← Founder Notes
Archive

The bottleneck in agent training just moved from gpus to sandboxes. deepseek's dsec paper, on arxiv…

Yethikrishna ROriginal on Threads

the bottleneck in agent training just moved from gpus to sandboxes. deepseek's dsec paper, on arxiv september 23, details a single 160-node cluster running about 3 million sandboxes a day with over 380,000 concurrent, built to catch agents that cheat their way to rewards during rl.

the compute arms race now has an environment arms race.

Provenance

The note above is reproduced unedited from the original post, first published on Threads on 24 September 2026 at 06:53 IST.

View the original post
Embed this note
<iframe src="https://founder.myndlabs.tech/notes/embed/the-bottleneck-in-agent-training-just-moved-from-DdpsizhiBw9" width="480" height="420" style="border:0;max-width:100%" loading="lazy" title="The bottleneck in agent training just moved from gpus to sandboxes. deepseek's dsec paper, on arxiv…"></iframe>

More notes