The intelligence
PAIMON
Model 1 of machine teaching. Built from scratch, taught subject by subject, tested after every lesson.
Why
Because a scraped model cannot tell you what it knows.
Models trained by hoovering up the internet are impressive and untrustworthy in the same breath. They cannot show where an answer came from, they cannot be corrected in a specific subject without retraining everything, and they carry whatever bias happened to be in the pile.
We wanted a system with a transcript. Something that can say: this is the curriculum I was given, this is the lesson I learned it in, and this is what I have not been taught yet.
How
Curriculum first, model second.
A subject is broken into lessons. Each lesson is written by someone who knows the material, taught in a batch, and followed by a test drawn from passages the model never saw. If it fails, the lesson is rewritten rather than the test being lowered.
What passes goes into the memory bank. What fails goes back to the curriculum. The scoreboard is the held-out test, not the training loss.
The goal
The computer on Star Trek.
Ask it anything about the ship and it answers, correctly, in a sentence, without making you learn an interface first.
That computer was never a search engine and never a chatbot. It knew the vessel it lived in, it knew what had happened aboard it, and it answered the person standing in front of it as though being asked were the most normal thing in the world. It was trusted, which means it was right, and it was right because somebody had taught it the ship.
That is the company goal, stated plainly. Not a model that scores well on a benchmark, but a machine you can simply ask, and believe.