Reproducing a Best-of-N jailbreak setup on smaller models, scaling experiments onto Colab, and thinking through what the results imply for robustness.
Writing
Notes, experiments, and systems write-ups
Mostly the things I wanted to understand well enough that I had to write them down.
A short note on Shepherd, Stripe's feature platform, and how it builds on the open-source Chronon model for point-in-time correct batch and serving workflows.