Earlier this week I gave a talk about some of my current thinking about generative AI. One of the points I made is that while these tools have tremendous power, we are still scrambling around looking for the best ways to invoke that power. This reminded me of the command line. With the command line, you need to know the right incantation that you have to type, in order to get the tool to do what you want. It can sometimes feel a bit like casting spells. (Sam Arbesman writes beautifully about this https://arbesman.substack.com/p/-coding-as-magic).

Language models are a little like that. You have to figure out the best kinds of prompts to ask them. Ask them to act as an expert, they perform better. Add in emotional triggers, they perform better.

After I gave my talk I noticed that the terminal that I'm actually using these days - warp (https://www.warp.dev/b), actually integrates with an LLM. When I can't remember the right command line query, I can ask warp, in natural language, what I want to do, and it translates that into the right command. It has been a really nice productivity gain for me.