vishal-dehurdle/hypersae
High-Performance Hyperbolic Sparse Autoencoders for Mechanistic Interpretability
10 best points · 0 comments · 5 Hacker News posts · 13 GitHub stars · View on GitHub
Hacker News discussions
- Show HN: At 16K features, flat autoencoders break. Curved space doesn't (10 points, 0 comments, by visha1v, 2026-08-13)
- Show HN: HyperSAE – Sparse Autoencoders, reimagined in hyperbolic space (6 points, 0 comments, by visha1v, 2026-08-19)
- Show HN: HyperSAE – Hyperbolic Sparse Autoencoders for LLM Interpretability (3 points, 0 comments, by visha1v, 2026-08-11)
- Show HN: LLMs think in hierarchies. Their interpretability tools should too (3 points, 0 comments, by visha1v, 2026-08-15)
- Show HN: See what an LLM learned, organised as a browsable tree (1 points, 0 comments, by visha1v, 2026-08-18)