AActive
Generative AI
Taking large models from an impressive demo to something dependable.
- Agentic systems that recover from their own mistakes
- Evaluation, grounding and guardrails
- ████████████████
Operating in stealth · Est. 2023
Generative AI,
Embodied AI
and Robotics.
From model to machine.
We work at the seam where frontier models meet physical machines. Most of what we are building is not public yet. This page is what we can say for now.
01 — Focus
Three pillars.
One continuous stack.
The interesting problems now sit between models and machines. We work on both sides of that line. Some of the detail stays under wraps.
BIn the lab
Embodied AI
Where foundation models meet sensors, actuators and the messiness of the real world.
- Vision-language-action models
- Learning from experience, in simulation and out of it
- █████████████████████
CIn the lab
Robotics
Putting intelligence into machines that have to work outside the lab.
- Perception and autonomy, edge to cloud
- On-device inference under real constraints
- ██████████████
02 — Principles
From model to machine.
We are not ready to show the work. We are happy to show how we think.
- i
The model is the easy 20%.
The demo dazzles in a week. The real work is everything around it: data, evaluation, recovery, integration.
- ii
Intelligence gets interesting when it touches the world.
Latency, safety and physics do not care how good the benchmark looked.
- iii
Ship the narrow thing.
One painful problem solved completely beats a general assistant that almost works.
03 — Status
Heads down.
- 2023
- Founded in California.
- 2024 — 25
- Quiet years. Research, prototypes, a lot of listening.
- 2026
- Opening a limited number of conversations.
- Next
- ████████████████████████
04 — Signal
More soon.
If you are building where models meet machines and want to compare notes before we are public, send a short note about what you are working on. We reply to the ones where we can be useful.