Analyzing abstractions, semantics and human language is a sorta famous research direction in alignment. (The researchers are John Wentworth & David Lorell & Gretta Duleba, Thane Rutheins, Jan Kulveit, Sam Eisenstat.)
So I wrote a draft where I tried to
classify ALL interpretable types of human abstractions,
explain where the usefulness of different abstraction types comes from,
explain how polysemantic words work (I analyzed “run”, “set” and “put”).
The idea is to give a fully general definition of what it means for a set of abstractions/computations to be “interpretable”.
I’d like to get feedback, because the topic is complicated and I don’t want to fumble it. If you want to give some, feel free to message me in LW DMs/on Discord or comment here/in the doc.
Analyzing abstractions, semantics and human language is a sorta famous research direction in alignment. (The researchers are John Wentworth & David Lorell & Gretta Duleba, Thane Rutheins, Jan Kulveit, Sam Eisenstat.)
So I wrote a draft where I tried to
classify ALL interpretable types of human abstractions,
explain where the usefulness of different abstraction types comes from,
explain how polysemantic words work (I analyzed “run”, “set” and “put”).
The idea is to give a fully general definition of what it means for a set of abstractions/computations to be “interpretable”.
Here’s the draft:
https://docs.google.com/document/d/1Q9Bvtt2xWYDMUm-DnlmAxFp5M6uGWiCsYilDiwtLZ20/edit?tab=t.0
I’d like to get feedback, because the topic is complicated and I don’t want to fumble it. If you want to give some, feel free to message me in LW DMs/on Discord or comment here/in the doc.