Everything, newest first.

What I read after Sora: an annotated list

DiT scaling, motion priors, flow matching — the papers that actually changed something and the ones that didn't.

Every AMM makes a bet. Here's what each one is betting on.

A walkthrough of invariant families from constant-product to concentrated liquidity — and which assumption breaks first under real order flow.

Taste is leverage. Most people don't use it.

The kind of judgement that only reveals itself in the tenth decision — and why it's worth cultivating even in technical work.

The three things a video tokenizer cannot afford to lose

A checklist I return to when evaluating video autoencoders. Short, opinionated.

Three months watching a MEV searcher: what I found

Sandwich decay, bundle inclusion patterns, and the slow death of naive frontrunning.

Diffusion for control: impressive method, uncertain edge

A paper using diffusion models for robotic control. Worth reading once; my brief take on why the bar is higher than it looks.

Eight things I was wrong about in 2025

A year-end accounting of the priors that didn't survive contact with reality.