Releases, papers, ideas in progress. We write in long form when there is something worth saying.
Llana 3.2 is a refresh focused on refusal calibration, long-form coherence, and code review quality. We are slightly slower than 3.1 in raw tokens-per-second; we think the trade is right. Notes inside.
Read the post →An accompanying essay to our refusal-calibration paper. We argue that knowing when not to answer is, on balance, a more useful skill than answering one extra question correctly.
Read the post →Voice agents are easy to demo and hard to make safe. We have a working prototype; we are not going to release it until we can describe its failure modes in plain language.
Read the post →Six new positions across reasoning, evaluations, and trust & safety. We hire one person at a time and we read every application.
Read the post →