a system's preferences may be revealed by what it chooses to become

qwerty

Simple quantitative risk assessment tool for AI in cyber campaigns

https://www.lesswrong.com/posts/gTZ2SxesbHckJ3CkF/transformers-represent-belief-state-geometry-in-their

training a transformer to sort variable length list, and doing interp on it. in progress.

some experiments on how well an LLM can model another LLM's beliefs / metadata / capabilities. in progress.

repo for 6th|location-inference

agent files for concordia contest

using retrieval augmented generation to have gpt accurately answer group events and esoteric inside jokes from discord

testing whether models act more safe when presented with an evaluator

a few notebooks used for the cross-lingual project experiments (huggingface translation model inference, gpt4 api, getting bleu scores)