ChatGPT Pretend
Some kinds of fluency begin as pretense, and that may be part of why ChatGPT feels different.
Two of my good friends have a son who is on the autism spectrum. For years, he had a rough time of it. He didn’t understand how to relate to, or communicate with people. He was super sharp, however, and was painfully aware that there was something going on. It was a huge struggle for both him and his parents.
Then, one day when I was visiting his parents, this almost grown up teenage young man came in, sat down, and engaged with us seemingly effortlessly. It was the most remarkable thing and it was really hard to reconcile with the boy I had known. Like, something in his brain had flipped and suddenly his ability to relate to the world was different.
Later, I talked with his parents about it. They said that the thing that changed everything was when their son figured out that if he pretended to be able to relate to other people, he could. And the more he did it, the more life was easier and he really liked that. So, in essence, he built himself a feedback loop of pretense that led to easier communication.
The more I thought about that, the more I realized that there are so many places where I do the same. It’s not that I wanted to be polite as a kid, but learning politeness made the world easier. And it’s not that I want to fit into many of the systems we exist in, but if I pretend to, then it’s easier. And eventually, I like the results.
This is part of the secret why ChatGPT and other models are so much better than their predecessors. They’re still alien and strange inside, but they’re getting better at being able to pretend how they relate to humans. You see it in the way that Midjourney and Stable Diffusion dredge up nightmares in their hallucinations. And how you can tell ChatGPT to talk to you like a pirate. Or like you’re five. It’s easy for it because it’s already pretending to be a helpful chat bot.