Generation 1 Legacy

GRU 1

The model that proved the idea could work. Small, limited, and somehow still charming.

15M
Parameters
256
Token context
2.7k
Conversations
GRU
Architecture

About this model

GRU 1 was trained purely on 2,700 real conversations with William — no pre-training on any other text, no general language data, nothing else. Just William. That's it.

The result is a model that has genuinely no idea how language normally works, but somehow still sounds a lot like William. It regularly quotes things he's said verbatim or produces completely incoherent strings of words. And yet — in a strange way — that's kind of his style. People who know William tend to rate it surprisingly high on resemblance, even when it makes no sense.

As an actual assistant it's nearly useless, but as a William impressionist it holds its own — sometimes beating GRU 2 on that front. It's kept alive because it's genuinely interesting, and because deleting your first model feels wrong.

A note

GRU 1 is a public shared space — conversations are not private and not saved between sessions. Everyone using it at the same time is talking to the same model instance.