9 November 2025

I'm working on building a blog-writing interface. I don't want to generate text like in Loom (well maybe just a little), because the fancy language machines are not actually trained to write! They are trained to read. 

I lost access to the GPT-4 base model back in May, so I'm using davinci-002 for now. I plotted the perplexity of each post (as it reads them in order), and it identified my two most predictable posts so far:
1. The recipe for pizza (saw that coming)
2. My chaotic ramble about crackpots (heh??) 

I'm surprised by davici's lack of surprise. 

I aim for the edge of chaos. Probably I should also normalize against GPT-2, to find the right kind of chaos. I could also use a classifier to ban words like "delve". What I really want is something to help eliminate evidence of gender, but anything I can think of right now feels way too hacky. 

I need it to push me. If derpy little davinci can tell me what I'm going to say before I say it, that's not a good sign! I want nudges to explore paths I wasn't considering, and constraints that help me think. We live in cages of our own cliches. I want to reach out and feel the shape of them. 

I was moved when I heard Karpathy say that he believes one of the main things slowing the mode collapse of our minds is interacting with other people. 

Maybe I should just get out more, instead of building fancy toys. . .