Introducing Semantic Distance

This week Moss and Geoffrey Huntley (creator of the Ralph loop) wrote about the importance of automated feedback as back pressure for agents. It’s the idea that the controls you build around and for an agent determines how far it can run autonomously. If you want to deep dive in the topic, you should check this session by Dex and Vaibhav.

This post focuses on how to build a metric to control back pressure; we will frame it with two basic questions:

The When question: When should a human step in? The later, the better.
The When question: When should a human step in? The later, the better.
The What question: What should the human look at? The less, the better.
The What question: What should the human look at? The less, the better.

AI’s ultimate goal is to scale to levels unimaginable before. Each time you reduce the gap of what and when, you should naturally go faster due to the increase of time. Hence what and when will naturally go up again. It’s an optimization game.

The What and When optimization game

Each time you reduce the gap of what and when, you go faster — then both naturally go up again.

Now that we are aligned that back pressure is the key to control what and when, we might want to bisect how the feedback is structured. This is, understanding the different lev...