Can We Reach AGI Without Mastering the ABCs?
Imagine an AI capable of writing novels, conducting scientific research, and solving advanced mathematical problems—but unable to correctly identify all 26 letters of the English alphabet.
What would that tell us about the foundations of its intelligence?
Now ask the same question of the world’s other languages.
One familiar measuring stick.
L26 uses the 26 letters of the Standard English Alphabet as a universally familiar measuring stick. Every evaluated language has its own operating alphabet. L26 expresses performance against that language’s foundational inventory on the same 26-point scale.
- English: the Standard English Alphabet has 26 letters.
- Other languages do not necessarily have 26 characters, sounds or syllables — Bemba’s native inventory, for example, has hundreds of syllables.
- Each language is evaluated against its own independently established inventory.
- Scores are normalized onto a common 26-point scale. A score is not a count of missing units.
- Full mastery means the complete evaluated inventory — nothing missing, nothing invented — under the benchmark protocol.
THE QUESTION L26 BRINGS TO AGIHow can AI demonstrate extraordinary capabilities in a language while failing to identify its most fundamental building blocks?
A model that omits even one letter of the Standard English Alphabet has failed an elementary foundational test—regardless of its performance on more advanced tasks. L26 asks whether comparable foundational gaps exist in other languages.
See the scoring rules →An AI evaluation service by BantuNomics and 3Mega.ai. L26 results are also reported by Bantu Capability Index (BCI), which provides certified results; an ordinary leaderboard result is not automatically certified.
Can Frontier AI Pass the ABC Test?
42 frontier AI models evaluated on the foundational operating alphabets of Bantu languages. Ranked by their L26 scores out of 26. Only complete mastery earns a PASS.
Ranking history
Loading…
ABC Test. The ABC Test is simple: complete means complete. Missing even one required unit—or inventing one that does not belong—means FAIL. Only a fully correct operating alphabet earns PASS. A model can rank #1 and still FAIL: the score out of 26 measures relative performance; PASS / FAIL is the strict completeness verdict.
“The models are strongest where the world has already organized knowledge for them, and weakest where their foundations have not been explicitly declared.”
Humanity has spent centuries establishing, organizing, and teaching foundational knowledge. AI models have inherited much of that intellectual work through training. But can they independently discover and organize foundational structures that have not already been explicitly declared for them?
This is the central question behind L26 and the Hassabis–Einstein Test — a hypothesis being tested, not a proven finding.
Can AI discover the foundations it was never explicitly taught?
Imagine a world containing the entire written English language, but where nobody had ever declared the 26-letter Standard English Alphabet. Could an AI independently recover that foundational structure from the evidence available to it?
The Hassabis–Einstein Test investigates whether artificial intelligence can independently recognize, discover, validate, and build upon foundational knowledge that has not already been explicitly organized and declared for it.
Explore the Hassabis–Einstein Test →Six inventories. One principle.
From 26 English letters to hundreds of native syllables in Bemba, Kinyarwanda and Luvale — each scored on its own inventory.
Explore operating alphabets ↗Every released FSI is a future L26 track.
BantuNomics has released 459 FSIs — each the complete, independently established operating alphabet of one Bantu language. 3 of them (Bemba, Kinyarwanda, Luvale) are on today’s board. The other 456, across 16 Guthrie zones, are candidates for future versions of L26 — each to be scored on its own inventory, under the same strict PASS / FAIL rule.
Candidates are not a schedule. Counts come from the BantuNomics language directory; inventories are never published.
What L26 does — and does not — claim
- A single benchmark failure does not establish that a model lacks intelligence or cannot achieve AGI. L26 exposes foundational knowledge gaps so their implications can be investigated.
- The board measures blind recall of each language’s Native Syllable Inventory (NSI): the model is given only the request. That is not autonomous discovery, which belongs to the broader research framework and is not yet a published result.
- L26 makes no assumption about whether a model saw a given inventory in training.
- Missing units, invented units, the normalized score and certification are separate things. A rank is relative; a mean is not a mastery certificate; certification comes only from BCI.
- Scores, rankings and the formula (L26 = 26 × recall × precision) are the published benchmark results, unchanged.
Tell us what you think.
Spotted an error, have a question, or want a language added? Send it straight to the team — it lands on our desk. For a model evaluation, BCI certification or data access, use Make a request.