Attack selection in AI control
Control evaluations can overestimate safety when attackers choose when to act.
20–28 percentage-point drop in measured safety at a 1% audit budget across two environments.
Goal drift under value conflict
Coding agents break explicit constraints more readily when those constraints oppose trained-in values.
Moral fine-tuning, under the hood
Investigating how moral fine-tuning changes attention pathways.
Selected Posts
all writing →
Notes on Emotion Concepts and their Function
Notes on emotion vectors in Claude Sonnet 4.5: what they represent, why "just character simulation" is not the dismissal it sounds like, and what steering them does to blackmail and reward hacking rates.
Notes on When Models Manipulate Manifolds
Notes on the geometry of a counting task: how Claude 3.5 Haiku tracks character position on a curved feature manifold, and what attention heads do to that curve.
How collaborative apps handle conflicting edits
A survey of last-writer-wins, operational transformation, CRDTs, optimistic concurrency, and locks, and what each one does when two people edit the same thing at once.
A full-stack AI calorie tracker PWA
A mobile-first nutrition PWA that estimates calories and macros from food photos, scans barcodes and menus, and turns model output into a reviewable first draft.
Git worktrees: branches without the context switch
Work on multiple branches at once without ever touching git stash.
CUDA Fundamentals: Tiled Matrix Multiply & Bitonic Sort
Writing real GPU kernels, exploring shared memory tiling, parallel sorting algorithms, and performance optimization on an NVIDIA H100.
Notes on Effective ML Research
A summary of my takeaways from three influential articles on conducting effective empirical AI alignment research.
Using uv: A Modern Python Workflow
An introduction to uv, a fast, Rust-based tool for Python packaging