A thread through the archive
Essays tagged “AI training”. Follow an idea across the archive.
Scott discusses three recent papers showing surprising patterns in how AI misbehavior does and doesn't generalize across different contexts, from emergent misalignment spreading across domains to reward-hacking staying confined to graded tasks.
Scott argues that AI 'hallucinations' should be called 'shameless guesses' because they work the same way as students guessing on tests - making their best attempt when uncertain rather than admitting ignorance, revealing an alignment problem.
Scott explores three approaches to 'writing for AI' - teaching knowledge, influencing beliefs, and enabling simulation - finding the first limited, the second theoretically confused, and the third creepy and ethically troubling.
Scott Alexander discusses Janus' experiments with GPT-3, exploring its capabilities, quirks, and potential implications.