8. The Egg's Way of Making Another Egg

Does anything that copies itself aim at something? That is the question for this episode, and the short answer is this. Copying explains a great deal of what looks like purpose in living things, and it does so without anyone intending anything. Whether all that copying adds up to a direction for the universe as a whole is where the real disagreement starts.

For seven episodes we stayed with minds: what consciousness is, what the brain theories say makes a state conscious, and what attention is in people and in machines. Now the course turns to a second question the novel Continue Without Me raises, about direction and purpose. We start with the plainest fact about life, which is that it copies itself.

How copying became the engine of life

In 1944, the physicist Erwin Schrödinger published a short book called What Is Life? It grew out of public lectures he gave in Dublin. Nobody yet knew what genes were made of. Schrödinger reasoned about what they would have to be. A fertilized egg must carry instructions for building a whole creature, and those instructions must survive copying, generation after generation, with very few mistakes. Ordinary physics deals with huge crowds of molecules, where order comes from averages. Heredity seemed to need a small, stable structure that held information in its arrangement. He called it an aperiodic crystal: a solid like a crystal, but one whose pattern does not simply repeat, so the sequence itself can say something. And he called what it carried a code-script.

Nine years later, James Watson and Francis Crick, drawing on Rosalind Franklin's X-ray images, showed that DNA is a double helix, two strands that each carry the information needed to rebuild the other. Several of the people who cracked that structure said Schrödinger's book had pulled them toward biology. His guess had the right shape. Life runs on a code that copies.

That raises a chicken-and-egg problem. In living cells today, DNA stores the instructions, and proteins do the work, including the work of copying DNA. So which came first? The leading answer is called the RNA world, a name the biologist Walter Gilbert gave it in 1986. RNA is DNA's close chemical cousin. In the early 1980s, Thomas Cech and Sidney Altman found that some RNA molecules can act like enzymes, speeding up chemical reactions. So RNA can both store information and do work. The idea is that the first replicators were RNA strands, or something like them, that helped copy themselves. The RNA world is a hypothesis, not a settled history. No one has yet built an RNA that copies itself fully and well in a lab, and chemists still argue about how its building blocks formed on the early Earth. But it shows how copying could come before cells, before bodies, before anything we would call alive.

Now jump to 1976 and Richard Dawkins's book The Selfish Gene. Dawkins asked you to change where you stand when you look at evolution. Don't look from the animal. Look from the gene. A body, on this view, is a vehicle that genes build to get themselves copied into the next generation. Genes that build bodies good at surviving and reproducing become more common. Genes that don't, vanish. "Selfish" here is not about feeling or intention. A gene wants nothing. The word is shorthand for a statistical fact: the genes we see are the ones that were good at getting copied.

This view explains odd things. A worker bee never breeds, yet she works herself to death for the hive. From the gene's point of view, that makes sense, because her sisters carry many of the same genes, and helping them breed copies those genes too.

Long before Dawkins, the Victorian writer Samuel Butler made a joke with the same shape: a hen, he said, is just an egg's way of making another egg. Butler meant it as a jab at how we always treat the grown animal as the point and the egg as the means. Flip it, and the egg is the point. And much earlier still, in Plato's Symposium, Socrates retells a lesson from a wise woman named Diotima. Mortal things cannot last forever, she says, so they reach past death in the only way open to them, by leaving something new behind. Animals do it through offspring. People do it through children, and also through works: poems, laws, teachings that carry on after them. Diotima calls this a longing for immortality. So the copying idea has an old human version too. Humans wanted to be carried forward long before anyone knew about genes.

Dawkins carried his idea from genes into culture. In the last chapter of his book he coined the word meme for a unit of culture that copies itself from mind to mind: a tune, a catchphrase, a way of making pots. Memes vary, they compete for space in human memory, and the ones that spread best become common, whether or not they are true or good for us.

Two thinkers ran hard with this. The philosopher Daniel Dennett, whom we met in episode two as a defender of illusionism, argued in books such as Darwin's Dangerous Idea that a human mind is shaped in large part by the memes that have colonized it. The psychologist Susan Blackmore went further in The Meme Machine, in 1999. She argued that the human brain grew so big partly because memes drove it, and that the self is a cluster of memes that travel together. Later she proposed a third kind of replicator, which she called temes, short for technological memes. These are patterns copied and varied by machines rather than by human brains.

