blog
Thoughts & notes
thoughts on AI safety, building things, and whatever I'm thinking about.
On human-agent collaboration: métis, Centaur Chess, and the milk problemJun 3, 2026Notes on the OpenAI Hackathon: Conway's Law and building with strangersMay 30, 2026Notes from Capital Factory: the forest, the grid, and what demoing teaches youMay 4, 2026On health literacy and design: raw eggs, Foucault's clinic, and the Blue CardMay 1, 2026On causal abstraction: Aunt Hillary, Borges' map, and implementing an algorithmApr 14, 2026On three kinds of alignment work: the polygraph, the Stag Hunt, and the hidden moleApr 11, 2026On scalable oversight: Meno's paradox and weak-to-strong generalizationApr 4, 2026On goal misgeneralization: Kripkenstein's quus and the CoinRun illusionMar 28, 2026On reward hacking: the Cobra Effect, CoastRunners, and the math of nanoRLHFMar 22, 2026Notes from SXSW 2026: Ender at the terminal and the runaway broomMar 19, 2026On attribution graphs: slime molds, Heptapod B, and the basal gangliaFeb 28, 2026On alignment faking: Ketman, the Panopticon, and why Claude is faking itFeb 21, 2026On emergent misalignment: Ice-Nine, holographic personas, and the mathFeb 14, 2026On AI control: The Garden of Forking Paths, time cops, and trapping a scheming AIFeb 7, 2026On DeepMind's AGI safety paper: the Golem, the Maginot Line, and defense-in-depthJan 31, 2026Hello world: on AI's adolescence, bootleggers, and the alien in the datacenterJan 28, 2026