Liad Bercovitch

Weizmann Institute of Science

Software engineer.

Interested in evaluation, efficient training and inference, and interpretability — how to tell if a model actually works, what it costs to run, and what happens inside it.

Posts 1

🧠 Learning to learn Oct 23, 2026

Why autoregressive and diffusion LMs can imitate each other (and why it still costs)

A claim from week 1, made easier to follow with a picture: the two model types can reach the same results, each at a cost.

Cross-lingual knowledge in pretraining