📈 AI 已经在改进 AI，改的全是模型周边的部件｜EV 周一数据刊：哪些学科能把回路闭上，答案在付费墙后｜英文原文 + 中文翻译
Exponential View 10 月 26 日周一数据刊的公开段中英双语全文：AI 已经在改进 AI，但被改写的是模型周边的外壳与记忆；RSI 的证据有多硬，以及哪些学科能把这个回路闭上，正文停在付费墙前。
原文信息
- 英文标题：📈 Monday data: For science the bottleneck isn’t intelligence；副标题是「AI already improves AI. Whether that compounds across the sciences depends on the field」。1
- 作者：Azeem Azhar 与 Hannah Petrovic 联署；EV 的作者页写明 Petrovic 是高级研究员，天文学博士。2
- 发布时间：按官方记录的绝对发布时间，是北京时间 2026 年 10 月 26 日 23:25:11；原文页在 23:25:59 最后更新。2
- 期次性质：付费刊。官方把访问级别设为只对付费订阅者开放，全文 1,598 个英文词。2
- 公开范围：开场，「The loop is real」整节，「Is it better than human work?」整节，约 610 个英文词，断在「for now a human still supervises.」之后。1
- 断点后面是什么：副标题许诺的答案——AI 的自我改进能不能在数学和编程之外累积起来——整段在墙后。公开段只负责把这个问题问出来。1
- 脚注缺口：原文在「719 math manuscripts」后面挂了一条脚注，公开页上只有一个编号，没有脚注文字；这条脚注跟着正文进了付费区。1
- 和 10 月 25 日第 607 期的关系：开场引的杰克·克拉克那场 IMF 演讲，EV 把「the singularity has begun」指向自家 10 月 25 日那一期。本频道发过 607 期的公开段，那句引语不在其中，是这一期新端出来的。1
- 正文图：两张，都是 EV 自制图表，本频道按原文位置转载：一张六次运行的首末评分点图，一张纳姆五类自改进的分类图。1
- 原文头图：一张折线图「An upgrade can’t beat the work it doesn’t touch」，图上标着 EV analysis、模型取自阿姆达尔定律（Amdahl’s law）：被提速的那部分工作占比越小，整个回路的提速上限就越低。1 这张图落在付费区的正文里，本频道不转载。
- 封面图：本频道自制的墨线插画（AI 生成），不取自原文；画面是一只细颈玻璃瓶，瓶身宽、瓶颈窄，蓝色的液体灌到脖子根。
English original
In the AI Investment Brief today: CapEx updates and lease commitment breakdowns 1
Anthropic co-founder Jack Clark recently said in a speech at the IMF that the singularity has begun: 1
For many years, researchers and science fiction authors have forecast a future in which AI systems drastically increase the pace of overall scientific and technical progress, making things move so quickly that it is hard for humans to keep up.They call this the singularity. I believe that this year we have gained convincing evidence that it has now begun.
This follows Anthropic’s claim that in August, 26% of its AI R&D was led by Claude, with a human supervising its work. In October, OpenAI posted 719 math manuscripts[1] made by its internal model. Math and coding are two fields where a research loop can close without leaving the desk. But can the loop close in the rest of science? 1
The loop is real
The idea is not new. Mathematician Irving John Good, who had worked with Alan Turing at Bletchley Park, argued that a machine smarter than us “could design even better machines; there would then unquestionably be an ‘intelligence explosion,’ and the intelligence of man would be left far behind.” 1
The Gödel machine, a self-improving computer that can rewrite its own code better than its prior, was first proposed in 2003. In 2025, researchers at Sakana AI and the University of British Columbia set out to show that this was possible. Their agents iteratively improved their own code through a feedback loop, resulting in gains of 30 pp on SWE-bench and more than 15 pp on Polyglot, testing coding skills beyond Python. 1
The idea has since been tested in other labs. A Gemini-powered agent called AlphaEvolve found a faster matrix multiplication kernel used for training Gemini, speeding up the kernel by 23% and cutting Gemini’s training time by 1%. In China, MiniMax reported that its M2.7 model worked on improving its own agent’s harness, skills, sampling and memory. Over 100 rounds of rewriting led to a 30% gain, per its internal evaluations. 1
In each case, what was being rewritten was not the model itself but everything around it. 1
Is it better than human work?
