Sublunar
Foundation
The research group of Sublunar. We study how to make small AI models that run on the devices people already own, and we publish what we find, including what fails.
Our thesis
Intelligence comes home. The most personal AI will not live in a data center far away. It will live on the devices you already own.
Research areas
- On-Device AI
- Small Model Training
- Evaluation
- Honesty
- AI Security
- Model Behaviour
- Offline Knowledge
- Private Search
- Personal Memory
Why it matters.
AI will soon know more about you than any app ever has. The open question is where that knowledge lives. Our research is about keeping it with you.
How we work.
- We publish the failures.Every report has a pass and fail table. What went wrong is part of the result.
- We test on real machines.Numbers come from the Macs people buy, not from a data center.
- We write the rules down.One document holds how Glim should behave, in order of priority.
- We keep it readable.If a finding can’t be explained in plain words, we’re not done with it.
Published.
One technical report and a running research log. Newest first.
- Glim: an on-device assistant with offline knowledge, and what our first tests caughtReport · Zenodo
- What Glim reads is information, never ordersSafety
- What our first tests caughtEvaluation
- Small models, perfectedModels
- How Glim should behave, written downBehaviour
- Knowledge for when the internet is goneOffline