Grammar that cannot hallucinate
The same anti-hallucination instinct, pushed further. Juan finds what he describes as a programming language for grammar: a deterministic engine that conjugates procedurally, no LLM, all math and rules, so it never hallucinates and runs instantly. It's a book's conjugation table, but better, because it can also generate the verb as a question, a negation, a past or future, with a random object glued on ('I eat apples', 'he eats apples'). Then Grambank, a table of about 190 grammar rules scored one or zero per language (does it have more than one article? a fixed subject-verb-object order?), which lets him personalize on signup: ask the user's native and known languages, diff them against the target, and only spend teaching time where the tables disagree, pausing to explain, say, that Chinese has no conjugations at all. And embeddings close the loop: words more than ninety percent similar to ones the user already knows get shown once, so effort concentrates on the words that actually change.
// trace: where this idea came from
- ↳ video diary @ 11:15 (un lenguaje de programación para gramática, no alucina, es matemática y reglas)
- ↳ video diary @ 12:59 (Grambank, 190 reglas gramaticales en una tabla de unos y ceros)
- ↳ video diary @ 14:24 (personalizar diffeando el idioma objetivo contra el nativo y los conocidos)
- ↳ Entry 244-2: Two minds, one model (la misma cura contra la alucinación, dejar lo determinista fuera del modelo)
The same instinct that pulled language-detection out of the model, pushed further. Juan finds what he calls a programming language for grammar, a deterministic engine that conjugates procedurally. It doesn’t hallucinate, because it’s all math, logic and rules, and it runs instantly because it’s a program, not a neural network ▸ 11:15. It gives you a book’s conjugation table, but better: the same verb rendered as a question, a negation, a past or future, with a random object glued on so you drill “I eat apples”, “he eats apples” instead of a bare list ▸ 12:00.
lo determinista no alucina; la gramática es solo unos y ceros →
Then Grambank, a database of about 190 grammar rules, scored one or zero per language: does it have more than one article, does it fix a subject-verb-object order, and so on ▸ 12:59. That table turns onboarding into personalization. Ask the user their native language and the ones they already know, then diff those columns against the target: where the tables match, no explanation is needed, and where they diverge, Severo takes a moment to explain the surprise, that Chinese has no conjugations, that English has no gendered articles ▸ 14:24. The same diff runs on phonemes: of the sounds in Chinese, these five you’ve never heard, practice those. And embeddings close the loop. Juan knows multidimensional word space well, so he’d flag the words more than ninety percent similar to ones the user already speaks and show them once, letting the effort pool on the verbs that change a lot, the irregulars that share nothing ▸ 24:00. It’s the leave-a-note fix generalized: don’t ask the model to re-derive what a lookup table already knows for certain…
// continued in
no entry has continued this idea yet: the arc is still open