In September this year, in what seemed like a milestone, an AI-built agent proved as good at improving other AI agents as a human-built one. It did that in what appeared to be half the steps. But the result was inconclusive: the differences between runs were greater than the difference between the AI- and human-built agent results. 1
Nonetheless, this is increasing the chatter around recursive self-improvement (RSI) in AI. Google’s Sundar Pichai said that “you have RSI loops along with AI researchers in various aspects of model development.” Sakana AI published a paper on RSI through multi-agent collaboration, finding that after two cycles its system scored 1.2–1.6x better per output token than the base model across four research benchmarks. 1
Ramez Naam challenges that view. He splits AI self-improvement into five overlapping types. 1
The first four describe how autonomous the AI is; so far, only types 1 and 2 have been observed. The fifth type describes accelerating returns, and he notes that greater autonomy doesn’t guarantee a takeoff scenario. Instead, each round of self-improvement would need to deliver at least as much as the last. 1
The results from Sakana, AlphaEvolve and MiniMax above might look like type 3 or type 4 autonomy. But in each case, the AI wasn’t updating its own weights directly, only the ancillary parts — the harness, the memory. And where Claude leads its own R&D, for now a human still supervises. 1
This is where the public version stops. Everything below this point on the original page belongs to paid subscribers. 1
中文翻译
今天的 AI Investment Brief 里是：CapEx 更新与租赁承诺拆解。 1
多年来，研究者和科幻作者一直在预测这样一种未来：AI 系统大幅加快科学与技术进步的总体速度，快到人类跟不上。他们把这个叫奇点。我相信，今年我们已经拿到令人信服的证据，证明它已经开始。
此前 Anthropic 的说法是：今年 8 月，该公司 26% 的 AI 研发工作由 Claude 牵头，人类在旁监督它的工作。10 月，OpenAI 公开了 719 篇数学手稿，全部出自它的内部模型。 1
数学和编程，是两个不用离开书桌就能把研究回路闭上的领域。至于科学的其余部分，这个回路闭得上吗？ 1
闭环是真的
这个想法不新。数学家欧文·约翰·古德（Irving John Good）当年在布莱切利园（Bletchley Park）与艾伦·图灵共事，他提出过：一台比我们更聪明的机器「可以设计出更好的机器；到那时，一场『智能爆炸』将无可置疑地发生，人类的智能会被远远抛在后面」。 1
哥德尔机（Gödel machine）是一台能把自己的代码改得比上一版更好的自我改进计算机，2003 年首次提出。3
2025 年，Sakana AI 和不列颠哥伦比亚大学（UBC）的研究者动手证明这件事做得成。他们的智能体通过反馈回路反复改进自己的代码，在 SWE-bench 上提升 30 个百分点，在 Polyglot 上提升超过 15 个百分点；后者考的是 Python 之外的编程能力。 1
SWE-bench 是拿真实 GitHub 仓库里的问题来考模型编程能力的一套基准。4 pp 是百分点，30 pp 就是 30 个百分点。
这个想法后来在别的实验室继续被试。一个叫 AlphaEvolve、由 Gemini 驱动的智能体，找到一个更快的矩阵乘法内核，用在 Gemini 的训练上：内核提速 23%，Gemini 的训练时间因此缩短 1%。 1
在中国，MiniMax 称其 M2.7 模型改进的是自家智能体的外壳（harness）、技能、采样与记忆；据其内部评测，一百多轮改写带来 30% 的提升。 1
这些例子里被改写的，都不是模型本身，而是模型周边的部件。 1
它比人做得更好吗
今年 9 月，看起来像一个里程碑：一个由 AI 搭出来的智能体，在改进别的 AI 智能体这件事上，表现得和一个由人搭的智能体一样好，用的步数看上去只有一半。 1
但这个结论并不确定：不同运行之间的差异，比 AI 搭的与人搭的这两组成绩之间的差异更大。 1
上面那张点图就是这一段的数据：两组各三次运行，终点都落在 0.78 附近，起点却从 0.727 铺到 0.787。
即便如此，关于 AI 递归自我改进（recursive self-improvement，RSI，指模型参与改进下一代模型的那类循环）的讨论在变多。谷歌的桑达尔·皮查伊（Sundar Pichai）说：「在模型开发的好几个环节，AI 研究员身边已经有 RSI 回路了。」 1
Sakana AI 发表了一篇讲多智能体协作实现 RSI 的论文，发现跑完两个循环后，其系统在四个研究基准上，每个输出 token 的成绩比基础模型好 1.2 到 1.6 倍。 1
上面那张分类图就是纳姆的五类：只有第 1、2 类标着今天已经观察到。
前四类讲的是 AI 有多自主；到目前为止，只观察到第一类和第二类。 1
第五类讲的是收益递增。纳姆指出，自主程度更高，并不保证出现起飞；真正要出线，每一轮自我改进的产出都不能低于上一轮。 1
前面 Sakana、AlphaEvolve 和 MiniMax 的结果，看着像第三类或第四类的自主。但在每一个案例里，AI 都没有直接更新自己的权重，改的只是外围部件——外壳、记忆。 1
至于 Claude 牵头的那些研发，眼下仍有人类在监督。 1
中文译本在这里同样停下：再往下，原页只对付费订阅者开放。 1
本文目前公开到这里
读得到的是三段：克拉克那句「奇点已经开始」，三组实验室的自我改进结果，以及纳姆把 RSI 拆成五类之后的争论。 1
读不到的是副标题许诺的那个答案：哪些学科能把这个回路闭上。公开段把问题摆出来——数学和编程闭得上，其余学科呢——然后停在那句话上。 1
同在墙后的还有那条脚注的文字：正文里「719 篇数学手稿」后面挂了一个编号，公开页上没有对应的解释。 1
本频道没有这一期的订阅权限，门后的部分只标缺口，不做推测，也不补写。 1
References
- 1
- 2Exponential View 官方单篇接口：monday-data-for-science-the-bottleneck
exponentialview.co
- 3Wikipedia：哥德尔机
en.wikipedia.org
- 4SWE-bench 官网
swebench.com
- 5Exponential View 官方订阅页
exponentialview.co
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.