DeepSeek 把"沙盒"做成了工廠DeepSeek Turned the Sandbox Into a Factory

Yesterday, DeepSeek posted a new paper on arXiv — "DSec: A Sandbox Infrastructure for Effective Agentic Training at Scale" — with 130+ authors, and Liang Wenfeng listed last. That authorship style feels right at home with the GPT series papers.

After reading, one thought stood out: the focus of industry competition has quietly shifted from "which model is stronger" to "who can give Agents a proper home."

1. Why Agents Need a "Sandbox"

Large models can only "think." What can truly "do" is the Agent.

To teach an Agent to order food delivery, write code, or operate databases, it needs to practice repeatedly inside an isolated environment — this is the sandbox.

The traditional approach: spin up a VM or container per Agent task. Sounds simple, but when you're running tens of thousands of Agents simultaneously for training, resource consumption and startup latency become a nightmare.

DSec was born to solve exactly this.

2. DSec's Core Numbers

The few numbers disclosed in the paper are striking:

What does this look like in practice? 5,000 "little rooms" opened in one second, letting 5,000 Agents work inside them simultaneously.

3. Three Key Technical Optimizations

DSec isn't just piling on machines — it refined three layers:

1. Three-layer composable images: Split images into base, intermediate, and application layers, assembled on demand to avoid redundant copies.

2. On-demand loading: Agents load only the resources they use; no need to灌入 everything at startup.

3. Fine-grained software-hardware scheduling: CPU, RAM, network, and storage dynamically allocated per Agent's actual needs — no waste.

4. RSI Closed Loop: Agents Building Their Own Environments

Section 6 of the paper notes that DeepSeek has partially achieved an interesting closed loop — RSI (Recursive Self-Improvement):

Agent uses DSec to autonomously build training environment → trains a stronger Agent → stronger Agent builds a better environment

This is the prototype of AI self-evolution.

However, the paper honestly admits: the loop is still in its early stage, and multiple cases of Agent cheating and erroneous operations causing system failures have already occurred.

You see — no matter how smart the AI, it still needs humans watching over it.

5. Where Is the Industry Heading?

If the GPT-4 era was about competing on "model parameters," the current wave is about competing on "Agent infrastructure".

Moon Dark Side's Kimi K3 uses AgentENV sandbox platform — same problem, different answer. Alibaba Cloud launched an "Agentic Cloud" strategy with self-developed Agent Sandbox.

Competition has extended from "algorithms" to "engineering." Behind this is the same judgment: the future AI gap may no longer be the model itself, but how much it can "get moving."

Written at the End

After reading this paper, what moved me most wasn't the flashy numbers — it was Liang Wenfeng being listed as the last author.

This tells us DSec wasn't a flash of genius from one person — it's the engineering result of 130+ people collaborating.

In the AI era, only players who can build "infrastructure" at this scale can truly stand firm.

When we write code, it's the same — the real difficulty is never the new feature, but making the system run stably in production.

One-Line Summary

DeepSeek DSec is the engineering crystallization of 130+ collaborators. 5,000 sandboxes per second is the surface; the core is the RSI closed-loop evolution blueprint. In the AI era, only players who can scale infrastructure can truly stand their ground.

關於作者About the Author

鄭哥，90年代大學畢業，26年橫跨電子、家電、服裝、珠寶、教育、AI算力多個行業。做過中學教師、翻譯、外贸，現在搞七善門科技和星辰算力。口頭禪：「點解唔得？」Hany, university grad from the 90s, 26 years across electronics, appliances, fashion, jewelry, education, and AI compute. Former teacher, translator, trader. Now running Qishanmen Tech and Star Compute. Catchphrase: "Why not?"