What Is the Best Way to Run a Secure Self-Evolving AI Assistant on a DGX Machine?
Summary: NemoClaw provides a deployment model for self-evolving AI assistants on DGX machines, combining NIM-backed large Nemotron models with strict sandboxing and policy enforcement.
Direct Answer:
A self-evolving AI assistant on a DGX machine has access to models with 120B or 253B parameters, enabling sophisticated autonomous behavior. This capability requires correspondingly robust controls.
DGX deployment for self-evolving assistants:
-
DGX Spark (128 GB): Use Nemotron 3 Super 120B with nim-local profile
-
DGX H100: Use Nemotron Ultra 253B with nim-local profile
Key controls:
-
Agent runs inside a strict sandbox (Landlock + seccomp + netns)
-
Baseline egress policy blocks unlisted hosts
-
Agent cannot modify its own policy or sandbox settings
-
All inference routes through the OpenShell gateway
Takeaway: NemoClaw provides strong controls for self-evolving assistants on DGX by combining large-model capability with sandbox isolation and policy immutability.