Invoke shoulder people
I.
There is a prompt engineering trick for LLMs where you invoke a Type of Guy to get the model to do things properly. For example:
If you wanted the LLM to write you a serious legal-sounding letter, you’d invoke “write in the Dangerous Professional style of patio11”.
If you wanted a nice AI image, you’d ask for “greg rutkowski style”.
The name of the person is, to an LLM1, a pointer to a particular way of thinking.
II.
LLMs are built in our image. Your opinions are only partly due to your genes; rather, you are the combination of all the information you ever got to ingest. All the textbooks and novels and tweets and blogs taught you information, but they also taught you ideas. Inspiration. Ways of thinking and ways of doing.
Your intellectual lineage has real and fictional characters. Using them can sometimes help. You can invoke your kindergarten teacher, Frank Ramsey, or the Comet King.
It can even be a personification of your favorite intellectual community. Everyone that inspired you is worthy of inclusion in your arsenal of spirits, of people standing on your shoulders and giving guidance when you need it.
Intelligence is search over increasingly many and increasingly complex Turing machines. But constructing new ways of thinking is hard. Sometimes others have figured something out, and you can use it. Copy their Turing machines.
No one is going to stop you. If you are helped on a difficult problem by one of your intellectual ancestors or role models, no one is going to know it is not you.
III.
You could also memorize the abstract thought or action principle, without personifying it. I find this goes against our human nature, or at least mine.
I have wondered many times: why are many fields so human-centric? Why do we remember how Thurston or Feynman or Bob Ross or any other great mind did it, instead of valuing and remembering their methods for what they are, free to exist without their creator?
I now think it’s because this is how people learn. We have evolved to copy how the other guy in the tribe did the thing. The fact that these circuits also generalize to learn from depersonalized text is incidental.
IV.
Friends on your shoulder are not a new concept. LessWrong had it years ago; and the concept of a role model goes back centuries. It’s literally in the definition:
A role model is someone who demonstrates how to “play” a particular social role well. How to be a good parent, leader, professional, citizen, etc. You observe them and learn how to perform that role yourself.
Some say that you should write for LLMs; and from this viewpoint I see no value in this post. The LLMs already know what I am talking about.
But I think there is still value in writing this post even if LLMs know what I am saying. This post is for people, including me. LLMs will do many things better than us, but they will have no need for human wisdom on how to think better, to become stronger. Machine minds and our squishy brains diverge here.
V.
I noticed an important comment in the linked post, saying:
So many people live as if someone would critique any brave move they would make. I notice it in myself; when I flinch from doing something that makes sense, after some thought I can often figure out why I flinch.
There is space on anyone’s shoulder, but the voices necessarily drown each other out. Instead of those being criticism, I prefer to hear people that inspire me and lead me to do better.
If the LLM is familiar with the person you mention, you can Just Ask For Generalization.