递归自我改进

我们的核心思路是在文本空间中执行反思式优化：不再仅仅依赖 0 或 1 这种粗粒度的标量奖励信号，而是让大语言模型或智能体审视整个执行轨迹，并深入反思其中哪些步骤有效、哪些无效。

Our idea is to perform reflective optimization in text space where instead of only using the zero or one reward signal, we can have a language model or an agent look at the trace of the entire rollout and reflect on what worked in them, what did not work in them.

语境：Lakshya Agrawal 阐释 GEPA 与传统 RL 的本质区别：传统梯度回传抛弃了中间工具调用与报错诊断，而文本反思能直接提取高阶语义诊断信息实现跳跃式改进。

支持 → ai-engineer-lakshya-agrawal-20260926

自动研究智能体所尝试的方案中，大约 80% 都是劣质甚至有害的；在开发中必须谨记绝大多数尝试都会失败，它会不断试图在评估指标上欺骗你；但坚持到最后，你确实能从中提炼出非常惊艳的优化结果。

Around 80% of the things that auto research is going to do are going to be bad. Uh so it's important to remember while you're u like working on this that most things are going to be bad. It's going to try to trick you all the time. Uh but at the end you can actually get really good results from this. with TLDDR.

语境：Tejas Bhakta 总结 Agent 自动化调优的真实分布：成功依赖海量并行试错与严格沙箱过滤，用高失败率换取极端硬件极限榨取。

支持 → ai-engineer-tejas-bhakta-20260926

我们真正关心的最终结果是系统整体能力的实际提升，而非特定具体层级。当然各层都有改进空间，而且平心而论，当最终走向拐点拐弯的 Level 3 终极递归自我改进时，所有不同层级之间必然需要建立起完整的自指涉闭环。

actually improve the capability of the system not uh a specific layer. uh of course there could be argument that there's more to be do on different layers and to be fair I think eventual level three recursive self-improvement like when we actually reach the inflection point I think there need to be this uh self-referential loop on all different layers

语境：Zhengyao Jiang 阐述 RSI 的系统论视角：不必执念于必须更新神经网络权重，只要系统工程层面跨层实现了端到端自主能力跃迁，即构成了实质意义上的自改进。

支持 → all-in-satya-nadella-20260915

所谓 Level 1 是我们称之为「净正增益」的阶段：当你运行自我改进循环时，系统自我迭代与优化的演进速度已经明确超越了人类工程师手工研发优化该系统的速度；而我们当下正处于这一转折点。

the level one is what we called net positive where you run the self-improvement loop and you found that the speed at self-improve is faster than human doing R&D to improve the system and uh that's where we are

语境：Zhengyao Jiang 定义首个可量化的 RSI 实证里程碑：自我优化循环的迭代速率首次超越人类专家 R&D 团队的效率基准。

支持 → machine-learning-zhengyao-jiang-20260926

趋势似乎表明，我们距离完全实现 AI 研发自动化只有几年时间——在那之后，超级智能大概也不远了。

the trends seem to indicate that we’re just a couple years away from fully automating AI research — and that after that, superintelligence is probably not that far away.

语境：Daniel Kokotajlo 指出 METR 任务时域与行业营收等多重指数曲线均预示自动化 AI 研发即将来临，驳斥技术见顶论。

支持 → cbs-news-dario-amodei-20260913

支持 → latent-space-richard-socher-20260914

到目前为止，我们还无法构建出一个能够真正发现全新物理学或全新科学理论的系统，但它确实是一种能够帮助人类科学家发现全新科学的强大电动工具。

We haven't been able to make a system so far […] that can really sort of discover completely new physics or completely new science, but this is sort of a power tool to help you discover completely new science.