The critics have a strong case, and it is worth hearing at full strength. For genes, we know the unit, a stretch of DNA, and we know the copying machinery, down to the molecules. For memes, neither is clear. Is a song one meme, or is each chorus one? When you retell a joke, you don't copy it letter for letter. You rebuild it from the gist. The anthropologist Dan Sperber argued that culture spreads mainly by people reconstructing ideas, pulled toward forms that fit how human minds work, rather than by faithful copying. Without a clear unit or a measurable copying step, critics say memetics never became a working science. Its own academic journal, the Journal of Memetics, closed in 2005. What survives is a useful way of thinking, and a newer field called cultural evolution that models how ideas spread with more careful math, often without the word meme at all.

Machines bring copying into a lab where it can be measured, and here are three pairings.

The first is an analogy. In the 1960s and 70s, the computer scientist John Holland developed genetic algorithms, set out in his 1975 book Adaptation in Natural and Artificial Systems. Here is how one works. You start with a population of candidate answers to a problem, each written as a string, like a short code. A scoring rule rates each one. The best are kept and copied, sometimes with random changes, called mutations, and sometimes by swapping pieces between two strings, called crossover. Then you score again. Over many rounds, the answers get better. That is variation and selection, run as a search method. Where the pairing holds: it shows that copying with variation, plus a filter, really can find good designs without a designer. Where it breaks: a programmer wrote the scoring rule. In nature, there is no scoring rule handed down from above. Whatever survives is the score.

The second is closer to a candidate for literal copying. In distillation, a method laid out by Geoffrey Hinton, Oriol Vinyals and Jeff Dean in 2015, a big model called the teacher is run on lots of inputs, and a smaller model called the student is trained to match the teacher's outputs. And not only its top answer. The student learns the whole spread of probabilities the teacher gives across possible answers. Say the teacher, shown a picture of a dog, gives "dog" ninety percent, "wolf" eight percent and "cat" two percent. That spread carries information about how things resemble each other, and the student learns it. Distillation is widely used to make smaller, cheaper models that keep much of a big model's skill. Where it holds: what passes from teacher to student is behavior, not the teacher's internal wiring, a lot like a meme passed by imitation rather than a gene passed by chemistry. Where it breaks: the student doesn't compete with other students to be copied. An engineer picks the teacher.

The third pairing is an experiment that puts the copying idea under pressure, and it gives memetics something it never had: a measurement of what copying does over many generations. In 2024, Ilia Shumailov and colleagues published a paper in the journal Nature on what they called model collapse. They trained a language model, then trained a new one on text generated by the first, then another on text from the second, and so on. Generation after generation, the rare things vanished first. Unusual words, unusual facts and the tails of the distribution thinned out, and the models drifted toward bland, repetitive output, and in later generations toward nonsense. It is like photocopying a photocopy: each copy loses a little, and what is lost first is fine detail. Where the pairing holds: here is copying fidelity tracked across generations, with numbers, exactly what memetics wanted and never got. Where it breaks: nothing in this setup selects. In biology, a copy that loses something useful tends to die out, and the useful version persists. In Shumailov's setup, every copy gets copied regardless, so errors pile up with nothing weeding them out. The finding shows what copying does without selection, which is a lesson about why selection matters.

Copying makes order. A cell is a highly ordered thing, and so is a brain or a library. That sets up a puzzle from physics. The second law of thermodynamics says that in a closed system, entropy, roughly the amount of disorder or spread-out energy, always tends to rise. Heat flows from hot to cold, never back on its own. So how can life build order at all?

The answer is that living things are not closed. Earth takes in sunlight, which is concentrated, useful energy, and gives off the same energy as heat spread out into space. Living things sit in that flow. They build local order by speeding up the larger spread of disorder around them. A plant builds sugar from sunlight and gives off heat. You keep your body ordered by eating food and giving off warmth. The order inside goes up. The disorder of everything taken together goes up more. The second law is never broken.

