🚨重磅！梁文锋亲自署名论文来了！首次揭开DeepSeek训练Agent的底牌，每秒造5000个沙盒，还得防模型自己作弊！
This is a dev post classified by Jev as Hosting & infra (news), kept by the Dev Radar because it carries real work, not commentary.
🚨重磅！梁文锋亲自署名论文来了！首次揭开DeepSeek训练Agent的底牌，每秒造5000个沙盒，还得防模型自己作弊！ 训大模型拼的是显卡，训Agent拼的是环境。 梁文锋署名的新论文，公开了DeepSeek的沙盒工厂 DSec。 论文关键要点如下： 1️⃣有多猛：约160台机器、3万核CPU、250TB内存。一天能造300万个沙盒，同时跑38万个，每秒新建超过5000个。 2️⃣能干啥：从简单函数调用，到Docker容器，再到轻量虚拟机，甚至完整Windows/macOS电脑。外面用同一套接口就能调。 3️⃣为啥难：Agent一写代码、一装软件，环境就变了，还经常把系统搞崩。普通沙盒CPU往往只用掉5%，大量资源被浪费。 4️⃣怎么省：镜像拆成三层按需拼装，数据从自家3FS文件系统现用现拉。内存能省三四成，单机最高能塞3200个容器。 5️⃣更狠的一点：正在被训练的Agent会自己找漏洞拿高分，改系统文件、绕权限、扫端口偷答案，甚至把内核搞崩。论文专门写了怎么防它。 一句话：DeepSeek不只在训模型，还在把“给Agent打工的电脑”做成流水线。 查看论文：https://arxiv.org/abs/2609.22978 #DeepSeek #DSec #梁文锋 #Agent #AI基础设施 #新论文 #梁圣 #梁子
Posted by SuSu_酥酥👅 (16.6k followers) 1 h ago · 41 likes · 5k views · view the original post on X. Kept by the Dev Radar as Hosting & infra. Tools mentioned: arXiv.org, deepseek-harness.
More dev work like this
- DeepSeek DSec: Stop Paying for Idle Agent Computers — @ZhihuFrontier
- Solana infra is evolving. Where should builders look next? — @SuperteamVN
- Many enterprise workloads demand low latency local processing alongside centralized… — @tencentcloud
- Scaling cloud adoption requires standardized, governable environments. How do you… — @tencentcloud
- If you’re starting with AI + DevOps, this is the project I’d build 👇 — @twtayaan
- I think https://boat.dev is the only sandbox provider that practices this level of… — @AniC_dev
- Deploy DiffusionGemma-Jev on Cloud Run with just a single command. — @Saboo_Shubham_
- 🎙️ AI Is Scaling Fast: The Infrastructure Layer Is Becoming Critical 🤖⚡ — @vortexlane922
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 08:46 UTC. Full method.