A translator, not a dictionary
The tap-any-word feature gets its requirements audit and loses its name: a dictionary defines (a plant is a being that gets energy from the sun...), but a learner tapping plant needs exactly one word back, planta. So the spec collapses to a translator, and the boring option wins: Google Translate's API, with a free tier and cheap per-million-character pricing that holds until there are real users. The image-exercise version of the same audit ends differently: Google's image-search API is too limited, Pixabay looks best, and the decision is to wait, because a cheap small-image generation model probably lands within months and deletes the problem.
// trace: where this idea came from
- ↳ video diary @ 10:26 (no necesitas diccionario, necesitas traductor)
- ↳ video diary @ 11:43 (el API de Google Translate, la opción simple)
- ↳ Entry 221-2: Guinea pigs by design (el roadblock del clic en cualquier palabra)
- ↳ video diary @ 34:12 (postscript: la decisión se invierte, gana Gemini)
The feature entered the roadblock list as “add a dictionary,” and the requirements audit kills the noun: “tú para aprender un idioma no necesitas un diccionario, tú necesitas saber qué significa la palabra en tu idioma. Entonces, lo que necesitamos es un traductor” ▸ 10:26. The proof is the definition itself: a dictionary answering a tap on plant would recite a being with roots that gets energy from the sun ▸ 11:07, when the learner needs exactly one word back, “yo necesito es darle click a plant y que diga planta” ▸ 11:29.
no necesitas un diccionario, necesitas un traductor →
With the requirement that small, the boring implementation wins: the Google Translate API, half a million characters free and a flat dollar rate per million after ▸ 11:43, sized honestly, good enough until there are many users, at which point a different solution earns its complexity ▸ 12:15.
The image half of the same audit resolves the opposite way. Testers keep asking for pictures, so he surveys the APIs: Google’s image-search option is too limited to serve ▸ 12:39, Pexels rations daily, and Pixabay, with larger limits and vectors, looks like the best of the shelf ▸ 13:21. And then he declines to build: a cheap model for generating small images is plausibly two months away, and the moment it lands, the whole integration becomes dead weight ▸ 13:35. Two features, two audits, two verdicts: build the smallest true thing, and don’t build the thing the future is about to hand you…
Postscript, video 235: the translator verdict reversed. Running the numbers, Gemini beat Google’s Translate API on both counts, cheaper past the free tier and strictly more transversal, one call that handles a lone word, a whole sentence, a context-dependent word, and the user’s follow-up question ▸ 34:12. The boring option lost to the one that did more with less.