Build for the world.
Start with an actual signal, an actual interaction, and a question worth answering. Intelligence should earn its place in the experience.
INSIDE INTERACTIVE MACHINES / 04
A future drawn by hand.
Built by science.
We are an independent AI research lab exploring new forms of intelligence. Our work brings sensory models, reflex models, and continuous world modeling into one shared ambition.
THE QUESTION THAT MOVES US
What else could
intelligence become?
We believe the next chapter of AI will be shaped by models with new capabilities, new structures, and new relationships to the world.
Our focus is the foundation: perception that keeps pace with events, judgments that fit the task, and state that carries understanding forward.
We’re building toward machines that feel more present, useful, and responsive in everyday life. That is an ambitious direction. The work begins with small, testable steps.
HOW WE WORK
Draft. Build.
Observe. Begin again.
A notebook is a place for possibilities. A laboratory is where we discover which ones hold up.
- 01
Draft
Draw the problem. Define the signal, the decision, and the failure that would matter.
- 02
Build
Make the smallest useful system. Put the idea in contact with real inputs.
- 03
Observe
Watch the interaction unfold. Measure what the system notices, misses, and gets wrong.
- 04
Revise
Follow the evidence back to the design. Improve the model and question the premise.
THE PRINCIPLES BEHIND THE WORK
Wonder, with
its feet on the ground.
We can be excited about the future and precise about the present. These are the commitments that connect the two.
Give each model a purpose.
Define its inputs, its outputs, and the conditions where it should ask for help. Use the right kind of model for the work.
Make uncertainty visible.
Keep measurements separate from interpretations. A useful model should communicate what it does not know.
Treat continuity as a responsibility.
State needs boundaries: what is retained, for how long, under whose control, and how it can be forgotten.
Test the whole interaction.
A promising result in isolation is a beginning. Evaluate timing, reliability, and failure across the system people will actually use.
Let evidence change the drawing.
A beautiful idea must survive contact with reality. Define what would disprove a claim, then go looking for it.
FROM THE READING TABLE
Ideas we’re
thinking with.
Selected research that informs the questions we ask. These papers describe their authors’ work; they do not establish the performance of our models.
01
WORLD MODELS
World Models
Learning compact representations of an environment to support a policy. A starting point for thinking about internal models of a world.
Open paper in a new tab02SELECTIVE COMPUTATION
RouteLLM
Research on learning when to route a request between stronger and weaker language models. A useful reference for allocating computation.
Open paper in a new tab03SEQUENCE & STATE
Mamba
Selective state spaces as an approach to sequence modeling. Part of the wider research landscape for processing long streams.
Open paper in a new tab04CONTINUOUS PERCEPTION
Stateful Conformer
Cache-based inference for streaming speech recognition. Relevant background for systems that process audio as it arrives.
Open paper in a new tab