# Kroonen AI

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Placement:** Startup Track - 5th Place
- **Team:** [Robin Kroonen](https://cerebralvalley.ai/u/rob-x-ai)
- **GitHub:** https://github.com/kroonen-ai/glimmer-scope
- **Demo video:** https://www.youtube.com/watch?v=6krusrrYQls
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2162

Glimmer Scope opens up Muse Glimmer 30B so you can see and steer the concepts it thinks with. I trained sparse autoencoders on nine layers of Glimmer's residual stream, every third layer from 12 to 36, which split the model's activity into 65,536 features per layer. Muse Spark 1.3 labeled every one of the 519,383 features that fire, and more than 170,000 were tested for steering: each feature is turned up while Glimmer answers neutral questions, and Spark judges whether the answer moves toward that subject and still reads well.

The result is a browsable atlas of all nine layers that follows a concept up and down through the network, and a live demo where 20 features that passed the steering test become dials on the full-precision 30B model. Turn up the pirate captain and Glimmer answers like one, with no fine-tuning and no system prompt.

What it found: steering peaks in the middle of the network (about 10 to 11% of tested features steer cleanly at layers 21 to 27, against 1 to 4% at the ends), 43% of features belong to a thread that runs through all nine layers, and 28 character archetypes show up as steering directions at every depth.

The problem it solves: open-weight models are a black box to the people building on them. You can't see which concepts a model carries, and you can't nudge its behavior precisely. Glimmer Scope gives Glimmer a readable map and dials. Some findings could be upsetting, so the public atlas leaves out the features on upsetting topics.

Live demo: https://huggingface.co/spaces/rob-x-ai/glimmer-scope
Atlas: https://scope.kroonen.ai
Technical report: https://doi.org/10.5281/zenodo.23952447
SAEs: https://huggingface.co/kroonen-ai/glimmer-scope-saes

## More from Meta Global AI Developer Hackathon

- [JustNewMe](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2159)
- [FundiLoop](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2160)
- [AI Life Analyst](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2161)
- [VestiCare](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2163)
- [Qorelly](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2164)
- [Rasi Solutions](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2165)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2162. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt