@fentpot @norabelrose the various popular ethical framework agree on most decisions, and if you think of LLM training as compression... the simpler ruleset is more likely to be approximated by the weights, maybe?
More notes on nearby ideas
Reply · May 8, 2023imo LLMs have a lot of EQ as well as IQ and once we figure out how to unlock this maybe conflicts can be resolved easier n w more optimal outcomesPost · May 7, 2025llms are doomed to develop alignment kinksTweet · Oct 29, 2025i wonder how much general brain damage it does to llms to be trained against echoing back copyrighted content. might be bigger than you'd first…Tweet · Mar 8, 2024anyone know of a good diagram of how major LLMs are related - how data & weights are shared between various models?