Interesting, reading GPT output with an expectation of human quality made its shortcomings a lot more obvious to me. Thank you for that experience. In particular, it stood out that significant parts are an oddly direct regurgitation of input text, suggesting they didn’t undergo translation into abstract form then back to text, and that it doesn’t pick out interesting or unexpected things to respond to, instead coming up with some comment on many parts of the input like it’s a checklist.