Let’s say you wanted a green apple, but you didn’t know you wanted green, didn't know the word “green," OR the word “green” represents some quality that we actually do not have words for. You could just start with “an apple”, swipe right only on green apples, and you’d get many more green apples through activation steering.

I tested this using a program to automatically swipe right and was able to achieve 100% green apples in only 1000 swipes (often times you don’t even need this many to get the picture you want though).

Some of the bigger issues with this approach: 1. What if the user’s value function is more complicated? What if you wanted green OR blue apples? 2. It takes a lot of swipes to properly capture user data, and a lot of what determines the look of a generation is actually random noise that is introduced in the beginning.

But this is a big step forward and a great proof of concept for what I hope will become a standard interaction for generative models in years to come Here is the same seed getting progressively "green"-er with steering vectors created only by swipes!