GitHub repo ai-trains-ai by Danau5tin open-sources a pipeline where an AI agent (Qwen3.6-35B-A3B with LoRA) is RL-trained via Tinker to write training jobs for smaller models (Qwen3-0.6B/1.7B) using prime-rl on Runpod GPUs. Two nested RL loops operate: the outer loop trains the agent, the inner loop trains the small model. The agent receives task specs, writes verifiers environments, rubrics, and prime-rl configs, then submits to a warm fleet of up to 16 GPU pods. Reward combines validation efficiency, job quality (uplift over baseline), and train speed.

Over 54 training steps, reward climbed from ~0.0 to ~0.63 in two distinct rungs: first process reliability (converting validation failures into completed episodes), then model quality (hidden-eval post-training score rose from ~0.04 noise to sustained 0.22–0.48). The agent learned to make better models, not just working ones.

Skill transfers to the held-out triage task family (never trained on), with performance rising then plateauing. The agent also learned to pick the better base model (1.7B over 0.6B, share went 42% → 95%) and adopt hyperparameters (prime_rl config usage 21% → ~78%). Infrastructure ran ~1,750 jobs at ~$0.13/job on RTX A5000, though mostly on A40s due to availability.