Cache-to-Cache: Direct Semantic Communication Between LLMs to Design Its Jalapeño Chip

But why would you want to be able to communicate with each other in Neuralese. It's not looking good for monitorability.