A world made of vocabulary

The diagnosis comes from watching real users, Julia and the partner: Severo’s skeleton works, but between exercises there’s a dead moment where you must think up what vocabulary you want, and those seconds of blankness kill the momentum, “no se siente interactivo” ▸ 36:26. Julia’s fix is curricular: a mixed system, an official track for the A1-to-B2 spine plus free-choice topics, because Duolingo’s rigidity is the anti-model, “en Duolingo no le van a decir cómo se dice machine learning” ▸ 38:35.

His fix is a fantasy he labels honestly, “una idea mala, pero la quería probar” ▸ 39:34: every learned word adds its object to a growing collage-world, bread, sandals, a table accreting into your personal city ▸ 39:56. The motivation inverts from passive to active: “a mi ciudadcita le falta un avión”, so you go learn plane ▸ 45:47, and the world doubles as a portrait, a skyline of machine-learning terms says who you are ▸ 40:46.

documenta por qué la idea linda no funciona; ese es el prototipo →

Then the engineering autopsy, which is the entry’s real value: no current model lets you place objects in an existing image, or knows that learning “building” means zooming the scene out ▸ 42:09; green-screen cutouts the user arranges themselves turn into tedious z-order homework ▸ 42:47; puzzle-island tiles come back at incompatible scales ▸ 43:29; icons lose the charm ▸ 43:47; and the economics close the door, about a cent per generated image, ten dollars per thousand words per user ▸ 44:20, unless everyone shares one image bank and the personalization dies with the cost ▸ 44:32. Verdict filed without drama: not the logic for now ▸ 45:18, the skeleton is done and what’s missing is the creative spark ▸ 46:29…