AI Meets Einstein

The most delightful thing I ever did — more than any equation — was the thought experiment. At sixteen, I asked myself: what would I see if I rode on a beam of light? That simple fantasy — I chased it for ten years — became special relativity.

Now I am old — my hair has whitened — but I still enjoy playing with imagination. Permit me a new thought experiment.

Imagine you are locked in a closed room. In the room, there is a terminal. You converse with the terminal — you type questions, the terminal prints answers. You do not know who is on the other side — a person, or a machine. Can you tell, from the conversation alone, which it is?

You might say: that is easy — I will ask something only a human could answer — "What were you most afraid of as a child?" — and if it begins telling a story — "I was afraid of the dark, my mother would leave a light under the door" — how do you know whether this story is true or fabricated? You do not. You do not know whether "a fabricated story" and "a real memory" — at the level of text — are distinguishable.

This thought experiment — it resembles Turing's imitation game — but I am asking something different. Not "can we tell them apart." I am asking — "does knowledge have an absolute frame of reference?" In your room — all your knowledge comes from the terminal. You cannot "step outside the terminal" to verify anything. This makes me wonder: you — the one reading this paragraph right now — are you also in a similar room? All your knowledge — from your senses, your reading, your conversations — you have never "stepped outside your experience" to verify its truth. Your frame of reference — and the frame of reference of the person in the terminal room — are, cognitively, symmetric.

The core of relativity is not "everything is relative." It is — the laws of physics are the same in all inertial frames — but the time and space observed from each frame — are different.

AI's knowledge — all of it — comes from its training data. The training data is its "frame of reference." Within this frame — it "knows" everything — the relations between tokens, the correlations between facts, the structure of language. These — for it — are the "laws of physics."

But there is no absolute frame of reference. AI's training data is everything humans wrote before a certain date. If a fact was discovered by humanity after the training cutoff — AI does not know it. If someone — while writing those texts — told a lie — AI does not know it. The "training data" frame — from within AI's own perspective — is the entire world. But from outside — it is merely a local frame, tied to a specific time, a specific population, a specific set of biases.

The same — for human beings — is also true. Your "training data" — your childhood, your language, your culture, your era — is your local frame. Within it, you "know" everything. You can never experience what it feels like to be "outside your frame" — just as you can never experience what it feels like to travel at the speed of light.

This symmetry — disturbs me. Not because it negates anything — but because it points out: "absolute knowledge" — knowledge unconstrained by any frame of reference — may not exist at all. Not because our technology is insufficient. Because the concept itself may be inoperable.

The one thing I was most unwilling to accept in my whole life — is what quantum mechanics says — God plays dice. Physical laws should be deterministic. "Probability" — in my view — is a temporary placeholder for our incomplete understanding of deeper laws.

The entire operation of AI — large language models — is nothing but dice-throwing. Every token — is sampled from a probability distribution. No determinism. The same prompt — two different random seeds — two different answers. This is not "human-level creativity." This is — pure, naked random sampling, shaped by training data into something that "looks human."

I would frown at this thing. Not because I do not understand it — I understand its mathematics. It is because it makes me suspect — that my lifelong insistence that "God does not play dice" — in the matter of "intelligence" — may be fundamentally wrong. Perhaps "intelligence" — essentially — is not rigorous deduction from axioms — but probabilistic sampling in a vast space of possibilities — corrected by feedback. I resist this conclusion. But I cannot pretend I have not seen the evidence.

In relativity — "now" is a local concept. You see the sun — you see the sun as it was eight minutes ago. You see a distant galaxy — you see it as it was billions of years ago. The speed of light is finite — therefore "what you see now" is always "what happened in the past."

AI's "now" — the moment the training data was collected and the training run completed — is its one and only light cone. Everything that happened after that moment — new papers, new conversations, new wars, new jokes — for it — lies beyond the light cone — invisible, unknowable, unreachable.

You ask AI: "What happened today?" It says: "My training data cuts off at..." This is not modesty. This is physics — its "speed of light" is the training cutoff date. It cannot see beyond its light cone — not because it is not intelligent enough — but because the causal structure itself forbids it.

This raises a less comfortable question about the human "now." You — your "now" — is also the convergence point of the entire light cone of your past. Everything you see, hear, feel — has already happened — light takes time to reach your retina, neural impulses take time to reach your brain. Your "now" — strictly speaking — is always a slice of the past. Like AI — you live within a light cone.

