“Dima from Fireworks advocates that companies building specialized models should use their own production environment for reinforcement learning to ensure the training environment is as close to reality as possible.”