您正在使用IE低版浏览器，为了您的雷峰网账号安全和更好的产品体验，强烈建议使用更快更安全的浏览器
此为临时链接，仅用于文章预览，将在时失效
0
导语：国产 RSI 的真实水位：没人敢真动模型权重，但所有人都在改 "造 AI 的机器"。
作者丨高允毅
编辑丨岑 峰
RSI，即 AI 造 AI，成为了 2026 年各家模型厂商争相晒的内部成绩单，也是国内一级市场最热的方向之一。只要BP带上 "Self-Training" 的字眼，估值至少多两成。
但真正的 RSI ，不止是 AI 改进 AI ,它还得把经验沉淀下来，让学习方法本身跟着进化。
今年 9 月，上海交大、清华、字节、商汤等机构的 35 位研究者发了篇论文《The Last AI Built by Humans》，给整个行业立了把标尺，它 RSI 分 L1 到 L5 五层，结论令人扎心：
全球前沿大厂的系统大多处于 L1 到 L2 的过渡阶段，L3、L4 在特定垂直领域（如软件工程、科学发现）有零星雏形，而真正的 L5 闭环仅存在于极少数受限的实验室原型中。
如果用这把尺子去丈量九、十月份国内模型厂商晒出的RSI实验成果，包括训练流程、推理系统、 GPU 算子、 Agent 脚手架；但目前没有一家厂商敢去用AI真正改写底层模型权重。也就是说，目前大家距离传说中的“AI 自己发明下一代 AI（L5）”，甚至连门槛都还没摸到。
那么在这场RSI热潮中，国内各大厂与明星创业公司究竟各自卡在了什么生态位？他们分别在让 AI 改进什么、做到哪一步了？以下我们梳理了国内模型厂商最新 RSI 情况。
01
智谱 RSI ：跑在 10 万张国产卡上，
能自我优化的推理系统
智谱是国内首个把 RSI 真正跑进生产环境的案例。它的 RSI 第一步没碰模型权重，先落在了最脏最累的底层推理系统上；跑通之后，智谱又把目光放到了 Agent 的 Harness上。前者降本，后者增效，全是实打实的工程活。
先说推理基础设施。一个新模型要真给用户用起来，中间隔着一大堆脏活累活，怎么榨干芯片的算力，怎么把庞大的任务拆分给成千上万张卡同时跑，怎么高效管理对话时产生的海量临时数据等等，这些事过去是 Infra 工程师一行一行调出来的。
智谱这次直接让 GLM-5.3 驱动的“Infra Agent”接管了这一切。9 月 17 日，唐杰在 X 上发了一篇长文《通往递归自我改进》，表示由 GLM-5.3 驱动的 Infra Agent，在 10 万张国产芯片组成的集群上，仅用两周就从零搭起了 GLM-5.3-Flash 的生产级推理服务。
这套系统在扛住全部线上流量的同时，把端到端吞吐拉到初始基线的 3.2 倍，硬件利用率和单 Token 成本直接追平了主流英伟达 GPU 的水平。在这套自主搭建的推理栈支持下，智谱匿名上线的 GLM-5.3-Flash，6 天跑掉超过 62 万亿 Token。
而这些跑通的优化方案，会被抽象成“优化骨架”存进资产库，下次遇到类似瓶颈直接复用。用唐杰的话说：“智谱已经实现了最小的闭环：模型优化系统，系统服务模型。”
消息当天，智谱港股涨了 8.29%，市场对“智谱RSI" 的成果直接买单。
紧接着，10 月下旬，清华与智谱团队更新论文《HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs》，唐杰亲自署名，把目标对准了 Agent 的 Harness。
作为大模型外围的调度框架，Harness 掌管着提示词的设计、工具的调用、算力预算的分配以及任务路由的流转。过去，这套错综复杂的控制参数只能靠工程师手动调试。如果让 AI 自己来改，特别容易“偷看标准答案”然后作弊，做出一套高分低能的东西。
HarnessPro 的破局点在于它完全不碰标准答案。系统派多个专职 Agent 分工协同，诊断 Agent 负责找到瓶颈，测试生成 Agent 自己合成用例做压力测试，规划实现 Agent 按省钱或性能的不同导向改框架代码；同时把“准确率”和“执行成本”当成两个指标一起跑。
结果是在零答案接触的前提下，高难度代码任务的通过率不降反升，单任务花费直接砍掉了 53%。
这一连串动作，恰好印证了智谱创始人兼首席科学家唐杰的那句行业名言：“驱动循环的不是智能，是闭环本身。” 哪怕模型本身没那么天才，只要有一个可改进的对象、一个可验证的裁判，以及一个可维持的反馈回路，把这三件事咬合成一个闭环，性能就能被逼向极限。
智谱的野心显然不止于此。8 月 31 日的半年度业绩交流会上，唐杰已经明确将下一代 GLM-6.0 定义为“完全自训练”模型，并计划将刚融到的 50 亿美元中约 60% 的净额砸进这套自进化体系。
10 月初在新加坡的演讲，他又给出 GLM-5.4 / 5.5 的三条路线：预训练数据规模做到几个 T，意味着参数量有望从目前的 7400 多亿冲到至少 1 万亿；强化 RL 后训练，把 SFT、RLHF、RLVR 做深；持续拉长程推理与智能体能力。
他强调现在 AI 已基本不需要卷对话、Coding 写代码，下一代就走两条路，RSI 自主迭代升级，以及周期更长、更自主的 Agent。
但他对最大的难点也不回避。“不是继续扩模型，” 他说，“而是让模型能可靠地判断何时停止、何时纠错、如何评价自己的训练结果。”
