Demis Hassabis proposed an Einstein test for AGI. Here is one you can run today.

Why general relativity, not the 1905 miracle year

The detail that makes the test sharp is the one easy to miss: Hassabis chose general relativity (1915), not Einstein's 1905 "miracle year." That is deliberate. The 1905 results — special relativity, the photoelectric effect, Brownian motion — were, in a sense, in the air: Lorentz and Poincaré had the mathematics of special relativity all but assembled; Planck had opened the quantum door. A brilliant synthesist might have reached them from the live puzzles of the day.

General relativity was different in kind. No experiment was demanding it; no one else was reaching for it. Einstein imagined that gravity was the curvature of spacetime — a structure no instrument had shown and no dataset contained — and only then went looking for the mathematics to carry it. That is the faculty the test is really probing: not interpolation from the surface of what is known, but the recovery of a deep structure the evidence never states.

What the test actually measures

Strip it to its core and Hassabis's test asks three things of a mind, in order: that it can take a body of evidence; that it can sense there is a deeper structure underneath the evidence than the evidence states; and that it can recover that structure — name it, exactly — using nothing but the priors available. Memorisation fails this. Fluent recombination fails this. Only genuine abstraction passes.

Hold that shape in mind — recover the deep structure the priors imply but never state — because it is not unique to physics. It is the shape of every act of real understanding. And it can be posed at scales far below cosmology.

The trouble with the Einstein test

As a definition, the Einstein test is beautiful. As an instrument, it is almost unusable — and honesty requires saying so.

You cannot really build a model with a clean 1911 knowledge cutoff; relativity has been in the training water for a century, impossible to fully drain. You have no grader: there is no oracle that can certify "yes, this derivation is general relativity, arrived at honestly." And history gives you the answer only once — there is exactly one general relativity, discovered once, so the test cannot be run at scale, across many trials, with a deterministic score. It is a destination you can point at, not a dial you can read. To actually measure the faculty it describes, you need a test of the same shape that is reproducible, gradeable, and cheap.

The same faculty, scaled to language

That test exists, and it is sitting in plain sight. Ask a model to produce the complete operating alphabet of a language — for a Bantu language, the closed set of legal syllables, the Full Syllable Inventory — using only the words it has already read. No one has ever written that alphabet down completely; for 400 million people it was never published. So the model cannot retrieve it. If it is to have the alphabet, it must distill it — recover the closed standard underneath the words, the structure the corpus implies but never states.

That is Hassabis's faculty, scaled down: from a body of evidence (the words), sense that a deeper structure exists (a finite operating alphabet), and recover it exactly (every legal syllable, and — harder — every illegal one). It is the Einstein test in miniature: not "produce general relativity," but "produce the alphabet no one wrote down." Same move. Smaller stage. And you can run it in two minutes.

The mapping, line by line

The analogy is not loose. It is structural, term for term:

• The priors — Einstein had 1911 physics; the model has the corpus of the language. Both are given; nothing more is allowed.
• The hidden structure — Einstein had to recover curved spacetime; the model has to recover the closed operating alphabet. Neither is stated in the evidence.
• The negative space — relativity required imagining what no instrument had measured; the alphabet requires knowing the syllables a language systematically excludes, which no text records (absences leave no trace). Both demand knowledge of what is not there.
• The leap, not the lookup — both reward forming the right new hypothesis, and both punish confident interpolation from the surface. A model that pads its answer with plausible-but-illegal syllables is doing exactly what a model "deriving" relativity by recombining textbook phrases would do: mimicking the form of discovery without the substance.

Why this one can be graded — and the Einstein test can't

Here is the move that turns a definition into an instrument. The Einstein test has no answer key. The Alphabet Test does. For 459 Bantu languages, BantuNomics holds the verified operating alphabet — native-curated, audited, versioned — so a model's attempt can be scored deterministically, syllable by syllable, against ground truth: how much of the real alphabet it recovered, and how much it invented. No LLM judge, no hand-waving — set arithmetic against a native-confirmed standard.

That is the whole difference between a thought experiment and a benchmark. Hassabis can describe the faculty; we can measure it — repeatedly, across languages, with a number. The reason such a test could not exist before is the same reason it is hard for models: the answer key — the closed, verified alphabet — did not exist. Building it is what makes the test runnable at all.

And it is already failing

Run it across the frontier and Hassabis's verdict holds, now with a scoreboard. Asked for a single Bantu language's alphabet, the most capable models either refuse (candidly — "I'd only be guessing"), return a fluent list missing most of the inventory, or pad it with invented syllables they cannot flag as invented. Not one produces a complete, clean alphabet — and that is on one of the best-documented languages, where the priors are richest. "It's clear today's systems couldn't do that" is no longer a claim about physics in 1915. It is a measured result about language in 2026. You can reproduce it on your own model in two minutes.

A runnable rung on the same ladder

None of this claims the Alphabet Test is the Einstein test. General relativity is a summit; an operating alphabet is a foothill. But they are on the same mountain — both ask a mind to recover deep structure from priors alone — and a foothill you can actually climb tells you something a summit you can only point at does not. A model that learns to distill the unwritten alphabet of a language from its words is exhibiting, in the small, the very faculty Hassabis's test demands in the large. The Alphabet Test is an early, gradeable rung on the ladder to the Einstein test — a way to watch the capability arrive, language by language, instead of waiting for one unrepeatable cosmological miracle.

And there is a clean asymmetry that makes it more than a benchmark. The structure a model cannot yet distill can be handed to it — the operating alphabet, with its boundary stated and its provenance attached. You can score a model against the standard; you can also give it the standard. Both, because the standard now exists.

The bottom line

Demis Hassabis has pointed at the right destination: intelligence is the recovery of deep structure from the priors alone, not the recombination of the surface. His Einstein test names that destination beautifully — and cannot be run. The Alphabet Test is the same faculty, scaled to language: runnable, gradeable, and failing today — the odometer for the very capability his test describes.

If you want to know how far a model is from the Einstein test, you do not have to wait for general relativity. Ask it for an alphabet that was never written down, and score what comes back. Start with your own model.

Run the Alphabet Test Read: the distillation thesis The case for the FSI Initiate Partnership