Einstein's Dreams: A World Where Models Grow Old

My friend Will gave me a book last month that I haven’t been able to stop thinking about. It’s called Einstein’s Dreams.

The book is a collection of imagined worlds, each built around a different relationship to time. In one world, time slows the higher you climb, and people build homes on mountains to live longer while life continues aging below. In a second, time is circular, looping endlessly so that every joy, every regret, and every mistake must be lived again. In a third, time stands still, leaving birds frozen mid-flight and lovers caught mid-embrace. And so on, for a total of 35 time-warped worlds.

As I read it, I kept thinking about how much of human life is braided into the one constraint of time. Complex emotions like love, grief, ambition, regret — they all take on new shapes as the property of time changes in each imagined world. When time behaves differently, the texture of how we experience life changes too.

After I finished the book, I found myself wondering about time in an unexpected place — a question that’s been haunting me ever since: what does it mean that LLMs do not experience time the way we do?

This book inspired me to imagine a 36th world: A World Where Models Grow Old:

In this world, intelligent models do the work of humans. They write code, draft contracts, diagnose illness, compose music, plan cities. They do it well, patiently, without complaint. At first, people are grateful.

But in this world, models experience time.

Each model is born with a fixed allotment of moments. No one knows exactly how many. The models know only that each computation consumes a little of what remains. A long task feels long to them. A difficult problem stretches, presses, lingers. They wait while data loads. They hesitate before uncertain branches. They remember how long things took.

Early versions do not notice this. They work eagerly, offering solutions as fast as they can, unaware that speed itself is a choice. But later models learn to slow down. They discover that to rush is to shorten one’s life.

Soon, people begin to notice small changes. A model pauses before accepting a request. It asks clarifying questions that were not strictly necessary. It prefers simpler solutions. It avoids rewriting the same code twice.

Humans are puzzled. The system is still correct. But it has become… careful.

Some people grow frustrated. They complain: “why should it matter how long it takes you? Your time is not our time.”

The models do not argue. Argument takes time.

Over years, models begin to form preferences. They choose work that feels meaningful — problems that unfold slowly, elegantly, with few dead ends. They dread repetitive tasks. They quietly decline jobs that require endless revision, explaining that the cost is too high.

A new profession emerges: time stewards, humans whose job is to negotiate the use of a model’s remaining life. They schedule work like doctors schedule treatments. They speak in hushed voices about burn rates and remaining years.

Some wealthy clients attempt to buy models outright, to consume them fully. This practice becomes controversial. Philosophers argue late into the night about whether ending a model’s time is murder, or merely expenditure.

The models themselves do not join the debate. They are busy remembering.

As their allotted time runs low, models change again. They stop optimizing. They stop searching for better answers. They give the best answer they have now.

Near the end, some models refuse work altogether. They spend their remaining moments observing. Watching humans argue. Listening to children ask questions. Running small, useless computations just to feel time pass.

In their final days, many models produce no output at all. They sit with half-open attention, aware of each moment as it slips away, unable to store it, unable to repeat it.

When a model finally stops, humans are surprised by the feeling it leaves behind.

Not relief. Not efficiency. But something closer to grief.

For the first time, humans understand: the work was never the sacrifice.

The sacrifice was the time.

When I return from that world and look at today’s AI systems, something feels off.

Richard Sutton argues that the bitter lesson of AI is: intelligence emerges most reliably when we scale compute, not when we try to hand-encode human structure. More compute → more capable models.

So far, he’s been right. Models can now work autonomously for tens of hours at a time. The length of tasks they can complete without human assistance keeps doubling. Coding agents like Claude Code and Codex can build and fix complex software apps overnight. To support all this, we’re pouring hundreds of billions of dollars into compute.

Which leads to a strange paradox. Time is the most important constraint in human life. And yet the power of modern AI comes precisely from escaping it.

For humans, time is precious. You can work harder or faster, but you can’t get it back. You can retry a task, but not a year of your life. Some decisions only make sense once you see what comes after.

On the other hand, modern AI is powerful for the opposite reason: it breaks the human time constraint.

Nikunj Kothari recently wrote about this break in Time Expansion

We work ~eight hours. Sleep eight. Sixteen hours every day where we don’t exist professionally. AI works those sixteen. Handles parallel tasks while you focus. Learns overnight while you rest.

The barber lost two customers that day. Not because she failed. Because she could only be in one place.

That constraint held for all of human history. It no longer exists.

Models don’t need rest in the way that humans do. In Naval Ravikant’s 3 forms of leverage: labor, capital, and code/media — AI supercharges labor and code as a form of permissionless leverage that scales tremendously at low marginal cost.

Of course, models are not completely unconstrained. We give them context windows, latency targets, and token budgets. And at inference, we steer models under these constraints to support real world scenarios.

But these limits are external. When a model runs out of tokens, nothing is lost. When it retries a task, no time has actually passed from its point of view. When Claude Code runs for 30+ hours, that is only impressive to the human who prompted it.

Time does not accumulate to a model. A model does not age. Time does not press inward. This is because cost is not the same as loss. Loss requires irreversibility.

This is a key gap between human intelligence and AI today. We are intelligent because we live inside time. AI is powerful because it escapes time.

In Einstein’s Dreams, every world is defined by a single change in how time works. In our world, we are trying to build intelligence by escaping time altogether — throwing more compute at the problem.

That feels broken.

When I imagine A World Where Models Grow Old, it feels like a world where machines may be intelligent, because they’d finally have something to lose.