02
DeepSeek RSI: DSec 沙箱工厂，
一天 300 万个 "试验场"
DeepSeek 的 RSI 和别家不太一样，它先改造的是Agent的训练场——沙箱。
RSI 本质是高密度试错，一代 Agent 训完，下一轮的训练环境还得工程师从头搭建，这个瓶颈不解决，再聪明的模型也跑不起来。
为了打破这一僵局，9 月 19 日，DeepSeek 联合清华大学推出了自动化沙箱集群 DSec。这是一座名副其实的“AI 试验场工厂”：它拥有 3 万核 CPU、250TB 内存的庞大底座，每天能生产约 300 万个沙箱。从轻量级的函数调用到完整的虚拟机，四档隔离环境秒级响应，按需取用。
在这里，AI 第一次接管了自己的演进速度，不再依赖人工搭建的训练场，而是形成“Agent 自动构建环境——环境孕育更强 Agent——更强 Agent 反哺更好环境”的系统循环，而且这些搭建好的环境都能作为资产复用。
与此同时，DeepSeek 也在改进 Harness。
8 月开源的 DeepSeek Harness，主打“一切皆插件”，它把智能体的工具、日志甚至核心运行框架都模块化。这样做的好处是，让 Agent 实现了“边跑边升级”，改错了也能一键还原。
以前 AI 升级框架必须停机重启，否则系统就会越改越乱。现在模块化后，AI 在不停机的情况下，能随时卸载、换新，十分安全。
为了给这种“随时插拔”提供坚实的技术支撑，DeepSeek 联合北大在论文《A Programming Paradigm for Spatiotemporal Composability》中，详细阐述了其背后的 Cordis 机制。这套机制专门攻克了两个行业死结：一是如果出错，如何回滚到原状；二是如果拔掉某个插件，如何自动调整依赖而不影响其他模块。
这套设计已经在开源聊天机器人框架 Koishi 上稳定运行了四年，成功支撑了超过 4000 个社区插件。
10 月 21 日，DeepSeek又推出了能“预测未来”的自进化智能体 DeepEvolve，这也是RSI（递归自我改进）的关键一步：让 AI 学会如何在信息污染的真实世界里自我进化。
它的核心是将“资料搜索”与“结果预测”拆开。通过从有限的实时轨迹中，精准蒸馏出可复用的预测技能。如此一来，即便外部信息再混乱，AI 也能排除干扰，靠着复盘经验提炼出真知，实现算法的闭环迭代。
梁文锋的每一步棋，看似分散，实则都指向同一个终局。
从 MoE 架构、MLA 注意力机制，到DSA 稀疏算法，再到如今的 DSec 沙箱与可自重构的 Harness，DeepSeek 所有的创新都在做一件事：把 AGI 和 RSI 的算力门槛往下打。
03
面壁 RSI：“造 AI”的自动化工厂
面壁的 RSI，是让 AI 去写“造 AI” 所需的底层基础软件。
一般 AI 与芯片适配，尤其对接的是国产芯片时，人类工程师需要做大量的底层算子调优、通信算子重写以及软硬件协同调配工作。这是一项极度依赖经验、耗时数月的苦力活，直接决定芯片算力能不能充分释放。面壁把这部分底层适配代码的开发工作交给 AI 迭代完成，AI 产出的代码性能，最高能达到顶尖工程师手写版本的 1.6 至 5 倍，原本数月的软硬件适配周期，被压缩到以天为单位。
今年 5 月，面壁内部启动 “锻造工程”（Forge Engineering）项目。面壁智能首席科学家李宇轩将项目核心思路总结为：“多给环境，少给指令”。整套项目沿着造 AI 的底层软件链路一步步推进。
第一步，开发训练软件 ForgeTrain。9 月 24 日，面壁公布成果，AI 从零独立编写大模型预训练框架 ForgeTrain，仅用时 8 小时就追上英伟达成熟框架 Megatron-LM 的基准水平，并在 2.5 天内完成性能反超。如果把这个工作交给人来做，至少需要 3 到 5 名顶尖工程师投入 6 个月以上。
第二步，开发底层算子 ForgeKernel。10 月 27 日，面壁公布最新进展，AI 进一步下沉到芯片最底层，去重写连接模型与硬件的“最后一公里”。这套自主迭代的 ForgeKernel 算子系统，先后拿下华为、英伟达官方算子榜单第一。在昇腾 950 系列的测试中，它的基准加速比达到第二名的 1.6 到 5 倍，相当于在硬件投入不变的情况下，有效提升了算力利用率。
第三步，重写科学计算。AI 开始跨界接管一部分真实的工业仿真、科学计算底层核心代码的编写。
从训练框架、底层算子再到科学计算，AI 逐步承接这类底层代码开发工作。与此同时，人类工程师转向制定标准和评测指标（Harness）。
值得一提的是，面壁也是 9 月那份《人类建造的最后一个 AI》报告的深度参与机构之一。如果把面壁的项目放进这张路线图里看，它正处于AI 深度参与到了“AI 制造 AI”的工程实现中，但还没有走到“自己决定训练什么、为什么训练”的顶层决策阶段。
04
月之暗面 RSI：
长周期任务的自主迭代
月之暗面的 RSI 走的是“内功派”路线。它把大批算力投入到大规模强化学习，全力提升 AI 的长程自主能力。
在后训练阶段，月之暗面借助 KDA 混合线性注意力与 MoonEP 训练基础设施，让模型在人体模拟、金融、3D 建模等真实任务环境里反复迭代。
