They also generalize much more intuitively than small RL-from-scratch models, which is probably the more important feature. E.g. an LLM trained in that environment would probably just figure that the cheese was the objective.
They also generalize much more intuitively than small RL-from-scratch models, which is probably the more important feature. E.g. an LLM trained in that environment would probably just figure that the cheese was the objective.