The pipeline that forgets your name

The paper is Anthropic’s Clio, the system built “para analizar anónimamente todos los las conversaciones” ▸ 30:55, and his walkthrough keeps the architecture’s logic intact. You can’t put a human on the chats, that’s a privacy violation by definition, so the anonymization is done by models distilling models. Haiku, the light one, summarizes each conversation ▸ 31:35 under orders to omit names, locations, and unique identifiers ▸ 32:10; his counterexample of why that instruction matters is domestic: “A Julia le gusta hablar sobre anime” is already too personal ▸ 32:19. The summaries get vectorized and clustered so related topics drift together ▸ 32:38, Sonnet audits what passes so only the generalized survives ▸ 33:47, then builds the taxonomy, programming, then data science, then machine learning ▸ 34:12, with a separate map for the unknowns that fit no category ▸ 35:03. Reported fidelity: about 96% ▸ 35:44.

las keywords cuentan clics; la intención necesita significado →

Why a fintech vlog spends ten minutes on this: “es precisamente un concepto que podemos utilizar para nuestro propio proyecto” ▸ 36:00. Traditional product analytics counts clicks; a text interface has nothing to count ▸ 36:34. And Reisi is intent based by design: people will say what they want to do with their money ▸ 36:58. Keywords are the tempting shortcut and his rebuttal is compact: “odio la programación” makes a keyword system recommend programming courses ▸ 37:49, and if keywords were enough, nobody would have built Clio ▸ 37:23. The privacy half is just as load-bearing for a bank: what Pepito Pérez wanted to pay for is nobody’s business; the system should only ever learn which payment rail he was missing ▸ 38:14…