longer pieces i've written — mostly things i became fascinated by and wanted to understand end to end. click read to open the full article.
what residual streams are, why they matter for interpretability, and how attention & FFN outputs accumulate into a shared communication channel across layers — worked through with a concrete "capital of France" example.
read.jpeg)
beginner questions i worked through while learning mechanistic interpretability — activation vector spaces, polysemantic neurons, bigrams & skip-trigrams, QK/OV circuits, and the "no privileged basis" of the residual stream.
readlooking back at four years of college — the robotics lab, learning ML alone when nobody else was, and the people who'll stay in my life long after. one last time.
reada short hello and a note on what i plan to write about here — thoughts on the parts of tech i'm currently fascinated by.
reada thread on the platonic representation hypothesis — how models trained on different data and modalities seem to converge toward a shared statistical model of reality.
read on twitter