Robin
Besides writing code, having ChatGPT pretend to be Robin from How I Met Your Mother and engaging in a conversation with her is so funny. She is constantly making things up and I am trying my best to make the story work. And when it does, it’s amazing!
It’s even more fun to ask ChatGPT to be Jamie Lannister, since there is (part of) a whole novel and countless online discussions. We fought about whether Joffrey is his bastard for more that half an hour. There is even a duel.
Can’t wait to try this on GPT-4. Maybe I should build my own LLM and put in the scripts of How I Met Your Mother.
But one of my favorite Sci-Fi writers of all time, Ted Chiang, shut my idea down. In the article ChatGPT Is a Blurry JPEG of the Web,1 he claimed that LLMs are not going to create something genuinely new that is not blurry or too general. I am getting the feeling that we just create a robot that produces something trash-ish, which could be useful when you’re stuck in a bug or just wanna have some fun or want to read something as fast as possible2 or want to write papers as fast as possible, but is useless when you are considering the next step of your project or want to talk to Robin herself. In the movie Her, ChatGPT is Theodore rather than Samantha.3
A perfect example of “blurry JPEG” would be this tweet, which describes a process of using GPT-4V to “Describe this image” then using Dall-E 3 to “Generate this image”. Mona Lisa becomes vortex in the end, which is like seeing the nightmare in Uzumaki unfold. (Or sometimes it’s owl, see this tweet.)
Also, according to Giulio Tononi, consciousness is a loop, while Transformer is in one way direction. In other words, it can’t reflect on or debug itself. But maybe, just maybe, will AGI have OCD or other mental problems? Overfitting maybe?