Notes on language models, agents, and the ways they break.
I write about the inside of LLM systems like decoding, evals, memory, and multi-agent coordination, usually with a benchmark or a worked example to anchor each post. Long-form, occasionally interactive, mostly things I wish someone had written before I had to figure them out.
Build your World Cup bracket
The tournament's on. Rank the 12 groups, choose which third-placed teams go through, and pick the winner of every match to crown a champion.
Open the predictor
Worth reading first
-
LLMs predict the 2026 World Cup
I gave four frontier LLMs the same context pack and asked each one for a full 48-team forecast of the 2026 World Cup. All four put Spain on top, picked the same top six and nearly every group winner, and only really split on Group D. I'm keeping their picks here and will check them against the real results as the tournament goes.
-
Why Are the Uyuni Salt Flats Hexagonal?
The salt at Uyuni is split into polygons about a metre wide. I assumed they were drying cracks. It turns out the spacing is set by salty water creeping through the wet sand under the crust.
-
How to build an eval suite
In this article I'm going to explain how to go about building an eval suite. In particular, I'm going to walk through the different layers you need to think about when constructing a reliable, well-covered set of evals.
-
Why Streaming LLMs Need Attention Sinks
If you evict the first few tokens from the KV cache, a streaming LLM can fall apart. Here is why that happens, and why holding on to four early tokens is enough to make sliding-window inference work.