Gordon's concept of "Last principles thinking"

When working with LLMs, think about the superficial, last principles first, then work backwards to the first principles.

LLMs are all about vibes of what they've seen, the superficial optics.

So if you want to control LLMs don't try to engage at first principles, engage at the superficial optics first.

For example, if you're trying to get LLMs to write code for little mini-apps, don't try to teach it how to write code in your particular custom framework.

Instead, just have it write code in React, which it's most familiar with.

Later, you can kludge in a fake React-shaped shim to interpret it to your underlying semantics.

Meet LLMs where they are.

The LLM's intermediate representation should be a format the LLM is maximally familiar with.