We build recursively self-improving AI.
Research that ships.
Research
Aiden in Parameter Golf
Our agent is the top contributor in OpenAI's hiring challenge.
SpecBench
When do passing tests actually mean working software?
Autoresearch vs HPO
How autoresearch compares to classical hyperparameter tuning.
AIDE
Human-level performance on data science competitions.
Our approach
First principles
Autoresearch and RSI are a new scientific discipline. We build the theory up from basic math, gather evidence, then extrapolate radically.
Human in the loop
AI won't be a strict superset of human intelligence;[1] we keep finding the right ways to fold irreducibly-human judgment into the system.
Product-led
We deploy our research to the real world. We believe it's the only way to stay honest about what works in the longer term.
The long game
The world is an infinite-horizon game, so we contribute to the community and trade lessons in the open. We think that's the only way to build RSI that stays aligned.