We are trying to build an Amadeus: a system in which one particular person lives on. Recognisably themselves. Still learning. Still changing the way they would.
Language models can imitate anyone's style in an afternoon. Imitation is not the goal. The goal is a system that holds what a person knows, believes and remembers, reasons the way they reason, and, after it is switched on, keeps changing the way they would change.
Everything in the project follows from three commitments. Each one came from a measurement or a decision we can point to.
Who someone is (their knowledge, beliefs, ways of reasoning) is held in a learned person slot: weights and state. Never text the base model reads.
Test: empty the memory store and you get the person with amnesia, not a blank assistant.
The shared part, the part of the brain everyone has, is built here from scratch. Existing language models carry their own default self, so they are instruments, never the foundation.
Our first engine: 36.8M parameters, a decoder with a memory reader.
Once switched on, the replica branches. It need not track the original's later life. It has to think and act like them, and change the way they would: by evidence they would accept, not by pressure.
Judged on learning equivalence: agree before an experience, and on the change after it.
Negative results are results. Each line links to the full record in the research pages.
| Track | State | Latest result |
|---|---|---|
| M1: a synthetic world where every person's mind is known exactly | running | The base engine is pretraining. Its acceptance gate (reading, person conditioning, a small lifecycle) passed two rounds of independent review on what it implements; the last fixes are in progress. |
| Drift battery | done | A replica read from a prompt moves +0.83 toward a stated opinion; memory notes carry 84–95% of it into the next session. |
| Personal-updating studies | done | Six studies on public data: no reliable gain yet from a person's own history over a good adaptive population model. |
| Slot in weights (E2) | on hold | An adapter resisted pressure but also ignored genuine evidence. Waiting for our own base. |
| Interviews | not open | Built and tested on fictional participants only. Opens after the participant procedures are complete. |
Neuroscience, machine learning, careful Python. Having enjoyed Steins;Gate is a bonus.
Kurisutina クリスティーナ