Writing
The long version.
Notes on post-training, interpretability, sequence modelling, and the occasional brain. Code-first, and honest about results, including the ones that did not work. Every post here is a chapter of the living CV, told slowly.
Readable shallow, steerable deep: the same split in three transformers
A direction I can read a composed sentiment concept off of is nearly perpendicular to the direction that steers it, and it is a worse lever at matched dose. The same split holds on three transformers and two normalization families, and a logit-lens diagnostic offers a candidate account of why the useful site is deep.
Does a single direction mediate refusal?
A small, controlled reproduction with norm-matched random baselines.
A probe at layer 0 is a lie detector for your experiment
The number is not the finding; the shape of the curve is.
Why SFT learned the words but GRPO learned the rules
Teaching a 14B model a proprietary building taxonomy with a hand-tuned reward.
From consuming a pretrained model to training my own
A sign-language Copilot trained from scratch, with a hybrid runtime.
Patient-specific filters as biomarkers
ICA, FOOOF and CSP for noisy EEG, and what they taught me about transformers.