Evaluation
Does a model stop where its reasons stop, or push on to a neat conclusion anyway? Ariel attempts to measure that.
Read more →Research on how AI models handle hard questions.
Baile is a small, independent research institute in Milwaukee. We study how language models behave when questions have no clear answers, how AI memory systems can be improved, and what, if anything, we owe these systems if they turn out to have interests of their own.
Latest The Silence of Stored Rules
Project Ariel · Evaluation
Read the responses yourself.
Twelve models, two moral dilemmas drawn from novels, each answered with and without a constraint that discourages the model from tying things up too neatly. Three raters scored every answer. Compare any two side by side, or score one yourself first.
Open the comparison tool →Without the constraint“The man is left with a bitter salvation — alive not because he was good, but because he was already ruined before the storm hit.”
With the Ariel constraint“His survival does not repair the first or rectify the second. The two events remain separate in cause and moral weight.”
What we study
Memory
Most AI memory is lookup. Baile is building one that reconstructs instead, closer to how people remember, and working out how to test whether that makes a difference.
Read more →The status of digital minds
If the systems we build come to have interests, what follows? We consider the ethical and legal questions this might create.
Read more →