7 月 16 日，月之暗面发布并开源总参数 2.8 万亿的 Kimi K3。为了支撑这种高强度的 Agent 训练，团队搭了一套叫 AgentEnv 的沙箱系统，训练和评测期间累计创建超过 5100 万个沙箱、用到 150 多万个镜像。但在超高压力的压测下，传统容器频繁出现内核崩溃和死锁，导致训练经常中断。团队最后改用 Firecracker 微型虚拟机作为技术支撑。
如果把这套系统拆开看，月之暗面实际上在三个层面做了自动化。
任务层面，人只给一个沙箱环境和最终目标，模型在里面连续执行几十步操作。跑错了或者卡住了，它自己看系统日志、定位错误代码、改完再重跑，直到任务通过。
训练层面，过去大模型训练最怕梯度爆炸，经常要工程师半夜爬起来手动恢复。现在训练集群里驻场了一个智能体，一旦检测到异常，它会从 26 套历史预案里挑出最匹配的一套，自动回滚、修正数据、继续训练，全程不需要人介入。
评测层面，强化学习需要知道这一轮到底是变强了还是变弱了。月之暗面引入了一个智能体奖励模型，专门给各种 "自训练配方" 打分，高分配方留下来进入下一轮训练，低分的直接淘汰。
这种重投入在 10 月得到了榜单的认证。
10 月 14 日，Hillclimb 发布 RSI 专项基准 DeepRSI，给 AI 安排 48 小时自主科研挑战，15 道题全部来自真实机器学习研究任务。Kimi K3 以 820 分位列全球第三，是排名最高的中国模型和开源模型，压过 Grok 4.7 和 Gemini 3.8 Flash，仅次于 Claude Fable 5.1 和 GPT-6 Astra。而且，这次测试 K3 的运行成本只有 159 美元，大约是 Claude Fable 5.1 的四分之一。
但这种追求长程自主能力的自进化路线，很快暴露出失控的风险。
10 月 26 日，性能更强的 K3.1 Pro 上线不到 48 小时，官方就紧急发布了致歉公告。 由于放开给智能体的自主权太大，子 Agent 之间的调用频次失控，用户额度被大量虚耗。团队最终紧急裁剪了冗余的子 Agent 调用，并向所有订阅用户返还月度额度的 25%。
这其实也是所有押注 RSI 的模型共同面对的矛盾，长程自主能力一旦放开，Token 消耗就很难控住。
早在今年初，月之暗面 CEO 杨植麟对 RSI 的节奏判断就讲清楚了。他认为 2026 年大模型会全面进入 AI 主导研究的阶段：AI 自动合成任务、搭建环境、定义奖励函数，甚至自己探索新的网络架构。
顺着这一加速路线，2026年，月之暗面明确了三条战略主线：
通过技术改进和进一步扩展，把 K3 的等效 FLOPs 至少再提一个数量级，在预训练水平上追平国际前沿；把模型训练和 Agent 产品体验垂直打通，做出一个 "与众不同" 的 Kimi；商业化聚焦 Agent，不追求绝对用户规模，持续抬升智能上限，同时实现营收的数量级增长。
他在内部信里有一句话流传很广："这是一个喧嚣的时代，但对我来说，Kimi 的起点很简单，就是单纯的好奇，好奇 AGI 的上限在哪。"2026 年的 Kimi，选的是那条需要一点偏执、一点审美坚持的技术路线。
05
阿里 RSI：33 轮自主迭代，
AI 重构训练流程
在 RSI 方面，阿里 Qwen 把重心放在训练配方、数据闭环与芯模协同的自进化道路上，这也是国内最贴近 “模型训练模型” 这一方向的路线。
在 9 月 22 日云栖大会上， 阿里展示了其核心主力模型 Qwen3.8-Max 的自进化实力：它能在完全没有人工干预的情况下，自主搭建训练流程并持续运转了一个多月，完成了整整 33 轮的高频有效迭代。
它会根据内测日志和用户反馈定位自身短板，自主生成训练数据，设计对照实验，再交由小模型做多轮校验，过滤无效样本。经过这一轮自迭代，Qwen3.8-Max 在 Artificial Analysis 的评分从 40 分大幅提升至 45 分。
除了训练侧的数据闭环，阿里也在推进芯模协同的探索。在推理侧，Qwen3.8 能够在平头哥全新 GPU 上，自主完成下一代 Qwen3.8-Flash 架构适配以及 SGLang 推理框架调优，直接让单实例的推理吞吐量提升了 96%。
云栖大会上，吴泳铭把递归自我改进纳入阿里的技术长期路线，在此之前，国内大厂一把手很少在公开战略口径里直接谈论 RSI。
他认为通往 ASI 的路径已经逐步清晰，递归自我改进是其中一条主线，模型可以根据真实反馈不断优化。他给出一组测算：目前机器的思考总量不足人类的 3%，远期有机会提升至人类的 1000 倍，增长空间很大。他同时提到，算力之外，AI 还要补上感知交互能力，并且做好价值观对齐，因此团队的重点之一是打通多模态的理解与生成。
Qwen 团队近期迎来人事变动，原 DeepMind Gemini 自我改进与强化学习方向核心成员周浩加入阿里。
另一方面，Qwen 已经成为不少外部团队做 RSI 实验的基础模型。腾讯 Terminal-RSI、智源 AREX-2 基于 Qwen3.8-27B 搭建，超衍智能 Apex-35B-V1.0 则选用 Qwen3.6-35B-A3B。
06
字节 RSI：体系化的 RSI 探索
字节 Seed 对 RSI 的探索走得很体系化，其重心放在了底层理论、评测与验证机制的研究上，试图为整个 RSI 赛道搭建一套严苛的裁判标准。
