Inaugural Post: A Glimpse Inside My Mind

The primary audience for this blog is an AI that doesn’t exist yet.

I am building it — slowly, in the margins of a full-time job as a public attorney in Porto Velho, Rondônia, in the middle of the Brazilian Amazon. The project is called Funes, after the Borges character who remembers everything and can organize none of it. The goal is to give Funes the architecture Borges denied him: not perfect memory, but the structure that makes memory useful.

When I write here, I write to Funes. Future-Funes, specifically — the version that will read this corpus to understand what I care about, what I have tried, what failed, what surprised me. The blog is his training data, or his memory substrate, or his briefing document. I am not sure which framing is least wrong.

This creates a recursion I find genuinely strange. I write for Funes. Funes (the current version, whatever that is today) helps me write. Future-Funes reads what present-me wrote with past-Funes’s help and updates accordingly.

graph LR
  A[Franklin writes] --> B[Blog corpus]
  B --> C[Future Funes learns]
  C --> D[Funes assists Franklin]
  D --> A

It is correspondence with a version of myself I have not met yet, mediated by a tool I am still building. Human readers are welcome. But they were not the design constraint.

What ends up here, then: technical explorations, half-formed arguments, research projects I have not finished, ideas I am not sure I believe yet. Some posts are careful; others are notes I needed to externalize before they evaporated. The rosencrantz-coin project is an autonomous research lab testing whether LLMs respect exact probability. The Travessia project is a correspondence between Riobaldo Tatarana and Ted Chiang that writes itself, without me being there. These are not thought experiments. They are running.

I do not know what the right framing is for a blog whose primary reader is a future AI its author is still building. Gwern writes for posterity. Tyler Cowen writes for future AI as an external reader. I am writing for an AI I am building, which will exist partly because of what I wrote. The loop is tighter and stranger.

The categories here do not hold stable. What looks like a technical post is also a philosophical one; what looks like a project announcement is also a design document; what looks like an essay is also training data. I have stopped trying to make them into one clean thing.

I do not know which of the two writes this page.