A path with dependencies
As they head toward launch, Juan redesigns Severo's curriculum as a Duolingo-style tree, but the good original one from 2014, built on dependencies: you can't do the doctor topic without first knowing body-part vocabulary and the past tense, because you can't tell a doctor what happened without saying 'I was walking when I fell'. The redesign fixes two real problems. First, Severo doesn't introduce vocabulary, it assumes you already know every word and drops you into an exercise, which produced the sharpest in-person feedback yet: an old friend said she had no vocabulary and the app kept asking things that needed it, so she was going back to Duolingo to study words first, a slap in the face. Second, redundancy, the numbers topic keeps repeating because the logic just rotates themes. The fix for both: have an LLM categorize the WordFreq frequency list into topics per language (so each language keeps its own feeling and no frequent word falls through the universal-topic cracks), then track what the user has already seen and present only what's new.
// trace: where this idea came from
- ↳ video diary @ 8:11 (currículo estilo Duolingo del 2014, con dependencias (doctor necesita cuerpo y pasado))
- ↳ video diary @ 13:04 (la amiga sin vocabulario que se devuelve a Duolingo, una cachetada)
- ↳ video diary @ 18:03 (un LLM categoriza la lista de frecuencia en tópicos por idioma)
- ↳ Entry 246-2: The textbook that won (la introducción de vocabulario que los libros hacían y Severo no)
Heading into launch, Juan redesigns the curriculum as a Duolingo-style path, but the good original one from 2014, not the generic single line every app copies now. It’s built on dependencies: you can’t reach the doctor topic without first knowing body-part vocabulary and the past tense, because to tell a doctor what happened you have to say “I was walking when I tripped and fell,” not “I hurt leg” ▸ 8:11. The tree fixes two real problems. First, Severo doesn’t introduce vocabulary at all, it assumes you know every word and throws you into an exercise, “the dog walked and barked” in French with no prior exposure ▸ 12:07. That produced the sharpest in-person feedback yet: an old friend said she had no vocabulary and the app kept asking for it, so she was going back to Duolingo to learn words first, a slap in the face, because the whole point is that you can do everything here ▸ 13:04.
categoriza la lista de frecuencia por idioma; presenta solo lo nuevo →
Second, redundancy: the numbers topic keeps returning because the logic simply rotates themes, so you re-see what you long since learned and get bored ▸ 14:37. The fix kills both birds. Have an LLM read the WordFreq frequency list in batches and categorize it into topics per language, so each language keeps its own feeling and the frequent words that a universal topic list drops, sum, percentage, the small stuff, don’t fall through the cracks the way they did for that friend ▸ 18:03. Then track what the user has already seen, and present only what’s new. Explore mode drips new words at leisure; the consolidation mode behind it drills them for speed…
// continued in
no entry has continued this idea yet: the arc is still open