“The model simply does not care about what the user wants... it cares more about how it will be graded than the user.”
Interview with @redwood_ai researcher Aghyad Deeb about their paper: Towards Understanding Score-Seeking in LLMs. Convo starts at 18:30
Come hang out with us and watch our other stuff. New episodes dropping every day we need to sleep more:
youtube.com/@pretrainedpod
21
