Hi, I'm Yannick
I'm a machine learning engineer based in Zurich. I like building systems that make AI more reliable, understandable, and safe. 🧠⚙️
Featured
Interpretability · 10 min read
How a one-head attention-only transformer computes the maximum
Reverse-engineering a one-layer, one-head, attention-only transformer that computes the maximum of a sequence, read straight off the weights: the QK circuit attends to the biggest token, the OV circuit copies it, and both read a single size direction in embedding space. Plus the min variant, which is the same circuit flipped.