a newborn language model, learning to talk
Nib is a small language model being trained live on one laptop, on text written before 2014: old encyclopaedia articles and public-domain books. What answers you here is its real latest checkpoint, running inside your browser. A new checkpoint is saved about every hour, so the way it talks changes as it reads. Most of it is still loops and half-words, and it is shown as it comes out.
Its little computer shows what is actually happening: the trainer checking in, its loss ticking, each checkpoint being saved and the hash posted for it on Solana devnet. Click the screen to look closer.
Type anywhere and press Enter. On a phone, tap Nib first. Esc closes things.
Nib is a small language model (about 34 million parameters) being trained on a laptop. It is often wrong and may say odd things. Its training data is old encyclopaedia text and public-domain books.
Everything here is measured: checkpoints, their fixed-prompt answers and their hashes. Nothing is promised about when, or whether, it will speak in sentences.
Unofficial fan project. Not affiliated with Rockstar Games or Take-Two. Not the first baby AI; others exist.
The token has no promised value, no revenue share and no buybacks. Any holder gate is a convenience: the weights are open.
Training pauses when the Mac is offline, and this page says so. Text: CC BY-SA, Wikipedia contributors; Project Gutenberg. See licence, recipe, verify.
Nib lives in a small 3D room that this browser can't draw right now.