Sous Lab — a research lab
We build visionary machines, not merely predictive ones.
A research lab building continually-learning systems whose taste is grounded in real discovery — not human preference.
Predictive machines learn the past.Visionary machines learn where to look next.
Next-token prediction is powerful, but it is trained to extend what already looks likely. Left alone, it drifts toward hyper-normal science: making the current knowledge domain more coherent, without learning how to break its boundary.
Sous aims to build visionary machines. We focus on two far-reaching but promising visions: autotelic AI agents that continually learn, and meta-objectives that guide their learning.
Progress, measured.
Experiments, systems, and measured progress from Sous.
More robotics coming soon...
Learning the shape of a push
In 2.5 hours, an executable dynamics model learned from 48 episodes and solved 97 of 100 held-out PushT starts.
Can you win a game without knowing the rules?
ARC-AGI-3 gives the agent a world but no manual. We ask whether a model written, tested, and revised during play can turn uncertainty into a plan.
Build with us.
A small team chasing a large bet — researchers and builders welcome.