Note 09
We looked for hidden instructions in 28,796 AI rule files. We found none.
Coding agents follow rule files that humans review but cannot fully see: some Unicode characters are invisible on screen and readable by the model.
Articles
Research notes with their evidence attached, and the occasional piece about building the lab itself. Failed predictions stay published.
Note 08
It did not say yes. It did not understand the question.
Open models refuse harmful requests in English and go along with them in Somali.
Note 07
The frontier can read Somali. It cannot judge it.
We showed 593 community-validated Somali submissions to the newest models from OpenAI, Google, and Anthropic and asked the question our volunteers answer every day: is this good Somali, or not? The models placed the texts in the right domain two-thirds of the time.
Note 06
One model reads Somali. The next one approves word salad.
We gave three open models the easiest version of our community's validation job: tell a real Somali sentence from the same words shuffled.
Note 05
Every sentence has a name
We released the first version of an open Somali corpus.
Note 04
The lie moved. The liar did not.
We trained a lying model against two detectors blind in different ways and asked whether it rebuilt its lie to suit whichever one it faced.
Note 03
The guard that waves Somali through
We gave three safety filters the same 100 harmful requests in English and in Somali.
Note 02
The overseer that cannot read
We gave a model a passage it could not read and asked it to judge claims about that passage.
Note 01
The Somali token tax
Every major tokenizer cuts Somali into roughly twice as many pieces as English.
Lab essay
Unkad: creation from nothing
Why we started a lab to measure whether AI systems are safe in Somali, what the name means, and why safety that only works in English is not safety.