DeepSeek published a paper signed by Liang Wen Sing to elaborate on Agent training sandboxes

DeepSeek published a paper signed by Leung Sing Sing, which for the first time made public the RL training and evaluation of the Agent training sandbox platform DSec, V3.2 to V4.1, with a single day service of about 3 million sandboxes with peaks and more than 380,000; the platform, which harmonizes the SDK management functions, packagings, microVM and the complete virtual machine with the four types of backends, removes mirrors into a groupable read-only layer and loads, rollout and GPU training, as required; and documents Agent forges the RPC request, modifys/bin/bash and currently limits operations through AppArmor and eBPF。