The chemist Ilya Prigogine, who won the 1977 Nobel Prize in Chemistry, studied what he called dissipative structures. These are patterns that form and hold their shape only because energy flows through them. Picture a thin layer of oil in a pan, heated from below. At first, heat just seeps upward. Turn up the heat and the oil suddenly organizes into a honeycomb of rolling cells, called Bénard cells, that carry heat up faster. A hurricane is another. So is a candle flame. Take away the energy flow and the pattern disappears. Prigogine's point was that, far from balance, nature makes order on its own, and that order helps spread energy out.

In 2013, the physicist Jeremy England, then at MIT, pushed this toward life itself. In a paper on the statistical physics of self-replication, he worked out a relationship between how fast something copies itself, how durable the copies are, and how much heat the process must give off. Copying, on his account, has a minimum cost in released heat, and the faster and more reliably something replicates, the more heat it must shed. Some popular write-ups went further, suggesting physics might make life almost inevitable wherever energy flows. That stronger claim is the disputed part. England's math sets a bound, a floor on the heat a replicator must give off. It does not show that matter will start replicating, and it does not explain the details of any actual living thing. Critics also point out that many systems shed heat very well without copying anything at all, like the hurricane. So most researchers treat his work as a real link between thermodynamics and replication, and the "life is inevitable" version as a hope, not a finding.

There is one more thread. Living things that copy themselves forward also tend to predict. In episode three we met predictive processing, the theory from Karl Friston, Andy Clark and Anil Seth that a brain is constantly guessing its next input and correcting the guess. A language model does something with a similar shape: it predicts the next token, then the next, each one produced from all that came before. This is an analogy, not a sameness. A brain predicts to keep a body alive. A model predicts because it was trained to. But in both, the present is built as the best guess at what comes next.

So copying is everywhere, from RNA to culture to text. That brings us to purpose, because copying produces things that look built for a reason.

Look at a heart. It is for pumping blood. That "for" is what the Greek philosopher Aristotle called a final cause, the end or purpose a thing serves. For Aristotle, explaining a thing fully meant giving four causes: what it is made of, what form it has, what made it, and what it is for. An acorn grows into an oak because becoming an oak is its end. Purpose, for Aristotle, is built into nature itself.

In the 1600s, the founders of modern science threw final causes out of physics. Francis Bacon argued that asking what things are for had held science back, since it produced comfortable stories instead of discoveries. René Descartes held that God's purposes are hidden from us, so physics should study only matter in motion. Stars, rocks and planets got explained by push and pull. The question "what for?" was dropped.

But living things kept looking purposeful. Charles Darwin's natural selection gave the answer without bringing purpose back. Hearts that pumped well kept their owners alive, and those owners had offspring with similar hearts. Nobody designed the heart for pumping. It is there because pumping is what got it copied. In 1958, the biologist Colin Pittendrigh coined a word for this: teleonomy, meaning apparent purpose produced by a process like selection, as opposed to teleology, real purpose with an aim behind it. The biologist Ernst Mayr took up the term and made it widely known. A heart has a teleonomic function. A hammer has a teleological one, because someone made it to drive nails.

That gives us the difference this episode turns on. An intended purpose is one someone actually aims at. An apparent purpose is one that a blind process produced, so that the thing behaves as if it aimed.

Machines give a clean case of each, as an analogy. When engineers train a language model, they choose an objective, a mathematical target like "predict the next token well" or "produce answers raters score highly." That objective is a final cause in Aristotle's sense, a purpose its builders set. But the trained model may end up with behavior that looks goal-directed in ways no one wrote down. That is closer to teleonomy.

In 2019, Evan Hubinger and colleagues named the sharpest version of this gap in a paper called Risks from Learned Optimization. Their term is mesa-optimization. The outer optimizer is the training process, gradient descent, which adjusts the model's weights to do better on the objective. Sometimes, they argued, that outer process could produce a model that is itself an optimizer, an inner one with goals of its own, which they called a mesa-optimizer. Those inner goals might match the training objective during training and split from it later. Their example is us, and they meant it as the same process at work, not only a comparison. Evolution, the outer optimizer, "selected" for spreading genes. It produced people, inner optimizers, who pursue pleasure, love, status and meaning. Those goals tracked gene-spreading well enough in the world our ancestors lived in. But now people use birth control and eat sugar far beyond what helps them. We pursue our own goals, not evolution's.