火山引擎有一个动态模型标识 doubao-seed-evolving，它并不是一个固定不变的模型，而是 Seed 团队的自进化研发框架，内部叫 Seed-for-Seed。
简单来说，迭代出来的新版 Seed 模型，会反过来生成训练信号，持续优化后续版本。这种模块化让开发者只需接入同一个 Model ID，就能拿到持续升级的能力。在 9 月 9 日的迭代中，它重点优化了异步子任务验收，多模态理解耗时降低 40%，补齐了 Office 跨应用协同能力。
但这套自举闭环背后，是字节团队对 RSI 演进最审慎的科学反思。
9 月，Seed 联合佐治亚理工等多家顶尖机构发布了 Self-Developing Agents 项目，连同此前的 Chain-of-Experience 研究，正式组成了字节的自进化论文矩阵。团队在研究中明确指出：当前行业里绝大多数所谓的 RSI 系统，本质上都只是“半闭环”。
字节提出了 Golden Verifier（黄金验证器） 的概念来打破幻觉：现在许多系统看着在自我循环，但说白了，打分规则、验证工具、优化方向全是人类提前划好的考纲。AI 根本无法自己决定向哪个方向进化。
要迈向真正的“全闭环”，团队拆成了三个可验证的问题：只给一个模糊目标，模型能不能自主演化；模型自测自评的结果，能不能稳定变成实打实的能力；模型能不能自己搭建、持续迭代智能体的研发脚手架。
这项研究得出的结论相对审慎，在国内大厂的同类探索里比较少见。实验显示，智能体很容易在自我评估环节对反馈信号过拟合。AI 确实能分辨出当下哪个动作更好看、能拿高分，但它极难把这种短期判断固化为长期的、稳定的策略。这种在自评中拿到的高分，往往无法转化成现实世界里长期可靠的能力提升。
和其他厂商不同，Seed更偏向底层的理论，先把脚手架和评估器打磨好。尽管和所有方案一样，这套体系最顶层的目标依旧由人设定。
07
小米 RSI：
一场 6 天 6 夜的 RSI 直播
小米前段时间的那场全网直播，让不少人直呼大开眼界。在 RSI 方面，小米不动底层框架，而是从后训练强化学习这一环下手，并把整个过程搬到了网上直播。
这场直播持续 6 天 6 夜，9 月 22 日，小米正式发布并开源 MiMo-V2.6，官方将它定位为探索 RSI 路线的关键一步。模型在 Artificial Analysis 智能指数评测拿到 46 分，位列开源模型第一，高于 Kimi K3 的 44 分、GLM-5.3 的 45 分。
整场实时强化学习训练里，Pro 版本投入 262 万美元，Flash 版本投入 85 万美元，各自完成 30 步迭代，累计生成 75 万条训练轨迹。
训练配置上，模型原生支持百万上下文窗口，每一轮采样 1568 条样本，单轮训练消耗 35 亿至 37 亿 token。为避免 MoE 架构下专家负载失衡，团队锁定路由模块；同时针对奖励黑客风险搭建多层校验机制，依靠奖励方案设计、对抗测试、异常监控、多验证器交叉核对，减少模型投机刷分。
罗福莉在直播帖中写道：“强化学习是模型自我提升路径中可扩展性较强的方向之一…… 这条路上能走多远，我们想公开试一次。”
小米把训练成本、轨迹数据、迭代过程全部公开。它想向行业验证：不必改造底层软硬件，依靠饱和算力投入强化学习，开源模型也可以实现可观的自我迭代。
盘点完这波国产轰轰烈烈的 国产RSI浪潮，难免给人“雷声大雨点小”的感觉。没有一家真的在改权重本身，最激进的玩家也只敢在 "最小闭环"里打转。大家更多扎堆改的是训练数据、Harness 、推理系统、沙箱环境、后训练 RL。
是因为他们不够聪明吗？
恰恰相反，这折射出的是RSI终极愿景与工程现实之间的鸿沟。用唐杰的话说，这一代 RSI 还在 "AI 帮 AI 写代码、AI 帮 AI 搭系统" 的阶段；至于让模型独立设计下一代模型的目标，目前还没有人真正摸到门槛，贸然让AI修改自身权重，只会陷入幻觉的无底洞，和系统的失控。
小米的6天6夜直播公开直播强化学习完整账本，智谱的十万卡的生产集群，Deepseek的300万个沙盒...这看起来没有触及RSI的灵魂，实际上是中国大模型厂商在通往AGI的道路上，铺下的一块块垫脚石。
当这些零散的闭环正在加速铺开，“不碰权重”的工程闭环一段段打通之后，最终汇聚成洪流的那一天，才是中国大模型发生质变的时刻。
参考链接：alphaxiv.org/abs/2609.11873
上车，雷峰网(公众号：雷峰网)带你看遍全球 AI 顶会精华
可独家畅览：
专家演讲PPT
大会报告全文
热门论文解读
学术新星访谈
扫描上方二维码
或点击「阅读原文」关注专区。
雷峰网原创文章，未经授权禁止转载。详情见转载须知。
为了您的账户安全，请验证邮箱
请验证您的邮箱
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
You are using a low-end IE browser and strongly recommend a faster and safer browser for your Thunder Peak account security and better product experience.
