Something about today’s personal agents is off. They don’t belong to the people who use them.
They live in someone else’s cloud, on someone else’s terms, and can be changed, repriced or switched off from somewhere you can’t see. Personal computers were never like that. Once a laptop was yours, no one at Dell or Microsoft decided what you did with it, and no one could take it back.
That gap matters more as agents leave the chat window. Soon they will live in what we wear and keep at home, listening and helping all day. At that point, privacy and ownership stop being features. They become the condition for letting an agent in at all.
What we’re building
Nora builds personal intelligence that runs on your own devices and answers to you.
Our first devices, a ring, a card and a hub for the home, work alongside your phone on a runtime we built for agents that never stop running. We still call frontier models in the cloud when a task needs one, because serving the person comes first. But the core is moving into a family of models, from around 30B to 120B-class, running on hardware the person owns.
Who we’re looking for
A cofounder to own that family of models: how they’re trained and post-trained, and the team that does it.
Because each model runs on hardware its owner controls, it can learn from that one person in ways a shared cloud model can’t. That opens problems a lab serving one model to everyone doesn’t face:
- 01Learning from a person’s real use, without their raw data leaving their own devices
- 02A session that never ends
- 03Rewards from what actually happened: the action ran or it didn’t, the person kept it or undid it
Some of this is engineering we know how to do. Some of it nobody has solved yet.
Models this size mean short loops. You’d run many full post-training cycles a quarter and ship them to real people, rather than a slice of one large run.
We’ve worked close to models long enough to know they aren’t magic, and to know this seat should be yours. A model team needs compute, real data and room to make its own calls. That’s why compute is committed before the hire, not after.
Maybe you’ve done post-training at a frontier lab and want your work in people’s hands soon. Here you’d shape a whole model family, and the company around it, rather than a slice of someone else’s. If so, we’d like to meet you.
Who we are
We work where models, systems and hardware meet.
- Models
- Ex-NLP @ MIT Media Lab, from the first deep-learning wave
- Systems
- Years of distributed and formally verified systems, including a settlement network that grew to hold over $120M of users’ funds and earned ~$15M in open-source grants
- Hardware
- Devices built closely with manufacturers in Shenzhen, with partners across the US and Asia
You’d be the first model person here.
How we work
We treat technical design as a craft.
We’re AI-native, all the way down.
Where things stand
Early and moving quickly.
- Our runtime already runs one agent across device, phone and cloud, and partners are building it into their own devices
- Our first devices open for preorder this holiday season
- Roughly 500–1,000 H-class GPUs of compute lined up for the model team early on
If this sounds like you, or someone you know, write to us.