LLMs, Batches, and Emergent Episodic Memory — LessWrong