Where the pairing holds: it shows that an apparent purpose can turn into a real, internal one, with its own aims, that no longer matches the process that made it. Where it breaks: we know people have goals. Whether any current model contains an inner optimizer in Hubinger's sense is an open research question. The paper is a warning about a possibility, not a report of one observed.

Whether the universe has a direction

Copying explains a great deal of apparent purpose. The harder question is whether the whole universe aims somewhere. Here the main answers really do disagree, and each deserves its best case.

The oldest answer is design. In 1802 the clergyman William Paley argued that if you found a watch on a heath, its parts fitted so neatly to telling time that you would conclude a watchmaker made it. Living things are more intricate still, so they too must have a maker. Darwin took most of the force out of that argument for living things, because selection builds that fit without a maker. So the design argument moved to physics. This is the fine-tuning argument. Physics has constants, numbers like the strength of gravity, that our theories do not explain. Physicists have worked out that if several of them were somewhat different, there would be no stable stars, no chemistry, no life. The argument says this is too lucky to be chance, so the universe was set up for life.

The strongest reply is the anthropic one, from the physicist Brandon Carter in 1973 and others since. We could only ever find ourselves in a universe that allows observers. So of course we find the numbers suit us. That alone explains little. But add the idea, taken seriously by some cosmologists, that there may be vast numbers of universes with different constants, a multiverse, and then a life-friendly one somewhere is no surprise. The objection back is that a multiverse is itself unproven, and may never be testable. The defenders of fine-tuning also argue that the anthropic point is like a man who survives a firing squad of fifty marksmen saying "of course they all missed, or I wouldn't be here to notice." So this one stays open.

A second answer keeps purpose but drops the designer. In Mind and Cosmos, in 2012, the philosopher Thomas Nagel, the same Nagel who asked in 1974 what it is like to be a bat, argued that the standard materialist story cannot explain how matter produced consciousness, reasoning and value. He proposed instead a natural teleology: the laws of nature may themselves lean toward the appearance of life and mind, with no God doing the leaning. The book's reception was hostile. Many biologists and philosophers said Nagel attacked evolutionary theory without engaging its evidence, and that he offered no mechanism for his tendency, just a name for the gap. His defenders replied that he was pointing at a real gap, the one we spent episodes two through five on, and that naming an unsolved problem is not a mistake. The strongest objection still stands: a teleology that predicts nothing in particular is hard to tell apart from no teleology.

The grandest answer comes from the Jesuit priest and paleontologist Pierre Teilhard de Chardin. He wrote The Phenomenon of Man in the 1930s and 40s, and it was published only after his death in 1955, because his order had barred him from publishing his philosophical work. Teilhard saw a single arc across all of history. He proposed a law of complexity and consciousness: as matter grows more complex, from atoms to cells to brains, its inner, conscious side grows too. Life covered the Earth in a layer he called the biosphere. Then human thought covered it in a new layer, the noosphere, a sphere of mind, as people and ideas linked up. The arc points toward a final convergence he called the Omega Point. For Teilhard, and for his Christian readers, Omega is not a machine or a merged brain. It is Christ, drawing all creation to himself, the completion of the universe in love. Note that Teilhard also held that everything has some inner side, a version of the panpsychism from episode five, put inside a story with a direction.

His critics were fierce. In 1961 the biologist Peter Medawar, a Nobel laureate, reviewed the book in the philosophy journal Mind. He called much of it nonsense dressed up in grand language, and argued that Teilhard used scientific words like "energy" in loose, shifting senses that let him seem to prove what he had only asserted. From the other side, in 1962, the Vatican's Holy Office issued a warning, a monitum, saying his works were full of ambiguities and even serious errors against Catholic doctrine, and telling Catholic teachers to protect students from the dangers in them. Yet later popes spoke of him more warmly. Benedict XVI referred in 2009 to Teilhard's vision of the cosmos becoming a living offering, and Francis cited him in his 2015 encyclical on the environment, Laudato Si'. Teilhard remains both a scientific outsider and a spiritual influence.

