Behavioral alignment is the next frontier of AI
Models learned the world by reading about it
Nearly everything they know came from text, and this took them remarkably far.
Agents learn judgment on the job
They have enough knowledge to do the work. What they lack is judgment, which humans learn on the job and rarely write down.
We build worlds for agents to practice
Each one is modeled on a real working environment, where information is hidden, user requests conflict, and not every requirement is stated.
We study what agents do
When an agent misses the point or cuts corners, we figure out why. Then we build the training data to correct it.
Every domain asks for different judgment
Research
Most research doesn’t lead to a breakthrough.
Experiments stall and results come back empty.
Each one rules something out.
Then one breaks through.
Research comes first
We study what’s just out of reach for models.
Come work at the limits of model behavior
View open roles
