“Learning a near-optimal in-context reinforcement learning algorithm via Algorithm Distillation requires multi-episodic contexts of 2-4 episodes.”