The search at work

Live run

The search at work: the LLM proposes policies, the simulator scores them. Model: Qwen3-8B via vLLM.

Recording started 2026-10-17 00:17:05 UTC. Showing 2026-10-17 00:17:06 UTC (0 s of 742 s).

Speed

0simulations

0 / 8round · fixed

0 / 0 / 0 LLM proposals valid / invalid / dup

—tokens/s

—best objective

Use 0% VRAM 93% Power 144 W Requests in flight 0

AMD Instinct MI300X · ROCm

AMD Instinct MI300X VF

The GPU works in bursts: it peaks while the LLM proposes a round of policies, then drops while the CPU runs the simulations that score them.

GPU use %

VRAM %

Power W