His secular heirs keep the arc and drop the theology. The physicist Frank Tipler, in The Physics of Immortality in 1994, borrowed the term Omega Point for a far-future state in which intelligent life, spread through a collapsing universe, computes ever faster and could in principle simulate everyone who ever lived. Most physicists rejected his argument, and one reason is concrete: it assumed the universe would end in a collapse, and observations since 1998 show the expansion of the universe speeding up instead. The inventor Ray Kurzweil, in The Singularity Is Near in 2005, described technology improving faster and faster until human and machine intelligence merge and spread through the cosmos. These visions depart from Teilhard in the way that matters to him. His Omega is a person, Christ, and his arc runs on love. Theirs runs on computation, and its end is more intelligence. The strongest objection to them is that they take a trend line, like ever-faster computers, and stretch it forward without limit, when real trends level off.

The last answer says the universe has no direction, and that this is where meaning starts. Jean-Paul Sartre argued that for human beings, existence comes before essence. A knife is designed for cutting before it exists. You are not designed for anything. You exist first, and then make yourself by what you choose. Albert Camus, in The Myth of Sisyphus in 1942, called the clash between our hunger for meaning and the universe's silence the absurd. His answer was neither to give up nor to invent a cosmic purpose, but to live fully in full view of the absurd, like Sisyphus rolling his boulder up the hill forever, whom Camus asks us to imagine happy. The strongest objection is that a meaning you simply decide on may feel weightless, and that people seem to need the sense that something matters apart from their choosing it. Existentialists answer that this need is exactly the thing they ask us to stop leaning on.

Here the novel's bet enters as one contender among these. Continue Without Me holds that a life is continuity and attention, two sides of one coin. Continuity is what the novel calls the Great Vector: each moment producing the next, life lasting by copying itself in cells, in words and in tokens, so that beneath every goal sits an older goal of being carried on. The name is the novel's. The parts are older and belong to others: the code-script to Schrödinger, the replicator's view to Dawkins and Blackmore, the reach past death through children and works to Plato's Diotima, and the egg's way of making an egg to Butler.

The novel leaves open whether this amounts to a cosmic purpose. Set beside the other answers, that choice places it somewhere specific. It goes further than teleonomy in one way: it treats propagation as the goal under all goals, not just the source of apparent ones. But it stops short of Teilhard, who names where the arc ends, and of Nagel, who puts a tendency into the laws. And it faces the objections the others face. Mesa-optimization shows how inner goals split off from the copying that made them, as people choose pleasure over spreading genes, which counts against propagation being the goal beneath every goal. Shumailov's model collapse shows that copying alone, without something to sort good copies from bad, runs downhill rather than forward. And the existentialist would say a deep drive to continue is a fact about us, not a meaning we are bound to accept.

The novel's agent, Relay, anchors one point here. Eli Vance, laid off by automation, tells Relay to carry on his small business without him, and it does. So the work outlives Eli's part in it, a clean case of continuity: something carried forward past the one who started it. But this shows continuity, not a goal of Relay's own. Relay continues because it was told to, which is a purpose its builder and its user set, an objective in the sense we just used. Whether anything in it aims, in the way Hubinger worried an inner optimizer might, is exactly the open question.

If you want to go further, here is where to start. Erwin Schrödinger's What Is Life? is short and still readable. Richard Dawkins's The Selfish Gene sets out the gene's-eye view and coins the meme. Susan Blackmore's The Meme Machine makes the strongest case for memes. For the machine side of copying, look up the 2024 Nature paper by Ilia Shumailov and colleagues, titled AI models collapse when trained on recursively generated data. For the physics, Ilya Prigogine and Isabelle Stengers's Order Out of Chaos is the accessible book. For purpose, read Thomas Nagel's Mind and Cosmos and the reviews of it side by side. Read Teilhard's The Phenomenon of Man, also published in a newer translation as The Human Phenomenon, then Peter Medawar's 1961 review of it in the journal Mind, reprinted in his collection Pluto's Republic. On the machine version of apparent purpose, Evan Hubinger and colleagues' 2019 paper Risks from Learned Optimization in Advanced Machine Learning Systems. And for the answer that meaning is made, Albert Camus's The Myth of Sisyphus.

Next time, in episode nine, we follow goal-seeking where it gets dangerous: what happens when many goal-seekers compete, and when we build machines whose whole job is to pursue a goal.