Show HN: Avoiding the Memory Wall by computing LLM inference directly inside RAM