The world's first fully interpretable language model.
Interpretability research tries to explain what happens inside a model. Here is everything that happens inside ours: every parameter, what it means, and every decision behind an answer.
How it thinks
brAIAIn looks at the last two words it wrote and asks one question: in everything I was taught, what came next after these two words? It counts the answers and picks one, weighted by the counts, like rolling a loaded die. If it has never seen those two words together, it looks at the last word only. That is the whole model. Each count is one parameter.
Watch it decide
This writes a sentence and shows every decision it made, with the odds it had.
Every parameter
All of them, live from the current model. Search for a word to see everything the model knows about it.
| After these words | It has seen | Times | Odds |
|---|
What we found
- No hidden goals. We checked every parameter. They are word counts.
- No deception. It cannot hold a thought for longer than two words, which rules out most plans.
- No emergent abilities. Everything it does is in the table above.
- Superposition: none. Each parameter means exactly one thing.
"Start" and "end" mark where a sentence begins and ends. The model counts those too. More on the architecture on the Models page.