How to Get Language Models to Do Things Like Chocolat One Up?

A paper from DeepMind that came out earlier this month, titled Human Time Scale Adaptation and Open-Ended Task Space. Here, they show that you can actually have the agent learned by trial and error very efficiently. So it's nice to see some more progress in this reinforcement learning and real embody the agent space where you don't have a passive model,. You have something that continuously interacts with the world to accomplish your goal.