Mapping ChatGPT’s ontological landscape, gradients and choices [interpretability] — LessWrong