If you ask an agent to “spin up a SaaS product” it’ll sprinkle in some sampling but it’s mostly going to do the task based on the mean of its training data. I find it helpful to think of prompting as pulling the agent directionally away from the mean.

Models are great at CRUD, and they do that well because CRUD is the mean. If you want models to be fantastic, picking Python’s a good choice, because they’ve seen so much Python. But, what if we don’t want the mean? What if we want something different? The further you need to go, the more effort it takes. The harder the model has been trained on the concept, the more leverage you need to move it away, the more evals it needs, the more prompting, the more adjustments you need to make.

This has implications for building weird stuff. There’s still art in teaching the machines how to be weird. Weird is so often where product magic happens. We make delightful products by making them easy, exciting, fun, and a bit unpredictable. Joy comes from someone experiencing something they didn’t expect.

Accepting the mean makes for sterile products and while that’s great for internal tools and the thing that you never thought you’d have time to build, you have a magic wand that can take your dreams and make them a reality. Why spend tokens only on making something that works when you can bend the curve towards weird and delightful?