Rethink the representation.
How can we represent what a model knows in a more useful form? We investigate the structure inside model weights and explore alternative representations.
Representations / Model structureIntelligence.
Built from
the foundations.
We build AI infrastructure and explore new ways to represent and run models. From the systems underneath to the intelligence ahead.
Our field of work
AI infrastructureExperimental researchFuture models
01 / What we’re exploring
What if intelligence
could work
with less?
Less memory. Less unnecessary computation. More room to explore what’s possible.
Explore our researchReconsider the computation.
What needs to happen for a model to produce an answer? We build and measure alternative execution paths, with memory and hardware in view.
Inference / SystemsBuild toward what’s next.
Infrastructure is our starting point. What we learn will inform our future work on model architectures and learning methods.
Models / Future direction02 / How we work
Curiosity starts it.
Evidence moves it forward.
We follow an idea into code, put it against a baseline, and look closely at what happens. The useful findings become tools. The open questions become the next experiment.
More about our approach01Question
02Build
03Measure
04Learn