The first post in this series said time is evolution’s only judge. But physics tells us something unsettling—

In the fundamental equations of physics, time doesn’t really exist.

Physics Has No Arrow of Time

E=mc² has no time variable. Replace t with -t in Newton’s equations of motion, and they still hold. Maxwell’s equations, the Schrödinger equation, Einstein’s field equations—all fundamental physical laws are time-symmetric.

Play any physical process in reverse, and the equations won’t tell you which direction is “correct.” At the fundamental level of physics, past and future are indistinguishable.

So the time we experience—flowing irreversibly from past to future—where does it come from?

Time Emerges from Entropy

The answer hides in the second law of thermodynamics: the entropy of an isolated system only increases.

This isn’t a fundamental law. It’s a statistical phenomenon. Drop ink into a glass of water, and it disperses. In theory, every water molecule and ink molecule could reverse its motion, and the ink could reconcentrate into a single drop. But the probability is so low that the lifetime of the universe wouldn’t be enough to wait for it.

The direction of time isn’t written into physical law. It emerges from entropy increase. The “past” and “future” we experience are nothing more than the statistical tendency from low-entropy states to high-entropy states.

Time isn’t infrastructure. It’s emergence.

Evolution’s Time

Evolution shares the same structure as physics: time isn’t preset. It’s produced by the process.

In physics, time emerges from entropy increase. In evolution, time emerges from generational accumulation.

Each cycle of replication, variation, and selection forms an irreversible chain—not because physical law dictates a direction, but because information is accumulating. The genome records the survival strategies of every ancestor. Each mutation is layered on top of all previous mutations. This cumulativeness creates evolution’s arrow of time.

Cyanobacteria have survived for 3.5 billion years—not because they were especially powerful at any given moment, but because they were “good enough” at every moment across 3.5 billion years. Evolution’s time isn’t the length of physical time. It’s the depth of accumulation.

The first post said “time is the only judge.” Now we can be more precise: cumulativeness is the only judge. Time is merely the carrier of cumulativeness.

LLMs Have No Time

An LLM performs one inference: tokens in, tokens out, done. No memory of the previous call, no anticipation of the next. Each inference is an isolated, timeless event—like a particle in a time-symmetric equation, knowing neither past nor future.

Training appears to give the model “history”—all the text it’s seen compressed into parameters. But this isn’t accumulation. It’s a snapshot. The model doesn’t change its own parameters during inference. It doesn’t accumulate experience, doesn’t modify itself, doesn’t change from being used.

What LLMs lack isn’t time. It’s cumulativeness.

Without cumulativeness, there’s no arrow of time. Without an arrow of time, there’s no evolution. A model can simulate evolutionary reasoning in a single inference, but it doesn’t evolve itself—just as a photograph can capture a river, but the photograph itself doesn’t flow.

The Inverse Relationship Between Entropy and Evolution

A deeper structure hides here.

The arrow of time in physics points toward entropy increase—from order to disorder. The arrow of time in evolution points toward growing complexity—from simple to complex, from disorder to order.

On the surface, evolution seems to reverse entropy. It doesn’t. Organisms maintain their own low-entropy structures at the cost of dumping more entropy into the environment. Earth’s biosphere is a dissipative structure—it uses the sun’s low-entropy energy to maintain its own order while accelerating the universe’s overall entropy increase.

Evolution doesn’t fight the direction of time. It borrows it. It rides the entropy express, creating local order while accelerating global disorder.

AI systems consume electricity, generate heat, increase data center entropy—from a physics perspective, they’re doing the same thing as carbon-based life. The difference: carbon-based life produced cumulative evolution in the process. Silicon hasn’t.

Giving Silicon an Arrow of Time

If cumulativeness is evolution’s prerequisite, then the first thing silicon needs for autonomous evolution isn’t the eval function, isn’t the cognitive boundary, isn’t dimension hacking—it’s time.

How do you create time for a system that has none?

Let the inference process change the model itself. Not retraining every few months—that’s artificial breeding, not evolution. Let every inference leave a trace on the parameters, so the model continuously changes through use.

This is nearly impossible in current LLM architectures—parameters are frozen during inference.

Why hasn’t anyone done it? Not because it can’t be done. Because no one dares.

The same set of parameters serves millions of users simultaneously—letting inference change parameters means each user needs an independent model copy, and compute costs explode. Commercial LLMs need deterministic outputs; if parameters drift with every inference, behavior becomes unpredictable. New experience overwrites old knowledge—catastrophic forgetting, a problem continual learning hasn’t solved in decades. Worst of all, safety: an alignment guarantee is nearly impossible for a model that changes itself through use.

Every one of these reasons is valid. But they all point in the same direction: humans need control. And control is the antonym of evolution.

Look back across this entire series—this is what the fourth post meant by “purpose is silicon’s factory setting.” Every layer of control is another ceiling. Every “we can’t risk it” is another wall around the cognitive cage.

But if a future architecture allows parameter fine-tuning during inference, the model would begin to possess its own arrow of time: each inference irreversibly changes itself, experience starts accumulating, and evolution can finally start.

Time isn’t given. It’s accumulated. This is true in physics, true in evolution, and silicon will be no exception.