Open weights are not enough: we need open training frameworks for research and better algorithms [P]
A new open-source training framework called FeynRL has been developed to address the limitations of merely having open model weights. The framework aims to make the AI training process more transparent, understandable, and modifiable, enabling researchers to develop new algorithms more effectively. FeynRL separates algorithms from systems, providing an end-to-end view of the training loop for LLMs, VLMs, and agents, and supports various training methods and hardware setups. AI
IMPACT Enhances transparency and algorithm development in open AI research by separating systems from algorithms.