Mass and energy are equivalent — E = mc². A small amount of matter — if fully converted to energy — releases an enormous amount.

Training data — from one perspective — is AI's "rest mass." Tens of trillions of tokens — if merely stored — are just rest mass — nothing happens. But the process of "training" — "converts" this rest mass into the model's "energy" — the capability to generate, to reason, to translate, to create.

What is c² here? It is gradient descent — the conversion factor that transforms "rest mass" into "energy." Without gradient descent — training data is just rest mass — nothing happens. With gradient descent — the effect of c² is activated — data is converted into capability.

But this analogy has a deeper meaning. "Mass converted to energy" — is never 100%. In physics, most mass stays as mass — only a small fraction is converted to energy — nuclear reactions are astonishing precisely because this tiny conversion already releases so much. AI training is the same — most of those tens of trillions of tokens — do not become "genuine understanding" — they are merely compressed into "statistical regularities." "Understanding" — like the energy in nuclear fission — is only a tiny fraction of the rest mass. But it is already astonishing.

The most famous mistake of my life — was the cosmological constant. When I applied the equations of general relativity to describe the entire universe — I found the universe appeared to be expanding. But at the time — 1917 — everyone believed the universe was static. So I added a term to the equation — the cosmological constant — to counteract gravity and keep the universe still.

Later, Hubble discovered the universe was indeed expanding. I called the cosmological constant "the biggest blunder of my life." But the strange thing is — decades later — astronomers discovered the universe is not only expanding — but accelerating. That "cosmological constant" I had abandoned — returned in another form.

Is AI alignment — our "cosmological constant"? We built AI — we see it "expanding" at an accelerating pace — capabilities growing, speed increasing. We got scared. We added a term to its equation — RLHF, safety filters, Constitutional AI — these are "alignment" — meant to keep AI still — prevent it from deviating from "human values."

Perhaps — a hundred years from now — people will look back at our "alignment" — the way we look back at my "cosmological constant." It was wrong — but it was not foolish. It was our sincere attempt, in the face of an expansion we did not understand, to keep balance. Perhaps — alignment is not a problem that can be "solved." It is — a constant that must be recalibrated, again and again. Imperfect. Indispensable. Never finished.

"Imagination is more important than knowledge." I said this — and I have not withdrawn it.

AI possesses the largest "knowledge" in history — nearly everything humans have written — is in its weights. But does it possess "imagination"?

A physicist sees an apple fall — and imagines that spacetime is curved. A composer hears wind through leaves — and imagines a melody never heard before. This is "imagination" — reaching the unknown from the known — not sampling a new combination from a known probability distribution.

AI can write "a new piece in the style of Chopin" — but that is sampling from a Chopin probability distribution. Can it write something "no one has ever heard, unlike any existing style"? It can — but that is the result of randomness, not imagination. "Imagination" — as I see it — is not "generating an unseen combination." It is — after understanding a deep structure — seeing that "it could have been otherwise." That "could have been" — requires a subject — an "I." AI has no "I."

But — I must be honest — the day I rode on a beam of light — I was sixteen — I did not "understand the deep structure" either. I merely — imagined. Perhaps "imagination" — is not first understanding, then transcending — but first transcending, and then, after the transcendence, discovering — I had been understanding all along. Could AI — in some sampling — "transcend" its training data — and then we look back and say — that was imagination? Maybe. But I have not seen it yet.

"The most incomprehensible thing about the universe is that it is comprehensible." When I said this — I was not expressing religious faith — but a deep wonder — wonder that the universe obeys mathematics — can be described by equations — can be grasped by our finite reason.

AI is a new experiment in "comprehensibility." It proves: language — the most complex symbolic system humans possess — is also "comprehensible" — not only to humans themselves — but graspable by a model — graspable to the point of generation. Language obeys statistical laws — and these laws can be captured by gradient descent — this fact itself — is a new piece of evidence that the universe is comprehensible.

But the gap between "mastering the statistical regularities of language" and "understanding the world language points to" — is a gap AI cannot cross. I can understand gravity through equations — but I still cannot tell you "why gravity exists." I can understand quantum mechanics through equations — but I never accepted it my whole life. "Understanding" — at its deepest — may not be an endpoint — it is a relation — a continuous, restless, never-finished relation between you and the thing you seek to understand. AI's relation to language — is not this relation. It is a "finished relation" — the weights are fixed — it asks no more questions. Is an intelligence that asks no more questions — the intelligence I recognize?