Best Work

This is the shortest path into the site.

Mechanistic Interpretability

• Your SAE looks solved. Your model disagrees
  An argument, with measurements, that reconstruction fidelity can look good while behavioral preservation gets worse.

• Transport Geometry as a Testable Hypothesis?
  The follow-up post: what survived, what disappeared with more training, and what geometry-aware proxy I think is worth testing next.

Representation Geometry / Computational Olfaction

• Robust but Not Unique: Topological Signal in Learned Odor Embeddings
  A technical post on the odor-topology project: robust H1 in OpenPOM, strong chemical baselines, and why the compression mismatch is part of the real story.

• odor-topology paper (PDF)
  The current compact paper artifact.

Speech Modeling

• Completed 12k Run: Early Speech Emergence in a 21.9M OLMo-Hybrid Speech LM
  The main speech result, with validation history, audio samples, and a clear non-claim section.

• OLMo-hybrid speech technical report (PDF)
  A compact artifact for the speech line.

• olmo-hybrid-speech code
  Public code for the speech experiments.

If you want the broader context, the Archive has the full site.