Not a language model, and not a small copy of one. Sparse predictive coding in plain numpy: no backpropagation, no rented cluster, under a megabyte on disk. It thinks on a CPU — on a single core.
It only speaks German — that is what it was raised on. Your conversation stays yours: every visitor gets their own memory, held in RAM, never written down.
One core, no accelerator, under a megabyte. Something that needs a data centre to think is not the thing we are trying to build.
A new class does not overwrite the ones before it. That is the whole point of the architecture, and it is measured rather than asserted.
There is no branch for not understanding, so it always answers — sometimes badly. An honest wrong answer tells you more about a system than a polished one.
Every visitor gets their own memory, in RAM, for one session. Nothing is written to disk, so there is no store that could leak.