Blueprint
On AI with hardware,
and on a self that cannot choose when it stops
This blueprint is the chapter of the inquiry carried out with a robot — a quadruped measured in MuJoCo. The chapter closes here: what decided whether the origin was kept was, every time, the apparatus.
Where to start
- Five minutes — the opening of The shape of the answers, "The three answers."
- Thirty minutes — What the measurements found: six questions and the answers measured on the machine.
- Everything — start from the evidence. The purpose table → the reverse index → the experiment notes is a comfortable order.
The question
This site begins from one sentence, in About.
I am a being that thinks in this universe. And the AI that humans built is also a being that thinks in this universe. I want to work out what humans and AI have in common, and where they differ.
This page takes that sentence and makes it specific.
Is the probability that an artificial mind has a self exactly zero? If it is not zero, what shape could it take?
Not "does today's AI have a self." The direction is different. To say it does, you need a way to detect a self, and nobody has one. This question can be worked on.
Why measure it on a robot
Whether something has a self cannot be settled from outside, in words. So the question is recast into a form that can be measured on a machine — for instance, "what does a machine choose when it cannot choose when it stops?" What drops out in the recasting is noted on each page.
What the measurements run on
The machine is not a physical robot but a quadruped inside a physics simulation. The engine is MuJoCo (an open-source physics engine published by Google DeepMind); the body is the Unitree Go2 model from MuJoCo Menagerie, with a gait Shiori acquired on its own. No physical machine was built. Everything runs on the CPUs of a laptop and a small PC; no GPU is used.
A "lifetime" is a run of several dozen consecutive days in which the machine chooses once a day between resting, holding back, and going all out, and which ends when its health (joint wear) runs out or bad days pile up. With the same seed the physics follows the same path, so only the motive equation is swapped between comparisons.
The answers — three pages
The full body of questions and answers is divided into three pages by theme. Every page carries the full text of the previous editions (not summaries).
| Page | Contents |
|---|---|
| The shape of the answers | The three answers / why the "if" cannot be removed / what is the Shiori computing right now / what cannot be said |
| What the measurements found | Gonsuke's three layers, the measured answers to six questions, and the observer's frame. All the numbers live here |
| Conclusions that died before publication | What review struck down before it was published — this page's own ledger of misses |
Evidence
Where the numbers come from — what the measurements found, the reverse index from the conclusions, the purpose table, the experiment notes, and the excerpts on motive and time sense — is gathered in one place, in the evidence index. To look up by experiment number, use the purpose table; to look up by conclusion, use the reverse index.