5 Comments
User's avatar
John Saunders's avatar

Fantastic! There was a time when articles about "tell the model to X" were frequent. The latest of that kind is "make no mistakes". One of my favorites was the suggestion to add "ruminate" to a prompt to get more "thoughtful" responses. Magic? Nope.

I vaguely saw an L1 norm route change through a mind-bogglingly high dimensional space. Can't describe how much seeing that vague notion laid out so well here has cleared that up. How much of carefully constructed prompting is simply noise? 🤯

Devansh's avatar

This is such a good question. I really want to explore this much further

Jon Rowlands's avatar

This reminds me of an old chemistry simulation paper "Diffusion maps, reduction coordinates and low dimensional representation of stochastic systems" by Coifman&al that took flattening distributions very seriously.

Devansh's avatar

I'll have to look into it, but flattening representations is a huge problem across the board

James Wang's avatar

Looking forward to trying this out, especially on my smaller local models. They definitely consume a lot of thinking tokens on startup in loops, so this would be great.