Writing
Notes on software craft, autonomous systems, and whatever language I am currently learning the hard way.
latest
BlueDot Technical AI Safety Puzzle #1 - Submission
A walkthrough of how a frozen MiniLM head represents the country feature nonlinearly at layer L: a dartboard of radius and angle, the ReLU that deletes the linear copy, and a weirder locker-code representation trained on top.
2026-06-15 / interpretability / 2 min
Archive2 entries
- 2026Jun 15 interpretability
BlueDot Technical AI Safety Puzzle #1 - Submission
A walkthrough of how a frozen MiniLM head represents the country feature nonlinearly at layer L: a dartboard of radius and angle, the ReLU that deletes the linear copy, and a weirder locker-code representation trained on top.
2 min - 2026Jun 15 notes
Hello, world!
A first field note, a small signal from a new corner of the web, and enough lorem ipsum to test the typesetting.
2 min