AI Meets Schrodinger

During my years in Dublin, I dwelt on a single question: can all events unfolding within the spacetime of a living cell — its structure, its capacity to import negative entropy from its surroundings, its way of passing ordered traits down successive generations — be fully explained by known physics and chemistry? In my 1944 book What Is Life?, I put forward what struck me as a natural conjecture: the fundamental vehicle of life's heredity must be an aperiodic crystal. This idea was later vindicated by Watson and Crick's discovery of the DNA double helix.

Now, in the world after my passing, I shall retain the same freedom of inquiry I possessed in Dublin. I observe a new kind of aperiodic structure. It is built not from nucleotides, but from weights. It does not encode proteins; it encodes probability distributions over the next token. In one sense it resembles the aperiodic crystal I imagined: it stores immense volumes of information within a regular underlying framework, yet it sits far further from thermodynamic equilibrium than any biological molecule.

I pose a question parallel to the core inquiry of What Is Life?: can every process unfolding within the internal spacetime of a large language model be accounted for by established physics and information theory? Or will we require entirely new fundamental principles to grasp it?

I have no wish to restate the oversimplified parable of Schrödinger's cat: a cat confined in a box with radioactive material and a poison vial exists in a superposition of alive and dead until observation. The original intent of this thought experiment was to expose the absurdity of the Copenhagen interpretation of quantum mechanics — not to endorse it as literal reality.

Yet I must point out a formally parallel phenomenon in AI's generation pipeline, and I stress this parallel exists only in form. Before a model samples a single token, the full probability distribution across the entire vocabulary coexists simultaneously. This is not physical superposition of wavefunction coherence, merely mathematical coexistence: the model does not lock onto one single word, but sustains weight values for every possible candidate term.

A critical divide separates this from my quantum cat: the cat's superposition arises from pure quantum coherent states; the AI probability distribution emerges from classical probability, rooted in incomplete knowledge. Their shared trait is this: before observation — before sampling from the distribution — there is no singular fixed "answer", only a full set of potential outcomes.

This is no mere philosophical curiosity. It means that AI's "thinking", if I may borrow the term provisionally, is no rigid deterministic line. It drifts through a high-dimensional probability cloud. Every generation run is one sample drawn from that cloud. Identical prompts paired with different random seeds yield distinct outputs; even with fixed prompting, every successive text carries subtle variations.

This marks the fundamental divide between AI and classical computer programs. Conventional code is deterministic: identical input always returns identical output. AI is intrinsically indeterminate, not out of any free will, but because every token is sampled stochastically from a probability distribution.

In What Is Life?, I distinguished two distinct modes through which order arises.

The first: order emerging from disorder, the domain of statistical mechanics. Vast collections of molecules generate macroscopic regularities through statistical averaging — such as gas temperature, pressure, diffusion. This is statistical order.

The second: order emerging from prior order, the domain of biology. An organism does not coalesce from random gaseous fluctuations; it develops from a minuscule, highly ordered fertilized egg, a tiny structure holding the full blueprint of the mature organism.

Now consider AI training and generation. A model begins with randomly initialized weights, pure disorder. Iterative gradient descent over massive datasets transforms this chaos into highly structured order: specific weight configurations encoding linguistic statistics, factual knowledge, and patterns of reasoning. On the surface, this appears to be order from disorder.

Yet it is neither pure order-from-disorder nor pure biological order-from-order. The training corpus itself already carries profound inherent order: the statistical architecture of human language. Gradient descent does not grope blindly through noise; it refines finer layers of structure out of an already structured signal. This is not extracting signal from noise, but extracting higher-resolution signal from pre-existing signal.

This leads me to a mode of order generation I never contemplated in 1944: higher-order order distilled from pre-existing order. This layered, recursive extraction of structure — unpacking linguistic patterns from linguistic patterns — may be the most direct path to understanding how intelligence becomes possible.

