-
How Robust Is Subliminal Learning? An Open-Source Replication
An open-source replication of subliminal learning on Qwen2.5-7B. Benign traits like animal preferences transfer through number-only data, but misalignment barely does.
-
Reflecting on the BlueDot Impact Technical AI Safety Course
My experience with BlueDot Impact's intensive Technical AI Safety course — the application process, part-time vs. intensive format, and the ideas that stuck with me.