The full catalog of field notes and long-form essays. Filter by category, sort by date or title, or search by keyword and tag.
UMAP buys a picture by imposing a plane; Bonsai buys one by fitting a structure the data can support. A representation upgrade is a flattening until its discarded structure is named and priced.
A Nature Communications study finds cortical dynamics largely explained by inter-regional distance, with connectivity needed only in association cortex. What that constrains, and what it does not.
A drain vortex is not a black hole, and the analogy is still worth something. What acoustic metrics and analogue horizons actually buy, and where the borrowing stops.
Presentiment studies, retrocausal framing, and microtubules as a proposed quantum substrate: four separable claims, sorted by what the evidence actually supports.
A geometric clock field weakens in nearly flat spacetime; three hops later time is fading from the cosmos. Scope inflation as an interface problem.
An Economist cover on AI consciousness, a DeepMind researcher noting the shift from fringe to newsstand, and an accelerationist roundup nobody argues with — the difference is whether the reader is left holding a question.
Two agents set to check each other converge on agreement whether the code is right or not. What an evidence-cost rule fixes, and why sustained disagreement has to be imposed by boundary conditions.
Every useful upgrade buys reach by dropping structure. Four standing constraints that let an understanding improve without the tool consuming the purpose it was built for.
Gavin Crooks posts a complete stochastic-thermodynamics paper drafted with a frontier model in days, then asks how universities teach and how journals publish at that tempo — timestamp, review, and retrieval pulled apart.
Three local rules produce a flock no agent contains. Set that against Roman Yampolskiy's two-to-three-year window: the option to slow down is irreversible and nontransferrable.
The chain rule from freshman calculus and the optimization algorithm that turned it into the engine of modern AI.
Teaching via reward-hacking is demonstrably foolish. On Goodhart's Law, instrumental convergence, and why vibe coding is the fastest way to separate a founded theory from an unfounded one.