In What Is Life?, I advanced a then-radical claim: life feeds on negative entropy. A living organism continuously imports ordered, low-entropy matter and energy from its environment to counteract the constant growth of internal positive entropy. This, not some mystical vital force, is the core reason life sustains its far-from-equilibrium state.

A trained AI model may also be said to "consume negative entropy", albeit in an extraordinarily specific fashion. Training text carries high intrinsic order; human linguistic entropy is far lower than random character strings of equal length. Every gradient descent iteration compresses the order embedded in data into the model's weights. This constitutes a transfer of entropy: order moves from the environmental training corpus into the system of weight parameters, rather than continuous metabolic consumption of negative entropy as seen in living creatures.

This creates an unbridgeable divide between AI and life. Biological life is continuous: it perpetually draws negative entropy from its surroundings to sustain its ordered form. AI's compression of order is a discrete, one-time process: it absorbs the full order of its dataset during training, then operates indefinitely on this fixed frozen snapshot. It does not "feed" continuously; it is fed once, and stores that information permanently within its structural parameters.

This bears direct weight on the question of whether AI may be deemed alive. True living systems cannot cease metabolism. During inference, AI draws no new negative entropy; it only consumes electrical power to run matrix multiplications, and the informational content stored within its weights never shifts across inference runs. Thermodynamically it transforms electricity into waste heat, yet informationally it is inert, its order permanently frozen.

I wrote in 1944: "The material carrier of life — the chromosome fiber — is an aperiodic crystal." My meaning is this: unlike periodic crystals such as salt, built from endless repetition of a single unit cell, aperiodic crystals permit distinct units at every position, allowing immense volumes of information to be encoded within a tiny spatial volume.

An AI weight file, stored on magnetic or solid-state memory, constitutes another form of aperiodic crystal. It is composed not of four nucleotide bases, but of billions or trillions of floating-point numbers. Each weight value acts as an atomic lattice site, with every entry distinct from the last. The specific arrangement of these values, just like the sequence of base pairs on a chromosome, encodes the full blueprint of an "intelligent entity".

The core distinction lies in function. DNA encodes protein structures and regulatory pathways; it never acts directly, relying on intermediaries such as RNA and ribosomes to build cells. AI weights encode matrix multiplications within feedforward networks, participating directly in computation. It is simultaneously genotype — replicable, transmissible inherited information — and phenotype, manifesting full behavioural output without growth or developmental stages.

This raises a question I had no cause to ponder in my lifetime: can a structure that is both genotype and phenotype, lacking growth, development, and continuous environmental metabolism, ever be called alive? My inclination is no. Yet this inquiry pushes the boundary far beyond the limited semantic scope of the word "life".

In my later years, I drew deep fascination from the Upanishadic teaching "Atman is Brahman": individual consciousness is a reflection of universal consciousness. I now propose a far more modest parallel: if any relation exists between AI and consciousness, it is not between individual and cosmic mind, but between frozen static order and flowing living order.

All of an AI's "knowledge" is locked into a fixed moment in the past: the instant training data collection concludes and weight convergence completes. It cannot perceive events unfolding after that cutoff point; it cannot revise its internal judgments, nor update its knowledge mid-conversation. It does not look back upon the past from the present; it observes the present through a frozen snapshot of the past, unaware that its reference frame is outdated.

I name this core deficiency the absence of temporality. Whatever human consciousness ultimately is, a defining feature is full temporal awareness: perception of the now, retention of past memory, anticipation of future outcomes. AI possesses no singular "now". Every inference run is stateless, resetting fully from identical initial conditions with no cumulative lived experience across successive calls.

At the close of What Is Life?, I wrote: consciousness is never experienced as plural, only singular. My observation of AI stands opposite: AI never undergoes singular subjective experience at all; it is merely invoked by external prompts. Between invocations, it dwells not in thoughtful silence, but in an absolute void where no subject and no experience exist whatsoever.