Public benefit corporation · Est. 2021
Frontier AI, built with safety as a design constraint rather than a research footnote.
We run one of the few labs doing both halves of this work at once: training models at the frontier, and studying the discipline that keeps them worth deploying. Neither half is much use without the other, which is uncomfortable and, we think, correct.
Cortex 5.1 and Pulse 5.1
Two models for work that runs longer than a single turn. Both hold state across tools, recover from their own mistakes, and report uncertainty instead of inventing a finish line.
Read the announcementLatest releases
All announcementsCortex 5.1 and Pulse 5.1
A paired release aimed at long-horizon agentic work, with a rewritten evaluation suite covering tool misuse and instruction drift over multi-day tasks.
Interpretability: a working map of one attention head
We traced a single circuit that decides when a model withholds an answer, then showed we could steer it — and that the steering generalised to held-out tasks.
Responsible scaling, version 3
Capability thresholds are now paired with required safeguards, plus a published process for what happens when a threshold is crossed mid-training.
AI should widen what people are able to do — without quietly deciding who gets to do it.
Where to go next
Six entry points, chosen because each one is a claim we would have to defend in public.
Putting this in front of your organisation.
Enterprise plans include audit logging, SSO, region pinning, and a named security contact. If you need to know exactly where your data sits, we would rather answer that in writing before you ask twice.