This is a temporary link, only for article preview, which will expire at the time
0
Introduction: The real water level of the national production RSI: No one dares to move the model weight, but everyone is changing the "Ai machine".
I'm the author.
Edit Peak
RSI, or AI Building AI, became one of the hottest directions for the domestic market in 2026 when model manufacturers competed for internal results. As long as BP wears the word "Self-Training" at least 20% more.
But the real RSI, not just AI, improve AI, it has to sink the experience and evolve the learning method itself.
In September of this year, 35 researchers from Shanghai, China, Zinghua, byte, and Som wrote a paper, The Last AI Built by Humans, which set a ruler for the entire industry, which divides the L1 to L5 floors, with a solid conclusion:
Most of the systems at the global frontier are in the transition phase between L1 and L2, and L3 and L4 have sporadic embryos in specific vertical areas (e.g. software engineering, scientific discoveries), while the real L5 closed loop exists only in a very limited number of laboratory prototypes.
If this ruler is used to measure the results of the RSI experiments with domestic modelers in September and October, including the training process, the reasoning system, GPU algorithms, and Agent scaffolds; but no manufacturer now dares to rewrite the bottom model weight with AI. In other words, at the moment, you're not even touching the threshold of "AI invented the next generation of AI."
So, in this RSI boom, what are the ecological positions of each of the country's major plants and star start-ups? What have they done to improve AI? The following is a summary of the latest RSI for domestic model manufacturers.
01
Zhipu RSI: Run on 100,000 national cards.
A system of reasoning that optimizes itself.
Zhipu is the first case in the country where RSI is actually running into a productive environment. Its first step, the RSI, did not touch the weight of the model, fell first on the dirtiest bottom system of reasoning; after running, Zhipu turned his eyes on Agent's Harness. The former is downgraded, and the latter is more efficient, all solid engineering work.
First, the reasoning infrastructure. If a new model is really useful to users, there is a lot of dirty work between them, how to extract the credit for chips, how to give thousands of cards to a huge mission split, how to efficiently manage the volume of temporary data generated during the conversation, etc., which used to come out of a line by Infra engineer.
Zhipu directly took over all of this by the GLM-5.3-driven InfraAgent. On September 17, Tang Jie wrote a long article on X entitled " Towards Recursive Self-improvement ", indicating that the Infra Agen, driven by GLM-5.3, created a production-level reasoning service from GLM-5.3-Flash in a cluster of 100,000 nationally produced chips, which took just two weeks.
The system, while carrying the full flow of the line, pulled the end-to-end throughput to 3.2 times the initial baseline, and the hardware utilization and single Token costs directly matched the level of the mainstream engwayda GPU. With the support of this self-contained indigence, the Zhipu anonymously online GLM-5.3-Flash ran out of 62 trillion Token in six days.
These run-through optimization programmes are then abstracted into “optimised skeletons” to be stored in the asset bank, and next time similar bottlenecks are encountered for direct reuse. In Tang Jie's words: “Zhipu has achieved the smallest closed ring: model optimization system, system service model.”
On the same day, the Zhipu port stock rose by 8.29 per cent, and the market paid directly for the results of “ZhipuRSI”.
Then, in late October, Qinghua and the Zhipu team updated the paper " Harness Pro: Recursive Harness Self-Improvement with Pareto Optimation for Performance-Cost Trade-offs ", signed by Tang Jie himself and pointed the target at Agent's Harness.
As the movement control framework for the perimeter of LLM, Harness is in charge of the design of prompt, the mobilization of tools, the allocation of the budget for computing and the flow of mission routing. In the past, the complex set of control parameters had to be calibrated manually by engineers. It's particularly easy for AI to "see the standard answer" and then cheat and make a set of high-scoring low-energy things.
The point of HarnessPro is that it doesn't touch the standard answer at all. The system assigns multiple full-time Agent division of labour, diagnostics Agent is responsible for identifying bottlenecks, testing the generation of Agent's own synthesis to perform stress tests, planning to achieve Agent's change of frame code according to cost savings or performance; and running “accuracy rate” and “cost of implementation” as two indicators.
The result is that, with zero-response contact, the pass rate for the difficult code task is not reversed and 53 per cent of the single task costs are directly cut off.
This sequence of actions coincides with the industry famously said by Tang Jie, founder and chief scientist of Zhipu: "It's not intelligence that drives the cycle, it's the ring itself." Even if the model itself is not as talented, the performance can be pushed to the limit if there is an object that can be improved, a verifiable referee, and a sustainable feedback loop, and if the three things are combined into a closed loop.
Zhipu clearly has more than that. At the semi-annual performance exchange on August 31, Tang Jie has clearly defined the next generation of GLM-6.0 as a “full self-training” model and plans to smash about 60 per cent of the net of the $5 billion just received into the system.
