Show HN: I RL-trained an agent that trains models with RL (for –$1.3k) (github.com)

A developer has unveiled an innovative project that involves training an AI agent to create models through reinforcement learning (RL), essentially forming an AI that trains AI. This open-source project, which cost approximately $1.3k to run, features complete transparency by sharing the trained agent’s weights, task setups, reward mechanisms, and thorough documentation of its training episodes. The method employs an outer RL loop to enhance the agent's performance, achieving a reward score that peaked at around 0.63 after 54 training steps, even successfully transferring knowledge to a previously unseen task family. This work is significant for the AI/ML community as it explores the potential of autonomous model training, representing a step toward self-improving AI systems. The project uses a dual-loop architecture where the outer agent formulates training jobs for an inner model, which learns from a variety of tasks requiring complex reasoning and multi-step interactions. By utilizing infrastructures like Tinker for RL management and Runpod for GPU resources, the project demonstrates how cost-effective scaling of AI training can be achieved. The implications of this research could pave the way for more sophisticated AI systems capable of adapting and optimizing their own training processes, moving the field closer to true AI autonomy.