Curtis Covington
Metrics first, caveats included, and enough context to understand the result instead of just the headline.
I trained a tiny decoder-only transformer from scratch on a finite-state-machine task, then probed its activations. The strongest evidence was a fragile mid-training next-state representation, not a robust learned state machine.
Read featured postLatest field note from the lab archive.
I ran a hardened Gemma SAE steering evaluation with holdout prompts, matched controls, wrong-hook checks, and seed extensions. Code held up; most other categories did not.
Read postA wiki-only SAE intervention lab on Pythia-70M: feature ranking, residualized feature knobs, and activation-gated minimal-pair behavior tests.
Read postLayer sweeps, top-k sparsity sweeps, and an interactive explorer for SAE feature discovery on a small open-weight model.
Read postA local-only SAE feature probing experiment on Pythia-70M: activation extraction, sparse autoencoder training, cross-domain generalization, and blog-ready artifacts.
Read postPairing with GPT-5-Codex to turn the SSA baby-name dataset into a polished Go CLI with weighted sampling, charts, and automated releases.
Read postAbout
This blog is where I write up hands-on experiments in interpretability, sparse autoencoders, feature steering, and related tooling with the methodology, results, and caveats left in.