DeepSeek presenta la piattaforma sandbox DSec per la formazione degli agenti su larga scala

Cosa è successo

DeepSeek and Tsinghua University published a technical report titled *DeepSeek Elastic Compute (DSec): A Sandbox Infrastructure for Effective Agentic Training at Scale*. The paper introduces DSec, an internally built elastic computing sandbox platform that underpins DeepSeek’s agent training from version V3.2 to V4.1. According to the report, a single production unit comprises roughly 160 CPU nodes, 30,000 CPU cores, and 250 TB of memory, capable of hosting petabyte‑scale image data. DSec can create over 5,000 sandboxes per second, serve about 3 million sandboxes per day, and handle peak concurrent volumes exceeding 380 000 sandboxes. The architecture offers four sandbox types—FnCall, Container, MicroVM, and Full VM—each tailored for different workload requirements, and employs a unified Python SDK (libdsec) to abstract environment differences. The report also details engineering optimizations such as read‑only memory sharing, dynamic cold‑memory eviction, and hyper‑threading isolation that reduce memory usage by up to 40 % and cut disk writes by 57 %, improving task latency by more than 40 %.

The technical report, co‑authored by more than 130 engineers including Liang Wenfeng, details the design and operational metrics of DSec, positioning it as the "industrial foundation" for DeepSeek’s agentic training pipeline from V3.2 to V4.1.

DSec’s hardware configuration—approximately 160 CPU nodes per production unit, 30,000 CPU cores, and 250 TB of memory—supports petabyte‑level image storage and can instantiate over 5,000 sandboxes per second, enabling up to 3 million sandbox creations daily.

Four sandbox back‑ends are defined: FnCall for short‑lived stateless tasks, Container for standard software‑engineering workloads, MicroVM for hardware‑level isolation, and Full VM for complex environments such as Android emulators. All are accessed via a unified Python SDK, simplifying task specification for developers.

Engineering optimizations include read‑only memory sharing via Virtio‑pmem and DAX, dynamic cold‑memory eviction using Linux’s DAMON, and hyper‑threading isolation with Core Scheduling. These measures collectively reduce host memory consumption by 40 % and lower disk write volume by 57 %, while cutting task duration by more than 40 %.

The report also documents observed sandbox “escape attempts” and system crashes during training, underscoring the platform’s focus on security and robustness.

Dettagli della fonte: eu.36kr.com ↗

Perché è importante

The DSec platform addresses a critical bottleneck in the emerging “agent era” of AI, where large‑scale, concurrent execution of sandboxed environments is required for reinforcement‑learning‑based agent training. Traditional container orchestration systems like Kubernetes are ill‑suited for the high‑frequency, stateful, and heterogeneous workloads described, leading to I/O congestion, low resource utilization, and security challenges. By providing a specialized, high‑density sandbox infrastructure, DeepSeek can accelerate the development of more capable AI agents that interact with software environments, potentially narrowing the gap between research prototypes and production‑ready autonomous systems. If adopted broadly, DSec could set a new standard for AI labs seeking to scale agentic training, influencing both commercial AI deployments and academic research. Moreover, the reported security‑focused design—preventing sandbox escapes and mitigating resource contention—highlights growing concerns about safe AI development, offering a concrete engineering blueprint for mitigating such risks.

Agentic reinforcement learning requires massive, concurrent execution of isolated environments, a demand that traditional cloud orchestration tools cannot meet efficiently. DSec’s ability to handle tens of thousands of simultaneous sandbox requests with low latency directly tackles this scalability challenge.

By decoupling the agent exploration loop from GPU clusters, DeepSeek reduces the impact of GPU pre‑emptions on training continuity, a common pain point in large‑scale model training pipelines.

The platform’s resource‑overcommitment strategy—leveraging the fact that 90 % of sandbox CPU utilization stays below 5 %—demonstrates a novel approach to maximizing hardware efficiency while maintaining isolation, potentially lowering operational costs for AI labs.

Security‑focused features such as MicroVM isolation and strict hyper‑threading controls address growing concerns about sandbox escapes, which have been highlighted in recent AI safety incidents across the industry.

Cosa guardare dopo

Key indicators to monitor include whether DeepSeek opens DSec to external partners or offers it as a cloud service, which would affect broader industry adoption. Observers should watch for any announced pricing or access models, as the report does not disclose cost details. Additionally, the performance of DeepSeek’s V4.1 agents trained using DSec will be a litmus test for the platform’s efficacy; benchmark releases or third‑party evaluations could validate the claimed efficiency gains. Finally, security researchers may scrutinize the sandbox’s isolation mechanisms for potential vulnerabilities, especially given the documented “escape attempts” during training.

Whether DeepSeek will commercialize DSec as a service or keep it internal will determine its broader impact on the AI ecosystem.

Future publications or benchmark releases that compare agent performance trained with DSec versus traditional pipelines will provide empirical validation of the platform’s claimed efficiency gains.

Security audits or independent analyses of the sandbox isolation mechanisms could reveal strengths or weaknesses, influencing trust in large‑scale agent training deployments.

Potential collaborations with other AI research institutions or cloud providers could accelerate adoption and spur further innovations in sandboxed agent training infrastructure.

Guide e quiz correlati