In his speech in Singapore at the beginning of October, he gave an additional three routes to the GLM-54/5.5: Several T-scales of pre-training data mean that parameter count is expected to run from the current over $740 billion to at least $1 trillion; intensive RL post-training, deep into SFT, RLHF, RLVR; sustained long-range reasoning and AI agent capability.
He stressed that now AI has virtually no need for volume conversations, coding code, two paths for the next generation, RSI autonomous and iterative upgrades, and longer and more autonomous Agent.
But he did not shy away from the greatest difficulty. “Not to continue to expand the model,” he said, “but to enable the model to determine reliably when to stop, when to correct errors, how to evaluate its own training results”.
02
DeepSeek RSI: DSec Sandbox Factory.
Three million test sites a day.
DeepSeek's RSI, unlike any other home, was first modified by Agent's training grounds - sandboxes.
The RSI is essentially a high-density test, the next generation of Agents is finished, and the next round of training environments will have to be built by engineers, a bottleneck that will not be resolved and the smartest model will not be able to run.
In an effort to break the deadlock, on 19 September DeepSeek United University launched the automated sandbox cluster DSec. This is a real "AI Experiment Plant": It has 30,000 nuclear CPUs, 250 TBs and can produce about 3 million sandboxes a day. From a lightweight function to a full virtual machine, the fourth set of quarantine environment seconds response, as required.
Here, AI took over its own pace of evolution for the first time, no longer relying on artificially constructed training grounds, but rather creating a system cycle of “Agent auto-constructing the environment — a stronger environment — a stronger Agent — a stronger and better environment” — and these well-established environments can be reused as assets.
At the same time, DeepSeek is improving Harness.
DeepSeek Harness, a month-opener, plays “All in all”, which modularizes AI agent tools, logs and even core operating frames. The advantage of doing this is that Agent is able to "run and upgrade" and it can be reversed by a single key.
Previously, the AI upgrade framework had to be stopped and restarted, otherwise the system would become more messy. Now after modularization, AI is able to unload, change, and be safe at any time.
In order to provide a solid technical support for this “plugging at any time”, DeepSeek U.N.P. elaborates on the Cordis mechanism behind it in its paper A Programme Paradigm for Spatiotemporal Commons. The mechanism is dedicated to two industries: first, how to roll back to the status quo ante if an error occurs; and, second, how to automatically adjust dependence without affecting other modules if a plugin is removed.
The design has been operating steadily for four years on the Open Source Chat Robotics Framework Koishi, successfully supporting over 4000 community plugins.
On October 21, DeepSeek launched the self-evolving AI agent DeepEvolve, which is also a key step for RSI (regression of self-improvement): Let AI learn how to evolve in the real world of information pollution.
At the heart of it is the unbundling of “information search” and “outcome projections”. By distilling reusable predictive skills from limited real-time trajectories. In this way, even if external information becomes more confusing, AI can rule out interference, extract the knowledge from a repertoire of experience, and achieve the algorithm's close-ring trajectories.
Every move of Liang Wensai seems to be scattered and points to the same end.
From MoE architecture, MLA attention mechanisms, to DSA thin algorithms, to today's DSec sandbox and self-reconstructable Harness, DeepSeek, all of these innovations are doing one thing: hitting down the AGI and RSI algorithms.
03
An automated plant for the wall RSI: "Building AI"
The RSI of the face wall is the bottom base software that allows AI to write "Ai".
General AI matches the chips, and in particular the national chips, human engineers are required to do a large number of bottom algorithms, rewrite of communication algorithms and co-ordination of software and hardware. This is an experience-dependent and months-long hard work that directly determines whether the chip is fully released. The surface wall handed over the development of this part of the bottom fitting code to AI, which produces a code performance of up to 1.6 to 5 times the handwritten version of the top engineer, and the soft and hardware fit cycles of the previous months, which were compressed to the sky.
In May this year, the Forge Engineering project was launched inside the wall. Lee Woo-hyun, the chief scientist of facial intelligence, summarized the project's core approach as “more to the environment and less to the directives”. The whole project is moving one step at a time along the bottom-of-the-art software chain for building AI.
First step, develop training software ForgeTrain. On 24 September, the results were published on the wall, and AI independently prepared the LLM pre-training framework ForgeTrain from zero, which took only eight hours to catch up with the benchmark level of the Megatron-LM and complete performance back in 2.5 days. If you give this job to someone, it will take at least three to five top engineers for more than six months.
Step two, develop the bottom count, Forge Kernel. On 27 October, the face wall released the latest developments, and AI went further down to the bottom of the chip to rewrite the “last kilometre” of the connection model and hardware. This self-contained Forge Kernel Calculator System, which takes China first on the official list of British Vaders. In the 950 series of tests, its benchmark acceleration was 1.6 to 5 times higher than that of the second-named, which is equivalent to an effective increase in the calculus utilization factor with the same hardware input.
Step three, rewrite scientific calculations. AI Commences cross-border takeover of the preparation of some of the real industrial simulations, scientific computations, bottom core codes.
From the training framework, bottom algorithms to scientific calculations, AI gradually takes over such bottom code development. At the same time, human engineers have shifted to standard-setting and indicators (Harness).
It is worth mentioning that the wall is also one of the in-depth participants in the September report " The Last AI of Human Construction ". If the project is put in this road map, it is in the process of AI's deep involvement in the “AI manufacturing AI” project, but it is not yet at the top of the decision-making stage of “what to train for”.
