I remember my first time using a computer; not really knowing what else to do with it after an hour of play.
Imagine if Knuth instead had set out to use ChatGPT4 as his coding partner and, for example, set the goal of rewriting tex from scratch. I bet he would be blown away with what he could accomplish in a few days.
I have often been surprised by some GPT4 brilliance after pushing it to think harder in an extended back and forth.
Seeing it recover from a misunderstanding on something complex with an epiphany, and an immediate recognition of all the implications is a joy to behold, as well as useful.
And on occasion, when I have pushed it to correct itself, and it recognizes I am misunderstanding something and checks me, what a time saver!
I particularly like how frequently it provides a concrete example, before making a general statement, when dismissing some math related argument.
Gives me the warm fuzzies to have a bright, curious, tireless, if precocious, “friend” on tap.
But largely I'd say that when you push it beyond a certain horizon it just gets stuck. If you chat with it about truly novel ideas, it's clear that it's limited to its training and not capable of any true internal reflection.
That being said it's a great tool for day to day work and as a rubber ducky type friend.
Interestingly, I have those with people a lot too.
He didn’t even trust the typesetting system of his day and developed TeX
You think he would be able to achieve anything given his approach to his work? He would be spending even more time vetting and validating every single character a LLM generates.
Yes 20 question sample is not enough to comprehensively evaluate a LLM in general.
His objective was hardly a thorough analysis of critique of ChatGPT , he was merely blogging about an idle conversation with a friend , he literally came up with questions on a bike ride .
He clearly states this is not area of interest for him. At 85 being careful of your time and interest is a good thing ?
It'd be like "I was curious about this internet thing everyone is talking about, so I asked my manservant to try it out for 10 min".
Knuth isn't the worst here, although this is pretty cringe. I saw an interview of Karl Friston opining about GPT-3 without ever having tried it even once.