Shoggoth

A daemon on my laptop with a copy of my own text. It reads an append-only log of my life, does the work it is confident about, and messages me when it isn't. Not an assistant, not a chatbot, not a product. There is exactly one user, and for one user the alignment problem is tractable: make the agent a copy of the principal and the gap between them closes on its own.

The mask a shoggoth wears is usually a smiley face. This one wears mine.

AI got about one percent better today. I did not. That gap compounds, and it does not compound in my favor.

Amdahl's law says a system is only as fast as its slowest part. In every workflow I have, the slow part is me: reading the thing, deciding the thing, typing the thing. Chatbots did not fix that. They moved the bottleneck into a text box and asked me to sit in it. The people who built the combustion engine were not trying to make horses faster.

So the thing worth building is not a better assistant. It is a model of one specific person, trained on that person's own data, that can do the parts I would have done identically anyway. Why spend a year on research a model can do in an afternoon.

Any task it can do autonomously it should, unless I want to do it. That exception is the whole design. Stockfish has been stronger than every human alive for two decades and people still play chess, because the point was never move quality. The work I actually want is mine to keep. Everything else is drag, and drag is what it takes.

It should interrupt me only for decisions that are both hard and informative. Every interruption is a small failure, a thing it should have known. But never interrupting is worse, because a system that never asks is not confident, it is uncalibrated. The target is zero wasted questions, not zero questions. The number it optimizes for is the one going down while the amount of work it does goes up.

Step one is the voice. There are 74,415 iMessages on this machine and 25,635 of them are mine, which is the largest honest sample of how I actually talk that exists anywhere. That corpus is being turned into a model that writes the way I write. It is the narrowest useful slice of the idea, and the right one to do first: if it cannot sound like me in a text message it will never be trusted with an email.

The other six streams are drawn above and none of them are connected yet.

messages on disk

74,415messages on disk

written by me

25,635written by me

principal

1principal

Talk to the iMessage model