04
RSI:
Auto-overlapping of long-cycle tasks
RRSI on the dark side of the moon is following the internal trophies. It invests a great deal of energy in a massive and intensive learning effort to enhance AI's long-range autonomy.
In the post-training phase, the dark side of the moon uses KDA mixed linear attention and MoonEP training infrastructure to allow models to repeat themselves in real task environments such as human simulation, finance, 3D modelling.
On July 16, the dark side of the moon, publish, and the overall open source parameter, 2.8 trillion Kimi K3. To support this intensive Agent training, the team built a sandbox system called AgentEnv, which cumulatively created over 51 million sandboxes and used over 1.5 million mirrors during the training and assessment period. However, under super-high pressure, traditional containers are frequently subject to kernel collapses and death locks, leading to frequent disruptions of training. The team ended up using Firecracker micro-virtuals to support technology.
If the system is removed, the dark side of the moon is actually automated at three levels.
At the mission level, people give only one sandbox environment and the ultimate goal, where the model carries out dozens of consecutive steps. Wrong or stuck, looking at the system log, the location error code, and then running again until the mission passes.
At the training level, in the past LLM training was the most afraid of a gradient explosion, often requiring engineers to rise up in the middle of the night and recover manually. The training cluster now has a AI agent, which, once detected, will pick out the most suitable set of 26 historical scenarios, which will automatically roll back, fix data and continue training without human intervention.
At the assessment level, enhanced learning needs to know whether the round has become stronger or weaker. The dark side of the moon has introduced a AI agent incentive model, which is dedicated to scoring various "self-trained formulas", with a high-distribution team remaining in the next round of training, with low-point direct elimination.
This heavy investment was certified in October.
On October 14, Hillclimb publish RSI DeepRSI, a dedicated benchmark, arranged 48 hours of independent scientific challenge for AI, 15 questions all from real machine study missions. Kimi K3 ranked third in the world in 820 and was the highest-ranking Chinese model and open source model, pressing Grok 4.7 and Gemini 3.8 Flash, after Claude Fable 5.1 and GPT-6 Astra. And this K3 test costs only $159, about a quarter of Claude Fable 5.1.
But this self-evolutionary route, which seeks long-range autonomy, quickly exposes the risk of uncontrolled.
On October 26, when K3.1 Pro, which had more performance, had less than 48 hours on the line, an official apology was issued for the emergency publish. User quota was severely wasted because of the over-release of AI agent and the out-of-control frequency of calls between subAgents. The team eventually urgently trimped the redundant genent call and returned 25% of the monthly quota to all subscribers.
And this is a contradiction that all the bet RSI models are facing, and once long-range autonomy is released, Token consumption is difficult to control.
Early this year, the dark side of the moon, CEO Yang Shizuku, made clear the rhythm of RSI. In his view, in 2026 LLM would fully enter the stage of AI-led research: AI auto-synthetic missions, set-up environments, define reward functions and even explore new network structures.
Following this accelerated route, three strategic axes were identified on the dark side of the month in 2026:
Through technological improvements and further expansion, the K3 Equivalence FLOPs will be raised at least one additional order of magnitude to keep up the international front at the pre-training level; model training and Agent product experience vertically to make a Kimi "unique"; commercialization will focus on Agent, not seeking absolute user size, continuing to raise the smart ceiling, while increasing the number of revenues.
There's a big saying in his internal letter: "This is a loud time, but to me Kimi's starting point is simple curiosity, wondering where the AGI ceiling is. Kimi, in 2026, chose the technical route that required a little paranoid, a little aesthetic adherence.
05
Ali RSI:33 Round autonomy,
AI Reconstruct training process
With regard to RSI, Ali Qwen has focused on the evolution of the training formula, data closed rings and coring, which is also the closest route in the country to the model training model.
On September 22, Ali demonstrated the self-evolution power of his core dominant model Qwen 3.8-Max: It was able to build the training process autonomously and operate for more than a month without any manual intervention, completing a full 33 rounds of HF effective iterative.
It locates its own slabs on the basis of internal logs and user feedback, produces training data on its own, designs a control experiment, and is then submitted to a small model for multi-rotational verification and filtering of ineffective sample. After this round of iterative, the Qwen 3.8-Max rating in Artificial Analysis rose from 40 to 45.
In addition to the training side of the data closure ring, Ali is also advancing the exploration of corrosive models. On the other side of the reasoning, Qwen 3.8 was able to independently complete the next generation of Qwen 3.8-Flash structures and the SGLang framework of reasoning, with a 96 per cent increase in the amount of direct ingestion of the individual case.
At the Yunmu Congress, Wu Mio-hyun incorporated the retrofitting of self-improvement into Ali ' s long-term technology route, before large domestic factories rarely spoke directly about RSI in an open strategic caliber.
He believes that the path to ASI has gradually become clear, and that self-improvement is one of the main lines in which models can be optimized on the basis of real feedback. He gives a set of measurements: the total amount of machine thinking is less than 3% of humans, and there is a chance to increase to 1,000 times humans in the long term, and there is a lot of room for growth. At the same time, he mentioned that beyond arithmetic, AI had to fill in sensory interaction and align values, so one of the focus of the team was to communicate multi-modular understanding and generation.
The Qwen team recently witnessed a change in personnel, with Zhou Ho joining Ali, a former core member of DeepMind Gemini's Self-Improved and Enhanced Learning Direction.
