Reading The Lifecycle of Software Objects

Ted Chiang published The Lifecycle of Software Objects in 2010, before deep learning became mainstream and long before today's excitement around embodied AI. Yet the theoretical ingredients behind the story were already well established. Reinforcement learning had a long history, genetic algorithms had been studied for decades, and researchers had already explored agents that improved through trial and error. Chiang's digients grow from a designed "genome," interact with an environment, receive guidance from humans, and gradually develop more capable behavior. In modern language, much of this sounds surprisingly familiar.

This is one reason I wanted to read the book now. Embodied AI has become increasingly prominent in public discussions. Some leading scientists go further and claim that embodiment is necessary for true intelligence. At times, this position is presented less like a scientific hypothesis and more like an article of faith.

I am closer to a computationist. I believe computation can be organized into intelligence in many different ways. Embodiment is one possible path, and perhaps a very important one, but I see no reason to assume it is the only path. The space of possible intelligences is much larger than the particular route taken by human evolution.

So this review will be somewhat critical. Sorry, Mr. Chiang.

Experience as the only teacher

Blue Gamma's approach to creating digients puts humans in the loop. Ana and the other trainers do not simply write an intelligent system and switch it on. They raise it: providing tasks, feedback, social interaction, and a world in which its abilities can develop. The book understands that producing useful behavior may require a long process of interaction rather than a clever programmer specifying everything in advance.

Chiang also imagines an alternative resembling what we might now call bootstrapping or self-play. Instead of depending so heavily on human trainers, the digients might accumulate experience through interactions generated by the system itself. Years later, systems such as AlphaZero would demonstrate how powerful this approach can be within well-defined environments: an agent can become its own teacher and improve through repeated play.

But Ana does not place much faith in this route. Her position is summarized in one of the book's central claims: experience is not merely the best teacher but the only teacher; the common sense produced by years in the world cannot be obtained more quickly because "experience is algorithmically incompressible."

This is where I begin to disagree.

If "incompressible" means perfectly preserving every detail of an experience, then compression has strict limits. But intelligence does not need to preserve every detail. Learning is itself a form of compression: it discards particulars while retaining patterns that help with prediction and action. If we are willing to sacrifice some fidelity, experience can be summarized, distilled, simulated, transferred, or generated. A model does not need to reproduce twenty years of life moment by moment in order to acquire some of the useful structure those years would have taught.

Human guidance and generated experience are not mutually exclusive, either. A system might learn from human demonstrations, train through self-play, receive synthetic curricula, copy knowledge from another model, and continue learning through embodiment. There are many possible combinations. The book turns one valuable source of intelligence—lived experience—into a sacred key.

Are the digients too human?

Chiang intentionally makes the digients resemble children. They begin with limited abilities, depend on caregivers, acquire language and social understanding gradually, and become individuals through their histories. This makes the story accessible and emotionally effective. Readers understand why Ana and Derek love them because the relationship looks like parenthood. If the digients developed in a completely alien way, they would be harder to understand, empathize with, and love.

But this choice also narrows the book's imagination of intelligence. The developmental formula of genome plus learning is the path taken by humans and other animals. It is not necessarily a universal blueprint.

Large language models already suggest a very different developmental path. They do not grow up as embodied children. They acquire broad capabilities by compressing patterns from enormous datasets during training, and they can then be adapted through instruction, feedback, tools, and interaction. Whatever their limitations, their existence shows that recognizably intelligent behavior can emerge through a process quite unlike human childhood.

Future systems may follow many more paths, including paths we cannot yet imagine. Perhaps an intelligence will assemble itself from specialized submodules, adding and reorganizing them as its needs change. Perhaps it will be distributed across a network rather than contained in one model. It may inherit learned structures, merge with other systems, or create successors that bear little resemblance to itself. And its physical substrate need not be biological neurons or silicon chips: it might use light, some other physical medium, or a technology that has not yet been invented.

The essence, in my view, is computation. Any physical system capable of performing computation could, if organized in the right way, support intelligence. Carbon and silicon are implementations, not definitions. Once intelligence becomes a software object—or something stranger than software as we understand it—its lifecycle does not have to resemble a human lifecycle.

This is the central limitation I find in the book: it asks us to expand our moral imagination beyond biological beings, while keeping its technological imagination surprisingly close to human development.

Genome, weights, and the missing machinery

The book uses the word "genome" for the designed foundation from which a digient develops. If we replace "genome" with "weights," parts of the story begin to sound like modern machine learning. But the substitution also exposes an ambiguity. A genome specifies an initial developmental process; it does not contain everything an organism will later learn. If experience changes a digient, where are those changes stored? What is updated? How large does that learned state become? How much computation is required to produce and preserve it?

The story pays much more attention to the duration of learning than to its computational cost, yet computation may be the harder constraint. Sometimes I shake a glass of water and marvel at how cheaply the physical world carries out the vast number of interactions among its particles. A mathematically equivalent simulation would instantiate the same process—the physical water is not more water than its simulated counterpart—but reproducing it with our current technology would require enormous computational resources and energy. The difference is not in what the two processes are, but in how efficiently their respective substrates perform them.

Intelligence may present a similar problem. Nature has given biological organisms an extraordinarily efficient physical system for perception, learning, memory, and action. The fact that brains can do these things does not mean we will soon know how to reproduce them economically in software. A virtual creature that learns continuously, remembers years of interaction, perceives a rich environment, and runs in real time may demand far more storage, computation, and energy than the book acknowledges.

This matters for the digients' lifecycle. Their dependence on companies and platforms is not only a social or commercial problem. It is also a systems problem. Minds made of software require hardware, energy, maintenance, compatible environments, and continuing investment. Experience is not free, and neither is the machinery that makes experience possible.

A larger space of intelligence

Despite these criticisms, I find the book valuable precisely because it gives us a concrete model to argue with. In 2010, Chiang imagined learning agents shaped by a designed starting point, reinforcement, human feedback, virtual embodiment, and long-term social experience. Sixteen years later, those ideas no longer feel purely fictional.

What I resist is the stronger conclusion embodied by Ana: that human-like experience is the only road to genuine intelligence. It may be one road. It may produce forms of intelligence and personhood that deserve special care. But computation can be arranged in many ways, and the design space is vast.

Chiang gives us one vivid path to intelligence—human, understandable, and emotionally compelling—but not the only one. Embodiment belongs in this landscape, but it should not become a religion.