物学朝一个方向突出,数学朝另一个方向。一个实验室可能花 20 年时间,用他们掌握的一种方法彻底研究一组基因,而邻近的基因和其他方法却无人问津。像 BootLoops 这样的框架可以把这些点连接起来,填补凸包。

A term I like that is sometimes used to discuss AI science is the **convex hull.** Take a shape and connect every pair of its points with a straight line; the new shape you get is the smallest convex shape that contains the original: its convex hull. Science is now very disjointed: there are jagged frontiers everywhere. Biology pokes out one way, math another. One lab might spend 20 years learning everything about a single set of genes with a single method they have mastered, while neighboring genes and alternative methods lie fallow. A harness like BootLoops can connect these points, filling in the hull.

事情变化很快,几乎不可能提前规划。为什么要申请三年资助去计算一个 AI 模型可能一夜之间就解决的问题?我仍然不知道该如何培养研究生。自从我发表了那篇关于“氛围物理学”(vibe physics)的文章以来,这种感觉变得更加强烈。在某些领域,比如计算机科学,这种颠覆简直令人恐惧。两年前,我会把“面向工程师的 Python”课程列为必修。如今,它已不再必要。我曾花大量时间研究机器学习,构建模型来研究各种物理现象。现在,所有这些模型都可以由 AI 构建,因此即便是学习神经网络的来龙去脉,用处也有限。只要告诉 Claude 去找最先进的机器学习方法,它就会替你完成。

Things are changing fast, and it's practically impossible to plan ahead. Why apply for a grant for three years of funding to compute a problem an AI model might end up solving overnight? I still don't know how to train graduate students. Since I published my piece on vibe physics, that feeling has only become more acute. In some fields, such as computer science, the disruption is downright scary. Two years ago, I would have recommended a course on "Python for engineers" as essential. Today, it's unnecessary. I have spent a lot of time on machine learning, building models to study various physical phenomena. Now, all of those models can be built by AI, so even learning the ins and outs of neural networks is of limited utility. Just tell Claude to find the state-of-the-art ML approach and it will do it for you.

我认为 BootLoops 表明,如果你找到了正确的问题,解决它的许多技术工作现在都可以自动化。从某种意义上说,这是一种解放。但人类仍然需要负责概念性的部分。正因如此,我希望我们能找到一种方式,让人类即便在“Claude 式科学”中也能获得应有的认可。奇怪的是,AI 有可能颠倒通常的商业模式:在那种模式中,位居顶层的人获得认可,而亲手解决问题的人却一无所获。我不知道这会如何解决,但假装人类的贡献只是打字,是解决不了的。

I think BootLoops shows that if you find the right problem, much of the technical work of solving it can now be automated. That is liberating, in a way. But humans are still needed for the conceptual part. Because of that, I hope we can figure out a way to give humans the credit they deserve even for Claude-shaped science. Curiously, AI runs the risk of inverting the usual business model, where the person at the top gets the credit and the people working on the problem hands-on get none. I don't know how this resolves, but it won't be resolved by pretending the human contribution was the typing.

那么,这给普通科学家(Joe Scientist)带来了什么?如果你问我,实际上是一个很好的境地。我们现在手头有工具,可以以闪电般的速度在看似棘手的问题上取得进展。我可以预见,在数据丰富但理论匮乏的领域,如系统生物学,将出现巨大的增长。我也很容易想象,人类思维将从繁琐的计算和分析中解放出来,去从事更深层次的智力工作——指导和引领。最重要的是,我预计科学家们将走出自己的领域,寻求协同与合作,就像我与 BootLoops 所做的那样。关注 Claude 式科学最令人耳目一新的结果是,它帮助我们欣赏人类式科学的样子。

So where does this leave Joe Scientist? If you ask me, in a great place, actually. We now have tools at our fingertips that can make progress on seemingly intractable problems with lightning speed. I can foresee huge growth on the horizon in data-rich-but-theory-starved fields like systems biology. I can also easily imagine human minds being freed from the tedium of calculation and analysis to do the deeper intellectual work of direction and guidance. And most of all, I anticipate scientists branching out from their home turf to find synergy and collaboration, as I have with BootLoops. The most refreshing outcome of focusing on Claude-shaped science is that it helps us appreciate what human-shaped science looks like.

关于 BootLoops 工具及其所支持的科学的详细信息,可以在 www.bootloops.ai 找到。BootLoops 工具已在 GitHub 上发布,供任何人使用和贡献。

Details of the BootLoops harness and the science it has enabled can be found at www.bootloops.ai. The BootLoops harness is on GitHub for anyone to use and contribute to.

互动版:针对本篇提问 →

接着读 · Read next