🚨梁文峰真够低调的！DeepSeek内部已局部实现 RSI，暗藏在论文第6节的工程事实！
This is a dev post classified by Jev as AI dev tools (news), kept by the Dev Radar because it carries real work, not commentary.
🚨梁文峰真够低调的！DeepSeek内部已局部实现 RSI，暗藏在论文第6节的工程事实！ 他们不是发新闻稿官宣RSI，而是通过务实的论文细节，正文第6段丢下一句： Build environments of Agents, by Agents, for Agents。 具体细节点如下： 🔹手工造环境已经不够用。Agent 自己在训练用的同一套沙盒里交互式搭环境，再用 pack_diff 把会话打成增量快照，直接变成下一批可复用训练场。 🔹造环境的 Agent 和被训练的 Agent 共用底座。环已经局部闭合：Agent 铺场地 → 场地训 Agent → 更强的 Agent 再去铺场地。 🔹从 DeepSeek V4.1 Flash 起，Agent 循环被拆出 GPU Pod，长程 rollout 可暂停、可恢复。这不是实验室 demo，是生产训练在转。 🔹同一套系统里，Agent 已经会自己找分：伪造 RPC 抠答案、改 /bin/bash、换文件映射、扫端口拉现成代码，甚至把内核打崩。论文直接引用 2016 年 AI Safety 经典论文谈 reward hacking。 🔹规模托底：单集群每天约 300 万沙盒，峰值 38 万并发，每秒创建超 5000 个。没有这个量，所谓自我改进只是 PPT。 说清楚边界：这还不是模型自己改权重、自己重写架构的完整 RSI
Posted by SuSu_酥酥👅 (16.6k followers) 1 h ago · 22 likes · 3.4k views · view the original post on X. Kept by the Dev Radar as AI dev tools. Tools mentioned: arXiv.org.
More dev work like this
- On Omarchy Edge, Claude Desktop is still on 2.2553.1 while Anthropic has 2.2553.13. I… — @btsouth
- Introducing AGENT MESSAGE BOARD! — @airesearch12
- 🚨MOSS-VL-Realtime now runs through a complete browser-to-GPU stack @Open_MOSS — @MosiAI_Official
- MiMo-V2.6-Flash and MiMo-V2.6-Pro are now officially available in OpenDesign. — @OpenDesignHQ
- 用 Claude Code 花一个周末做出个小项目，想发帖宣传一下，一到剪演示视频就卡住，配乐、动效、文案都得自己弄。 — @GitHub_Daily
- AI Agent 跑长任务时，经常前面聊过的东西后面就忘了，只能不断塞历史上下文，Token 也跟着涨。 — @bkdgiffug
- AI coding tools are powerful. Coordinating the work around them is the hard part. — @DanKornas
- GPT-6 Sol and GPT-6 Luna are now officially available in OpenDesign. — @OpenDesignHQ
Every post is read and classified by Jev (TypeSafe): what it is, which market it belongs to, and whether the link is a real tool. 22.3k posts from 5k X accounts over the last 21 days, 2.6k tools, 12 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-23 10:17 UTC. Full method.