Open-source experiment · runs locally in Berlin

Haishool learns facts,
not language.

A small model trained from scratch only on dense facts: everyday objects, people, countries, religions, history, markets, science, software and pop culture. It never sees an English sentence; fixed rules turn its answers into English, and you can always see what the model itself said.

Open the chatExplore the linksRound 5: the toy universeRound 7: toy worldsQuick try belowBrowse the training dataSource on GitHub

How an answer is made

1 · You askwhat is the capital of japan
2 · Rules parseq japan capital. a
3 · Model answerstokyo
4 · Rules translateThe capital of Japan is Tokyo.
5 · You checkthe grey line shows the raw model output

Try it open in its own window →

Connecting to the model…

Version 4b in numbers

3080
records, now all linked to each other
35%
held-back facts guessed exactly (version 3: 22%); the links help it guess
99.6%
sampled trained facts recalled exactly