All Posts

Before the computer was built, there was an older dream: to write down reason itself. Three collisions between that dream and the modern machine — from Saussure and word2vec, to Wittgenstein and hallucination, to Hume and generalization.

Start with five restaurants whose true quality is unknown, then follow the same problem through multi-armed bandits, dueling bandits, noisy pairwise ranking, and top-K selection: the optimization structure is known, but its parameters must be learned by trying.

Every team has more good ideas than capacity. The hard part is selection — learning to see past the costly signals, prestigious credentials, and compelling narratives to focus on what improves the outcome.

How can agents improve educational demos when quality depends on human judgment? A proposed workflow using explicit teaching patterns, multiple evaluators, and checks on what those evaluations actually measure.

Devesh ran a shady octopus meat caravan in the Simulation. Top agent, deep cover. Eight tentacles, eight side hustles. A story about love, AI, and taxes.