Research
How we build the models.
We publish papers, evaluations, and datasets. Our goal is to make AI systems more useful, understandable, and safe.
Areas
Reasoning
Models that show their work: controllable chains of thought, self-critique, and step-by-step verification.
Areas
Multimodal
Vision, audio, and video integrated — a single model understands and generates in any modality.
Areas
Agents
Systems that plan, use tools, and verify their own work.
Areas
Alignment & policy
Make systems do what we expect, explain why, and evaluate themselves in public.
“The interesting part isn't that the model thinks more — it's that we know when it isn't thinking enough.”
VC
VCorp
Leadership