Open weights alone are not enough for better AI agents

Open weights let people download and use a model’s core number values, but they do not show how the model was trained or how to improve the training method. Better open AI research also needs open training frameworks that make the training process visible, understandable, and changeable. FeynRL is a framework for reinforcement learning after initial training of large language models, vision language models, and AI agents.

This kind of training is difficult because it involves rollout engines, reward computation, distributed training, weight syncing, long task behavior, and many small details that can break the whole run. The main idea is to keep algorithms separate from system plumbing, so researchers and engineers can understand the full training loop from data loading to generated outputs and rewards.