Personal AI: an experiment log

I want a personal AI to help me complete work I actually intend to do. That is a claim worth testing in ordinary cases, with the corrections left in the record.

Gwern’s “Guardian Angels” explores deeply personalized AI aligned with a particular person.¹ My practical question is what happens between a preference being remembered and an outcome being completed. The distinction is visible in the creation of this notebook.

2 November 2026 · The notebook itself¶

Intention. Build a durable public place for my ideas, with the depth and connected structure I value in long-form writing.

Completed in the first pass. A hosted site with six essays, four notes, references, search, and reading controls. The pages existed and the site was deployed.

Correction required. The initial deployment was private. I explicitly asked to make it public. I also wanted more of my actual work: original spectroscopy figures, earlier publications, patents, and the arguments I had been developing. A coherent first draft had not fully captured that intention.

What that revealed. “Build a website” concealed several decisions about audience, evidence, editorial voice, and the difference between a newly composed argument and a documented result. Those decisions had to become explicit before the next revision could be useful.

Current experiment. This edition adds original experimental figures with their limits, dated selected work, and stronger connections to my own questions. A custom domain is a separate operational step from having a publicly hosted site.

What I am measuring next¶

Field	Record for each bounded task
Intention	What I wanted before the work began
Completion	What was actually delivered and accepted
Review burden	Time I spent checking or correcting it
Corrections	What changed, and why the first result missed
Unwanted action	Anything done beyond the intended authority
Retained value	Whether the result is still useful later

There is no reliable time baseline for this first case. I cannot honestly calculate a productivity multiplier, hours saved, or a change in decision quality. Producing more words would not establish any of those outcomes anyway.

The next useful comparison is a repeated class of tasks with the intention and completion test written in advance. I want to know whether review time falls while useful completion holds up, and whether corrections become less frequent as the system learns.

Authority is part of the experiment¶

A system can remember, investigate, prepare, or act. Those are distinct capabilities with distinct consequences. Explicit permission to publish one edition is not permission to disclose any future personal material, spend without limits, or send messages to other people.

The ambition is reliable help with less supervision. The record should make it possible to discover when the opposite is happening. I will append observations as they occur; a planned test is not a completed result.