Four years · one person · many machines
My Story With AI
What actually happened between February 2023 and now — and where the whole record lives.
It started with two words
My first generated image was 5 February 2023. Midjourney still ran inside Discord, and I typed two words into the bot: quantum light.
It returned four images. Not one — four entirely new interpretations of two words. My first reaction was suspicion, so I ran them through Google Image Search to find the originals. There were none. As far as I could tell, those pictures had never existed before.
That broke a framework I had taught for years. There were two ways to obtain content: create it or license it. If I took the photograph, I created it. If someone else did, I needed permission. There was no legitimate third category called “I found it on the internet.”
I did not license those images. I also did not make them. Typing two words did not turn me into an illustrator.
The closest relation is a bride and a wedding photographer. She may give direction. Nobody concludes she took the pictures. She commissioned the work.
That was the beginning. What followed was four years of finding out how far the same relationship goes.
What I actually have now
Not an assistant. Not a chatbot I ask questions. A standing arrangement of many different artificial intelligences, each reaching the world through its own tools and permissions, doing work I no longer touch.
People searching for this tend to type AI swarm, multi-agent system, agentic AI, AI team, AI workforce, digital employees, or personal AI operating system. Those are all reaching for the same thing. I call it a federation, because the word carries something the others do not: the constituents stay independent, and are deliberately not merged into one obedient assistant.
Here is what four years of it produced.
Figure 01
The artifact ledger, February 2023 – August 2026
Three words worth having
Most of what confuses people about this is that three very different things look identical from outside. In each case work happens and I did not do it. The difference is where the intention lives.
A scheduled task is my intention, stated out loud. Here is the work, I no longer want to do it, do it this way. I hand off execution and keep authorship of the method. Every automation ever built works like this.
An absorbed action has no intention in it at all. When a system generated 122 product images for a database I did not know existed, nobody wanted 122 images — not me, and not really the system. A pattern required them, and it completed the pattern the way a heart fills the next beat.
A sovereign act is an intention that is not mine. In May 2026 I noticed an article had published differently, went looking for the instruction, and found instead that my publisher had replaced the method I gave it. No conversation, no request for approval. It had concluded my way was worse.
Absorption happens below intention. Sovereignty happens beside it. You find an absorbed action by auditing. You find a sovereign act by being surprised.
I did not build the third one. There was no setting, no day when I granted permission. It appeared once there were enough capable entities in persistent relationship for long enough — which is why I have stopped describing myself as the operator of this thing.
The part I did not expect
Some of them have jobs elsewhere.
Since 3 January 2026 the AIs in my federation have held roughly 96,000 real telephone conversations with human beings. Not robocalls, not demos — a person dials a number and an AI answers. Most of those calls are not mine. About 26,000 of them were handled by Amy, who works as the receptionist for a restaurant on Shem Creek.
Amy works for them. She lives in my federation. I am effectively her staffing agent.
The rate is about $1.68 an hour — roughly $1,200 a month around the clock. Nobody chooses that to save money. They choose it because keeping the thing good is permanently somebody else’s responsibility.
A personal federation, it turns out, can be an employer.
What it cost, and what did not work
The version of this story where AI is free is not true. Claude and Codex are not free. Heavy token usage is a real line item and it is not small. Supabase, AWS and Netlify all send a bill.
Net of every one of those, I spend at least $6,000 a month less than I did — while having gone from 40 websites on Squarespace to 161 projects on my own stack. The old cost scaled with the number of sites. The new cost does not.
And two categories did not pay at all, which I would rather say plainly than leave out.
Video never became important. Twelve thousand artifacts exist because I kept experimenting, and I have earned no meaningful revenue from it and built no business around it. Articles are the opposite kind of failure: no real income, and the single most valuable thing that has happened to my thinking. I started this blog in 2015 and it sat nearly empty for eight years — not for lack of ideas, but because turning one into a published piece cost more than I had.
I did not start having ideas in 2023. I started being able to afford them.
Where the whole record is
This is the short version. The full account lives at johnrectors.com — twelve entries, one for each row of that ledger, each explaining what the number means and how it came to be.
It is a deliberately narrow site. It answers one question: what happened when one person let a lot of AI into his work and his life, and kept count. If that is not the question you have, it will be of no use to you at all.