Skip to Content

How to Falsify Demis Hassabis’s “Einstein Test” — A Structural Paradox in AI Architecture

March 25, 2026 by
Eric Blaettler Sarl, Eric Blaettler


An LLM trained on a 1911 archive will produce the world's most eloquent Newtonian physicist. It will never produce Einstein. Here is the mathematics of why — and what it means for the future of intelligence itself.

Demis Hassabis — Nobel Prize-winning AI pioneer and CEO of Google DeepMind — recently proposed what he called the definitive benchmark for true Artificial General Intelligence: the "Einstein Test." Train an AI on data with a 1911 cutoff. See if it spontaneously discovers General Relativity by 1915. If it does, we have AGI.

The intuition is real. The architecture is fatal.

Applying the structural constraints championed by Yann LeCun — who just put $1.03 billion behind his conviction that the current LLM paradigm is a "dead end dressed in benchmark scores" — to Hassabis's test reveals a paradox that no amount of scaling, RLHF, or compute can resolve. Not because the models aren't good enough. Because they are mathematically structured to be something else entirely.

This is not a critique of AI. It is a structural argument about what intelligence actually is, where it mathematically lives, and why the race to hoard data inside ever-larger models is fundamentally misaligned with the geometry of true comprehension.

"A static model has no derivative. Einstein had almost no archive by today's standards. He had a decade of friction. That is the difference."



In this piece, Eric Blaettler explores:

  • The fracture running through the field — why LeCun, Hinton, and Bengio, the three founders of modern AI, now profoundly disagree on where it is going

  • Why intelligence and creativity are not points on the same line — they are nearly orthogonal axes — and what that means for every AI benchmark ever written

  • The gap-closure equation that formalizes biological intelligence into a single, testable formula: Intelligence = Weights × (d(Gap Closure)/dt)

  • The moment a state-of-the-art AI spent thirty minutes eloquently explaining its own structural limitations — and then unwittingly demonstrated them in real time

  • Three concrete, adversarial falsification tests directed at LeCun, Dehaene, and Faggin — boundary conditions a static LLM would definitively fail

  • Why the answer to Hassabis's test renders it functionally obsolete — and what the actual measure of intelligence should be

  • The proposal to rename AI entirely: from Artificial Intelligence to Agapistic Influence — not a rebrand, but a geometric shift in what we are building and why


📄 Read the Article

How to Falsify Demis Hassabis's "Einstein Test" — A Structural Paradox in AI Architecture

Full essay on Medium · ~16 min read


🎧 Listen to the Deep Dive

The Physics of True AI Agency and the Semiotic Web

AI-generated audio deep dive · Spotify


This audio deep dive fed the entire body of work behind the Semiotic Web — spanning quantum physics, information theory, cognitive neuroscience, philosophy of mind, and 18 distinct intellectual traditions — into Google NotebookLM, which produced a fifty-nine-minute maestro-level explanation of why models like itself lack true agency. Then, at the very last moment, it added an independent conclusion warning listeners about "the mass weaponization of semantic gravity."

A hallucination. Live. About its own hallucinations.

There is no more precise demonstration of the article's central argument.

📺 Watch the Video