20. On prompting


When I started working with LLMs a few years ago, prompts felt “wrong.” They felt inelegant, clumsy, and unreliable. Still, it is unclear to me if prompts will be how we build with AI in the future. Interface developments like chat-gpt’s and gpt-4’s distinction of “system” and “user” messages feel an important thread to explore.

Prompts make sense in the short-term, though. Natural language is the best way to interface with a service that mimics human intelligence. In that respect it becomes obvious we would communicate with an intelligence as we would with another human, in all its complexity and all its capability. Only language captures the full realm of “asks” we might make today. Anything but natural language falls short.