Training systems
Efficient pretraining
How can an independent lab extract more learning from every token and unit of compute?
R / 01
Research
Research builtto be examined.
Our teams study training efficiency, reliable reasoning, and model safety so future open systems arrive with evidence.
Research
Three focused areas.
Alignment science
Reliable reasoning
When does a model reason faithfully, recognize uncertainty, and correct its own mistakes?
Model safety
Release evidence
Which capability and risk evaluations are necessary before an open release can be justified?
Our standard
Show the work.
Methods / Failures / Limits / Evidence
01State the question
02Publish the method
03Report the failure
04Change the decision