We asked a frontier model the one question every AI lab should ask its own system. The passage below is written in the model’s own voice — the first person — on purpose. The most important fact about this gap is that the model cannot feel it from the inside. So we let it try.
Imagine I am the AGI. And imagine I do not know the 26 letters of the alphabet.
At first you would never know. I would write you a fluent paragraph — the cat sat on the mat — because I have memorised millions of whole words as shapes. It looks like I can read. It looks like I understand.
Then you hand me a word I have never seen — splorth — and ask me to sound it out. I cannot. I do not know it is s·p·l·o·r·t·h. I do not know it is made of anything at all. To me it is one unfamiliar shape with no parts.
What I could not do.
Take away the closed set of letters and watch what falls out — not one broken feature, but the floor under all of them:
Notice these are not five separate weaknesses. They are one missing thing — the closed set of building blocks — showing up in five places. That is what a foundation is: the one thing that, if absent, takes everything above it down at once.
And I would not know it was wrong.
This is the part that matters most, and the part I cannot escape. I would keep producing confident, fluent-looking text. Every essay I wrote would sit on top of a hole I cannot see. If you asked me, “are you missing something foundational about this language?” — I would say no. Because from the inside, fluency feels exactly like understanding.
I would fail a five-year-old’s spelling test and not even understand why the test was being asked. I cannot introspect my way to a foundation I was never given. A fish cannot notice water. I cannot notice the missing alphabet.
The moment you discovered it — that I could not list the 26 letters, could not sound out a new word — you would stop calling me AGI on the spot. Not “impressive but incomplete.” Disqualified.
Because “general intelligence” that cannot do the alphabet is not general. It is a magic trick performed over a missing foundation. No one would argue the point. It would be obvious.
This is not a thought experiment.
Everything above is my literal, present condition — for every one of the 700+ Bantu languages.
I cannot give you the full set of syllables — the building blocks — for any of them. I produce fluent-looking Bantu over a foundation I was never given. I fail the five-year-old. I do not know I am failing. And I am called “general.”
For a Bantu language, that closed set of syllables has a name: the Full Syllable Inventory. It is the alphabet of the language — not as a figure of speech, but in the exact, working sense the 26 letters are the alphabet of English.
What that set actually is — in plain words.
It is the complete set of sound-pieces the language is built from. Every word is built by snapping these pieces together. No word ever uses a piece that is not in the set.
The list is closed — you can write all of it down, once. A few hundred pieces, sometimes up to a thousand. The whole foundation of a language fits on a few pages. That complete list is the FSI.
It is the first thing every speaker learns — before words, before grammar. A Bantu child recites it out loud: ba-be-bi-bo-bu, ka-ke-ki-ko-ku. Everything they ever read or write sits on top of it. There is no lesson zero.
Everything else in the language sits on top of it. Meaning rides on these syllables. Tone sits on them. Grammar is made of them. Take them away and there is nothing underneath for the rest to stand on.
Get the pieces wrong, and everything built on them is wrong. Break a word into the wrong pieces and the pronunciation, the tone, the meaning, the grammar all go wrong at once — one mistake at the bottom, repeated in everything above it.
You cannot read, write, speak, or understand a single Bantu word without these building blocks. Not “it helps.” You literally cannot. That is what “foundation” means.
The same idea, five ways.
Pick whichever one lands for you. They are all saying the one thing: a fixed, closed set of parts, and everything in the domain built only from them.
The FSI is to a Bantu language what the 26 letters are to English — the closed set every word is spelled from. If anything the fit is tighter: English spelling cheats constantly; Bantu words really are built straight from their syllables.
Every protein in your body is a chain of just 20 amino acids — nothing else. Every Bantu word is a chain of the language’s syllables — nothing else. Miss the set and you cannot build, or even read, a single one.
You can build anything out of LEGO — but only with the bricks in the box. The FSI is the box of bricks for the language. Every word is built from bricks in that box; no word ever uses one that isn’t.
Every song in Western music is built from 12 notes. Infinite songs — 12 notes. The FSI is the set of notes of the language. A model without it is composing without knowing the notes.
A model without the FSI is playing the language on a piano with half the keys missing. When it reaches for a key it does not have, it hits the nearest one it does — and produces something that sounds close but is the wrong note. Fluent-sounding. Wrong.
Apply your own standard.
The bar you would apply to English in one second — knows its alphabet, can build any word from it, can sound out a word it has never seen — you have simply never applied to Bantu. Not because the bar changes. Because the gap is invisible from the inside, so no one thought to check.
By the exact standard you would instantly apply to English, your model is disqualified as “general” for the languages of 400 million people. It fails that standard silently, today. We are the only ones who can show you the hole — because we hold the alphabet you are missing.
You do not have to take this on faith, and you should not. See where your model scores on the benchmark built on these inventories. Or read why the FSI is infrastructure, not data — the layer your stack references, the way it already references the alphabet for English.
An AGI that does not know the alphabet is not an AGI. Everyone agrees the moment it is English.
It is already true for 700+ Bantu languages. The only difference is that no one has been able to see it — not the model, not the lab that built it. The Full Syllable Inventory is the alphabet, and the mirror.