DeepSeek Introduces In-House System for AI Training

[Financial News] Chinese artificial intelligence (AI) company DeepSeek introduced its in-house system for training AI models in a new paper.
According to Pengpai and other outlets on the 24th, DeepSeek published a paper titled 'DeepSeek Elastic Compute (DSec): A Sandbox Infrastructure for Large-Scale Agent Training' on the preprint platform arXiv on the 19th. The paper, co-authored by more than 130 people including DeepSeek founder Liang Wenfeng, introduced DSec, a sandbox mechanism designed for large-scale agent training and evaluation.
A sandbox is an isolated environment for running computer programs and other applications. In the AI field, it refers to an isolated virtual environment where numerous commands can be executed safely for the large-scale training of AI agents.
According to the paper, one basic unit of DeepSeek's DSec consists of about 160 nodes (servers), with 30,000 central processing unit (CPU) cores and 250 terabytes (TB; 1,024 gigabytes) of memory. It also creates 3 million virtual workspaces (sandboxes) every day where AI agents can test code and can run more than 380,000 sandboxes simultaneously. The system maintains an efficiency of generating more than 5,000 sandboxes per second.
The paper also addressed the security risks associated with operating AI agents on a large scale.
 


[email protected] Kang Gwi-gui Reporter