What?

I work on things I find interesting. more recently spanning interpretability, representation geometry, and sequence and speech modeling. This site is a research log for experiments, reports, technical notes, and a smaller amount of personal writing.

The homepage foregrounds research work. Diary entries live on a separate page. If you want the shortest path into the site, start with Best Work. The Archive includes everything.

Lights

• Your SAE looks solved. Your model disagrees
  A concrete interpretability result about why reconstruction metrics can overstate behavioral faithfulness.
• Transport Geometry as a Testable Hypothesis?
  A follow-up that tightens the claim, reports what disappeared with more training, and lays out the next diagnostic step.
• Topological Signal in Learned Odor Embeddings
  A representation-audit paper on POM/OpenPOM, persistent homology, chemical baselines, and why compression mismatch matters for interpretation.
• odor-topology paper (PDF)
  The compact artifact version of the odor-topology result.
• Completed 12k Run: Early Speech Emergence in a 21.9M OLMo-Hybrid Speech LM
  A speech-modeling report with metrics, samples, and a narrow viability claim.
• OLMo-hybrid speech technical report (PDF)
  The compact artifact version of the speech result.
• olmo-hybrid-speech code
  Code for the current speech modeling line.

Recent Research

• Mar 14, 2026

  Completed 12k Run: Early Speech Emergence in a 21.9M OLMo-Hybrid Speech LM

• Mar 10, 2026

  A 21.9M hybrid recurrent attention to learn speech on 1.8k steps

• Mar 6, 2026

  OLMo Hybrid for Autoregressive Music Token Modeling: A 300M Single-A100 Baseline

• Mar 4, 2026

  Janky memory cache over mamba hybrid for music?

• Feb 22, 2026

  Transport Geometry as a Testable Hypothesis?

• Feb 9, 2026

  Your SAE looks solved. Your model disagrees

• Apr 25, 2023

  Building out a mini-compute cluster

• Feb 11, 2023

  Review on audio generative systems

Essays And Notes

• Apr 8, 2026

  Topological Signal In Learned Odor Embeddings

• Mar 28, 2026

  Hmm1

• Apr 24, 2023

  Consumer llms are coming (here?)

• Dec 25, 2022

  A Dirge for the Romantics

• Oct 18, 2022

  Neural Fields are cool

• Oct 18, 2022

  Latent Diffusion models are Jungian artifacts

Diary entries are collected on the Diary page.