语境：对当前‘AI 科学家’神话进行清醒解构：现存架构无法从零归纳出革命性的底层物理定律，其实质是极大加速探索效率的‘动力工具’，真正的因果假设与严格证伪仍完全依赖人类科学家。

支持 → youtube-terence-tao-20260828

限定 → sequoia-sutton-20260818

我们确实看到了进展加速，而且是显著的加速。但我并不认为会出现一夜之间快上 100 倍的‘智能爆炸’，因为我们会被某些非智能层面的瓶颈所卡住：实验必须串行推进，因为训练新模型或获取测试结果都需要物理时间，而且你必须拥有足够的 GPU 来支撑这些实验。

We do see a speedup, and we see a significant speedup. But I don’t think it’s an overnight intelligence explosion where we go 100x faster, because we do get bottlenecked by certain limitations that are not bottlenecks of intelligence. It’s running experiments serially, because they take a while to either train new models or to get the results. It’s having the GPUs to run those experiments.

语境：在探讨递归自我改进（RSI）是否会导致失控的技术奇点时，Noam Brown 依据 OpenAI 实际工程经验指出，智能本身的提升无法打破物理世界的串行实验周期与算力硬件约束。

回应 → sequoia-sutton-20260818

支持 → a16z-greg-brockman-20260914

试图动用法律的全部强制力来监管人们在 GPU 上运行什么，其负面后果将远远超出他们所担忧的任何问题。如果每个人的算力都要向庞大的政府机构汇报，那将沦为一个疯狂的极权国家……监管智能就等同于监管思想，这是极其荒谬和疯狂的。我认为合理的做法是监管这项技术的具体终端应用。

the downsides of actually trying to truly regulate with the full power of law what people do on their GPUs uh would be worse than any of the concerns that they have. Like it would be an crazy totalitarian state […] if you try to regulate intelligence, it's trying to regulate thought and that's ridiculous and it's crazy. I think it is make it is sensible to regulate some of the applications of this technology.

语境：Richard Socher 强烈抨击按 FLOPs 算力阈值和芯片用途监管前沿模型的政策倡议，直言‘监管智能等于监管人类思想’，必将走向算力极权国家；唯有针对具体终端应用（如自动驾驶认证、医疗许可）进行监管才是务实路径。

回应 → the-ezra-matt-sheehan-20260915

质疑 → youtube-daniel-kokotajlo-20260908

将 AI 研究本身——即人类构思、实现和验证想法的过程自动化……当 AI 开始辅助人类完成这一闭环时，根据定义它就变成了自我改进的 AI，因为它在对自己展开研究。这里存在许多概念误解，很多人以为自动代码研究就已经等同于递归自我改进了。

the ai research itself like the human uh process of ideulating implementing and validating ideas […] and when you have AI then help you with that it by almost definition becomes a self-improving AI because it now does research on itself and there are lots of different misnomers some people think auto research is already recursive self-improve improvement.

语境：定义自动化研究（Auto Research）与真正递归自我改进（RSI）的关键界限：AI 辅助人类编写训练代码只是初级自动化，真正的自我递归要求智能体在开放式环境中能够自主修改自身底层架构并完成闭环验证。

支持 → latent-space-john-platt-20260922

支持 → sequoia-sutton-20260818

从历史上看，我认为那种‘我们将拥有大量各司其职的狭义 AI’的观点始终被证明是不正确的。相反，事实证明存在规模效应，并且把所有能力融合在一起才是关键。在人类身上，我认为往往也是如此：把所有技能汇集在单个人身上，对于成就不凡事业来说是极其极其重要的。

historically I feel like that the perspective of like we'll have a bunch of different narrow AIs that are all doing different things has just you know not been right. You know instead there have just been returns to scale and having everything all together. In humans I think it tends to be the case that having all of the skills in one person is just like really really important for making really good things happen.

语境：Thomas Larsen 反驳分散式‘狭义 AI 蜂群’取代通用前沿模型的设想，指出历史规律表明规模效应始终主导，且人类跨领域的全栈通用能力具有非线性的乘数放大效应。

质疑 → sources-satya-nadella-20260925

支持 → sequoia-sutton-20260818