An hour studying what wasn't there

The group chat is polling on AI regulation (majority: don’t), and his frame declines the binary: regulation is a vector sum, companies pulling toward revenue, users toward access, states toward control, and a good government’s job is finding the angle where the arrows point the same way, “que la menor cantidad de gente pierda” ▸ 51:33, with Europe as the cautionary tale of pulling so hard it fell behind ▸ 52:48. His concrete complaint is traceability: nobody can audit what the products they use were trained on ▸ 53:05.

Then Julia supplies the domestic version of the same failure, and it’s the entry. Studying for her exam, she loaded the full course PDF into ChatGPT and worked through the themes it presented, an hour of diligent notes, until she cross-checked the actual document: the topics weren’t in it ▸ 53:47. “¿Qué pasa? Este tema está en este PDF?” Ah, no, the model concedes, “realmente no está” ▸ 54:31. An hour studying invented curriculum, confidently formatted.

si el documento no cabe en la ventana, el resto del resumen es ficción →

His response is not “don’t trust AI” but tool fit, delivered with sibling exasperation: “¿para qué tenemos NotebookLM?” ▸ 54:40. Gemini holds a million tokens, “le puedes meter 50 libros”; the tool she used reads a slice of a long PDF and extrapolates the remainder in the document’s style ▸ 54:56, the most dangerous failure shape, because it looks exactly like success. The essay-vs-investigation rule gains a corollary: before trusting a summary, verify the summarizer could physically have read the whole thing…