Sparse autoencoders on nine layers of Muse Glimmer 30B. Report: doi.org/10.5281/zenodo.23952447 · Atlas: scope.kroonen.ai
AI & ML interests
Interpretability, sparse autoencoders, activation steering, pretraining, evaluation, open models
Recent Activity
Organization Card
Kroonen AI
Kroonen AI is an independent research lab and product studio. We study language models end to end, from pretraining on our own hardware to interpretability on open frontier-scale models, and we release the weights, code and write-ups, including what didn't work.
Research
Glimmer Scope (November 2026). Sparse autoencoders on nine layers of Muse Glimmer 30B: 519,383 labeled features, more than 170,000 steering tests, and every feature followed from layer to layer. SAEs · Live demo · Atlas · Report
Genesis 1B. A 1B-parameter foundation model pretrained from scratch on two RTX 4090s: 80,000 steps, about 42B tokens. Playground
Focus
- Interpretability and steering of open models
- Pretraining and fine-tuning on modest hardware
- Evaluation and model safety
- Local-first and hybrid deployment
