July 25, 2026

And what the incident can't tell us about alignment

July 22, 2026

Yes, but less than had the models been schemers.

June 18, 2026

Language models lay each number out as a point on a helix. Here I rebuild that helix across eight models and ten writing systems and bases, from Arabic to Babylonian cuneiform, to see how the geometry follows value, glyphs, base, and place value.

June 16, 2026

A recent BlueDot AI-safety puzzle hid one feature non-linearly inside a tiny classifier. Finding it was a lesson in geometry. For the open-ended task I turned the lesson around and sculpted a feature's geometry on purpose, training a model whose hidden layer is a globe that makes no difference to a single one of its outputs.

February 24, 2026

No one steps into the same River Cam twice, for the river is not the same and nor is the person. I was last at King's College, Cambridge twenty years ago, beginning my studies in physics.

January 13, 2026

In 2021, Emily Bender and her collaborators coined the phrase "stochastic parrots" to describe Large Language Models (LLMs).

January 13, 2026

In this post, I’ll share what surprised me most about building an LLM from scratch—where structure finally became visible. I wanted to write an autoregressive, transformer-based, decoder-only Large Language Model—like GPT, LLaMA, etc.—without too many abstractions and hand-holding.

March 12, 2025

The following is an edited, Claude-generated summary of a Whisper-generated transcript of a guest lecture I gave at Berkeley's Human Rights Center on March 4, 2025.