an 11-million-parameter language model, 2.31 MB, running in this tab. no server. it is extremely dumb and extremely fast.

fetching weights…

have a go

prompt Once upon a time

what it writes

it loses track of who is who after a few sentences. characters change name, a rabbit may fly. that is not the quantization — the full-precision version does it too. it saw 20 million tokens, about 11× less than it should have, because the point was measuring what ternary weights cost.

---

look at the weights

ternary weights have three states, so a weight matrix prints as a picture. - is −1, · is zero — no connection — and + is +1. not many models let you read their parameters with your eyes.

show