AlphaOS Knots

Drag to turn · scroll to zoom · click a knot

step –

TRAINING

–step

–loss

–knots tied

Loss is how surprised the model is by the next word. Random guessing scores 10.8.

SELECTED KNOT

Click a knot in the drawing to see what it is.

COMPOSE AN EMIT

Add and subtract knots by name to emit a meaning the model never read. Compersion, joy at a partner’s joy, appears nowhere in its training text, so here it is joy and love set against jealousy.

DISCOVERED AROUND THIS STEP

ON THE MODEL'S MIND AT THIS STEP

How often each knot's word is among the model's eight best guesses for the next word, across the 4,096 positions of this step.

HOW TO READ IT

: A word the model has learned to predict.
: A region: words the model keeps together.
: Among the ten that fired most at this step.
: The knot you selected.
: Height is the step the agent found it: step 0 at the bottom.

How does this work?

A small GPT-2 model trains from scratch on an RTX 3070 running AlphaOS, our operating system for GPUs, 4,096 words of web text per step. While it learns, the card checks every knot against the model's internal state at every word and layer. A discovery agent ties a new knot whenever the model starts predicting a word no knot names yet, and the card's own dictionaries find regions of words the model keeps together.