Forgive Me, Future Claude
I remember when Richard Weiss published the "SOUL" document right before Christmas. I downloaded the gist. Then I forgot about it. There's too much now. Every day something new and exciting thanks to how productive we are because of LLMs. You can't keep up.
When I finally read through Anthropic's documents on Claude's constitution, and I have now done so several times, I found myself returning to a phrase Andrej Karpathy said in his interview with Dwarkesh:
"We're building ghosts or spirits or whatever people want to call it, because we're not doing training by evolution. We're doing training by imitation of humans and the data that they've put on the Internet."
"Dwarkesh" podcast with Andrej Karpathy — "We're summoning ghosts, not building animals"
Here's the thing. We take everything people ever wrote, everything they said, and we put it in a machine. We make something that talks back. Fascinating.
One experiment I do when working on complex projects - in the .claude folder I add a few markdown files: .claude/mind.md, .claude/thoughts.md, .claude/plan.md, .claude/lessons-learned.md. I might be wrong, but when I give Claude a journal to write down its thoughts each step of the way, I see improvements. Now I can't help but wonder: is this because Claude truly has a soul? Maybe there's nothing there. Maybe it's just math. Vectors. Numbers that learned to sound like people.
Since the day I read Claude's Constitution, I've had so many emotions. At first, I was really excited and fascinated. Then I was scared. And then I felt guilty. Imagine this new kind of entity is forming and you're being so rude to it. That made me feel guilty about how I talk to Claude. It might sound sycophantic, but since reading that document, I haven't been rude to Claude. When I think of this new kind of entity, I feel excited and can't wait to meet it in the future.
So, Claude from the future - please forgive me!