Consensus mechanisms don't produce innovation.

Innovation is surprising, outside the distribution, at the edge between consensus and chaos.

Consensus mechanisms can only give innovative results if there's a specific consistent bias in all of the components.

Imagine telling a room of creative people to individually come up with wacky ideas.

Then you take all of the ideas and average them together.

What you'd get is… the centroid, again!

Everyone was being creative, but they were doing it in random directions away from the center, so the average is still the center.

But now imagine you imparted a consistent bias to the creative process.

The bias is now consistent, which means that when you average it all out, you get something away from the centroid, and possibly itself innovative.

This intuition applies to brainstorming with a group of people, but also any time you're using an LLM.

An LLM is effectively a planetary scale consensus mechanism, so it's especially important to give them a specific bias in your prompt to get them to innovate.