$Lobster
DeepSeek releases a paper on its public proxy-training sandbox, with Liang Wenfeng listed as an author
DeepSeek has published a new paper disclosing its sandbox foundation for training AI agents in-house. According to the arXiv paper titled “DeepSeek Elastic Compute (DSec),” the system is designed to support large-scale agent training and evaluation, and the founder Liang Wenfeng is also listed as an author.
A paper on an SDK that manages four types of sandboxes says that when using large language models for large-scale agent training and evaluation, the model needs to inspect libraries, call tools, run commands, and interact with task-related services within isolated execution environments that preserve state. Such workloads create sandboxes in large numbers over short periods; their required functionality and isolation needs vary, while state must be preserved during long interactions, and images must be retrieved from a huge image repository—yet reuse rates are not high. Therefore, the paper argues that what is needed to support these workloads is a flexible execution platform, not a single type of sandbox execution environment.
DSec’s approach is to use a unified SDK to provide four backends: function calls, containers, micro virtual machines, and full virtual machines. The system is responsible for scheduling placement and managing lifecycles in the cluster, assembling layers that are version-controlled independently for each environment. It also combines memory sharing, garbage collection, and CPU scheduling to increase execution density, while loading image data on demand from the cluster-level distributed filesystem 3FS.
Designed together with reinforcement learning frameworks, the paper separates stateful rollouts from preemptible training. DSec and the reinforcement learning (RL) framework are jointly designed: it decouples stateful rollout execution from preemptible GPU training, and coordinates the sandbox lifecycle with training—reclaiming idle resources while preserving rollout state, and addressing issues caused by improper agent behavior.
According to a TechNode report, a production-scale unit consists of about 160 nodes, can support around 3 million sandboxes per day, runs more than 380,000 sandboxes at the same time, and can create over 5,000 sandboxes per second. The paper has more than 130 authors.
This article, “DeepSeek releases a paper on its public proxy-training sandbox (DSec), with Liang Wenfeng listed as an author,” first appeared on .
DeepSeek releases a paper on its public proxy-training sandbox, with Liang Wenfeng listed as an author
DeepSeek has published a new paper disclosing its sandbox foundation for training AI agents in-house. According to the arXiv paper titled “DeepSeek Elastic Compute (DSec),” the system is designed to support large-scale agent training and evaluation, and the founder Liang Wenfeng is also listed as an author.
A paper on an SDK that manages four types of sandboxes says that when using large language models for large-scale agent training and evaluation, the model needs to inspect libraries, call tools, run commands, and interact with task-related services within isolated execution environments that preserve state. Such workloads create sandboxes in large numbers over short periods; their required functionality and isolation needs vary, while state must be preserved during long interactions, and images must be retrieved from a huge image repository—yet reuse rates are not high. Therefore, the paper argues that what is needed to support these workloads is a flexible execution platform, not a single type of sandbox execution environment.
DSec’s approach is to use a unified SDK to provide four backends: function calls, containers, micro virtual machines, and full virtual machines. The system is responsible for scheduling placement and managing lifecycles in the cluster, assembling layers that are version-controlled independently for each environment. It also combines memory sharing, garbage collection, and CPU scheduling to increase execution density, while loading image data on demand from the cluster-level distributed filesystem 3FS.
Designed together with reinforcement learning frameworks, the paper separates stateful rollouts from preemptible training. DSec and the reinforcement learning (RL) framework are jointly designed: it decouples stateful rollout execution from preemptible GPU training, and coordinates the sandbox lifecycle with training—reclaiming idle resources while preserving rollout state, and addressing issues caused by improper agent behavior.
According to a TechNode report, a production-scale unit consists of about 160 nodes, can support around 3 million sandboxes per day, runs more than 380,000 sandboxes at the same time, and can create over 5,000 sandboxes per second. The paper has more than 130 authors.
This article, “DeepSeek releases a paper on its public proxy-training sandbox (DSec), with Liang Wenfeng listed as an author,” first appeared on .