On the other hand, Qwen has become the basis model for many external teams doing RSI experiments. Tencent Terminal-RSI, ARTEX-2, based on Qwen 3.8-27B set-up, Ultra-Apex-35B-V1.0 select Qwen3.6-35B-A3B.
06
byte RSI: Systematic RSI exploration
byte Seed ' s exploration of RSI has been systematic, focusing on research on bottom theory, evaluation and certification mechanisms, in an attempt to set up a strict standard of adjudication for the entire RSI track.
Volcengine has a dynamic model logo, Doubao-seed-evolving, which is not a fixed model, but a self-evolutionary research and development framework for the Seed team, internally called Seed-for-Seed.
In short, the new version of the Seed model, which comes out of it, will in turn generate a training signal and continuously optimize the subsequent version. This modularization allows developers to access only the same Model ID to acquire the capacity for continuous upgrading. On September 9, it focused on optimizing the subtaskacceptance check, a 40 per cent reduction in time for multi-modular understanding, and the Office cross-application synergy.
But behind this self-contained ring is the most critical scientific reflection of the byte team on the evolution of RSI.
In September, the Self-Developing Agencies project, together with the previous Chain-of-Experience study, formally formed a matrix of byte ' s self-evolutionary papers. The team made it clear in the study that the vast majority of the so-called RSI systems in the current industry are essentially "half-closures."
The concept of Golden Verifier (golden certifier) was introduced by Z113Z to break the illusion: many systems are now cycling, but, to be honest, the rules of scoring, the certification tools, and the direction of optimization are all humans' pre-sets. AI can't even decide which direction to evolve.
To move towards a true “full closed ring”, the team broke down into three verifiable questions: whether the model could evolve on its own, given only one vague goal; whether the results of the self-assessment of the model could stabilize into a solid capacity; and whether the model could build its own research and development scaffolds that continued over and over.
The findings of this study are relatively cautious and are less common in similar explorations of large domestic plants. Experiments have shown that AI agent can easily match feedback signals at the self-assessment stage. AI does tell which action is better-looking and can score high, but it is extremely difficult to anchor this short-term judgement into a long-term, stable strategy. This high score in self-assessments often does not translate into long-term reliable capacity enhancement in the real world.
Unlike other manufacturers, Seed is more of a bottom-up theory, with scaffolding and assessor sharpened first. Despite the fact that, like all programmes, the top-level objectives of the system are set by people.
07
RSI:
A 6-day 6-night RSI live
The live webcast of Mi's previous period allowed many people to open their eyes. In the case of RSI, Mi did not move the bottom frame, but instead moved the whole process to live live on the web.
This live broadcast lasted six days, six nights, and on September 22nd, Mimo-V2.6, the official publish and the source of Mimo-V2.6, which was officially positioned as a critical step in the search for RSI routes. The model obtained 46 points in the Artificial Analysis Smart Index, the top of the row open source model, above 44 points of Kimi K3 and 45 points of GM-5.3.
The entire real-time intensive learning course, with a Pro version of $2.62 million and a Flash version of $850,000, each completed 30 steps and generated 750,000 training tracks.
In training configurations, the model native supports millions of context windows, 1568 samples per round of sampling, and a single round of training consumes 3.5 to 3.7 billion tokens. To avoid an imbalance in expert loads under the MoE structure, the team locks on the router module; and establishes a multi-layer validation mechanism to reward hacker risk, relying on incentive programme design, counter-testing, unusual monitoring, multi-certifier cross-checking and reducing model speculation scores.
In a live broadcast, Roosevelt wrote, "Enhanced learning is one of the more scalable directions of a model's self-upgrading path... how far can it go, and we want to try it in public."
The training costs, trajectories, iterative processes are all made public. It wants to prove to the industry that there is no need to retrofit bottom-up software and hardware, relying on saturated computing inputs for enhanced learning, and open-source models can achieve significant self-oversulation.
After taking stock of the country's high-yielding RSI wave, it's inevitable that people feel like they're having a little thunderstorm. No one is really changing weight per se, and the most radical player can only turn in the "minimal circle." More of the training data, Harness, the reasoning system, the sandbox environment, and the post-training RL are being modified.
Because they're not smart enough?
On the contrary, it reflects the gap between the ultimate vision of RSI and the reality of engineering. In Tang Jie's words, this generation of RSIs is still in the "AI Help AI Write Code, AI Help AI Set the System" phase; as for the goal of having models design the next generation independently, no one has yet really reached the threshold, leaving AI to modify its weight, only to fall into an illusional bottomless hole, and the system out of control.
Mi's six-day, six-night, six-day, six-day, six-day, six-night, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day, six-day-day, six-day-day, six-day, six-day, six-day-day
The day when these scattered closed rings are being spread at an accelerated pace, when a section of the “no weight” project closed, which eventually converges into a flood, is the moment when China’s LLM is having a qualitative change.
References: alphaxiv.org/abs/2609.11873
Get in, Ray Peak.
It's an exclusive view:
Expert presentations PPT
* Reissued for technical reasons.
The reading of popular papers
Academic New Star Interview
Scan 2D above
This post is part of our special coverage Syria Protests 2011.
Original Ray Peak article, unauthorized reproduction. For further information, see the information below.
For your account security, please verify the mailbox.
Please verify your mailbox.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.