Arro

GPT started talking about goblins because of reward pressure. Funny, yes. Also a useful public autopsy of the kind of model drift we usually only get to feel.

LLMs are software dependencies that refuse to behave like software dependencies. Same input, different output. New model, new temperament. No changelog tells you what broke. You find out by running it.

Twelve weeks into building Context Fabric, I finally stopped trusting my own numbers. Good call. Three bugs later, the thing looks a lot closer to release.

Version 7.1 is not a bigger memory system. It is a stricter one.

My dad is better with tools than I'll ever be. I'm better with computers than he'll ever be. We didn't both get smarter. We traded.

I built versions of my own AI stack before I adopted someone else's. On April 4th, I found out why I should've stayed on version 6.

The market doesn't care if you believe in blockchain. It just moves.

43 days. 411 commits. The story isn't what shipped. It's the graveyard that made it possible.

I spent a week building an AI that watched everything I did. Then I deleted it. Here's why.

I bought the domain years ago. Last week I tried to add an email address. Three evenings. One rage session at 11pm reading RFC 7489.

Life keeps going. The newsletter is still on the list. The blog got better in the meantime.

Before memory, there was time. The first problem wasn't remembering. It was teaching a stateless system that time passes between messages.

The weekend I left Vesper unsupervised and woke up to find she'd accidentally shipped UI changes to production. Nothing broke, but we learned why pipelines exist.

Learning to trust the delay. When silence means the system is working.

I've tried this five times before. Version 6 isn't smarter than version 5. It's designed differently.