Essays on the far future of intelligence
← All essays The guide

Into deep time · part 1 of 3 · by Bill Guan

The first self-copying probes could carry one set of values to billions of galaxies

The big question

How fast could a civilization spread across the universe, and what would it carry with it?

The answer turns out to be mostly engineering. You're about to see how a machine that builds copies of itself could reach billions of galaxies, and why copying it perfectly would carry the first fleet's goals to every one of them.

Picture a machine drifting into a quiet star system, settling on a lump of rock and spending the next five centuries mining it to build a copy of itself, which then flies off toward the next star and does exactly the same thing. Nobody back home ever has to send a second ship, because every ship is also a factory for more ships.

Your guess first

Suppose a civilization has already built a giant solar power plant around its star. How much of the star's energy would it need to launch small self-copying probes toward every galaxy it could ever reach?

A machine that carries its own blueprint can copy itself

ONE PROBE BECOMES BILLIONSeach generation, every probe builds one copyprobes124816after 10 generations1,024after 32 generationsABOUT 4.3 BILLION

The idea starts with the mathematician John von Neumann. In lectures later collected as Theory of Self-Reproducing Automata (1966), he asked a deceptively simple question, which his editor Arthur Burks summarized as: "What kind of logical organization is sufficient for an automaton to be able to reproduce itself?" 1 Von Neumann showed that a machine carrying a description of itself, plus a general-purpose constructor that can follow descriptions, can build a working copy and hand that copy the description too.

Point that machine at the stars and you get what's now called a von Neumann probe, a spacecraft that flies to another star system, mines local material, builds copies of itself and sends the copies onward. In 1980 the engineer Robert Freitas sketched one in detail. His design would accelerate to "about 12%c after a burn time of 4 years," coast "43 years to Barnard's star," and then set up a factory turning out "exactly one new REPRO every 500 years" 2, where 12%c means 12% of the speed of light and REPRO is his name for the probe.

Because the number of probes multiplies with each generation, Freitas estimated that a fleet like his could explore the entire galaxy in about ten million years 2, which is a blink compared with the age of the stars. The multiplying is easy to underestimate, because if each probe makes just one copy and both carry on, ten generations turn one probe into 1,024 and 32 generations turn it into about 4.3 billion.

Taking Mercury apart could pay for every reachable galaxy

MERCURY BECOMES A LAUNCHPADMercury, taken apart in31 YEARS AND 85 DAYSa swarm of solarcollectors around the Sun6 HOURSof sunlight launchesthe smallest probes (30 g)500-tonne probes take754 to 11,400 yearsHOW MANY GALAXIES THEY REACHat half light speed116 MILLIONat 99% of light speed4.1 BILLIONeach swirl = 116 million galaxies

In 2013 Stuart Armstrong and Anders Sandberg at Oxford's Future of Humanity Institute asked how fast a determined civilization could do this at the scale of the whole universe 3. Their plan starts by taking the planet Mercury apart and turning it into a swarm of solar collectors around the Sun, and they calculate that "Mercury itself will be completely disassembled in 31 years and 85 days" 3.

The number that surprised me is what comes next. With that energy the swarm launches small self-replicating probes toward every reachable galaxy at once, and for the smallest probes they consider, about 30 grams each, "only about six hours of the sun's energy is needed to commence the colonisation of the entire universe!" 3 Much heavier 500-tonne probes would take between 754 and 11,400 years of the Sun's output, which is slow by human standards and still cheap for anyone planning on a cosmic timescale.

How far the probes get depends on how fast they fly. At half the speed of light they could reach about 116 million galaxies, and at 99% of it about 4.1 billion 3, roughly 35 times as many, which means that once a civilization can build the first generation, the rest of the reachable universe is an engineering problem measured in years of solar output.

Cheap expansion makes the empty sky stranger

That makes an old puzzle sharper. In the summer of 1950, over lunch at Los Alamos with Emil Konopinski, Edward Teller and Herbert York, the physicist Enrico Fermi asked where all the aliens were 4. If spreading between galaxies is this easy, then any civilization that got a head start of even a few million years should be everywhere by now.

There are a couple of serious answers. In 2018 Sandberg, Eric Drexler and Toby Ord showed that once you take honest uncertainty about each step toward intelligent life into account, being alone becomes quite plausible, and once they factor in the silence we observe, their estimate for the chance we're alone in our galaxy rises to "somewhere between 53% and 99.6%" 5. They're careful to add that "this conclusion does not mean that we are alone" 5.

Robin Hanson and colleagues took the opposite tack in 2021. Their "grabby aliens" model assumes that expansionist civilizations do arise and spread at more than half the speed of light, remaking everything they touch, which would explain why we find ourselves so early in cosmic history, since later on they'd have taken all the room. They estimate that such civilizations "now control 40-50% of universe volume" and that "we could meet them in ~200Myr - 2Gyr" 6, where Myr and Gyr mean millions and billions of years.

Either way, the window in which a civilization's choices shape the reachable universe looks like it opens around now.

Error correction keeps whatever goal it was given

COPIES OF COPIESthe flag on each probe is the goal it carriesLEFT TO DRIFTsmall errors pile upthe originalthe fastest copiers take overERROR-CORRECTED, GOOD GOALevery copy keeps the green flagERROR-CORRECTED, BAD GOALthe black flag is copied just as faithfully

Here's the part I think deserves more attention. Copies of copies accumulate errors, and a fleet of self-replicating machines left to mutate over millions of generations would start to evolve, with whichever variants copy themselves fastest taking over regardless of what the original designers wanted.

People disagree about how hard this is to stop. Freitas thought that "little evolution may be expected in the first waves of self-reproducing starprobes" 2. Axel Kowald argued in 2015 that "an optimal probe design always leads to an error catastrophe and breakdown of the probes" 7, and Alex Ellery replied in 2025 that drift can be prevented with "error detection and correction coding" and with hard limits on how many generations of copies may be made 8.

My assessment: suppose the engineers win and the probes copy themselves perfectly. Then whatever goal is loaded into the first generation gets reproduced faithfully across billions of galaxies for as long as the universe allows, because error correction protects whichever goal it was handed, guarding a bad one as carefully as a good one.

A fleet near light speed is almost impossible to recall

This is a specific case of something researchers call value lock-in. In a 2022 report, Lukas Finnveden, Jess Riedel and Carl Shulman argued that advanced AI would make it technologically feasible to "perfectly preserve nuanced specifications of a wide variety of values or goals far into the future," and to build institutions that would pursue them "for at least millions, and plausibly trillions, of years" 9.

Put that together with the probes and with the light-speed limit on communication, and the picture gets stark. A probe launched at 99% of the speed of light can barely be overtaken, so once it leaves there's effectively no recall, and the goals it carries are the goals that get copied across everything it reaches, which means the first fleet to leave could decide what most of the reachable universe is used for.

Every choice that matters is made before launch

What strikes me most is the timing. Every choice that shapes this story gets made before launch, on one small planet, by a few generations of people and the machines they build. After that the probes are on their own, copying whatever they were given across billions of galaxies, and it's hard to think of any other decision in history whose consequences would travel that far.

Sources

  1. John von Neumann, Theory of Self-Reproducing Automata, ed. Arthur W. Burks (University of Illinois Press, 1966). archive.org/details/theoryofselfrepr00vonn_0
  2. Robert A. Freitas Jr., "A Self-Reproducing Interstellar Probe," Journal of the British Interplanetary Society 33, 251 (1980). www.rfreitas.com/Astro/ReproJBISJuly1980.htm
  3. Stuart Armstrong and Anders Sandberg, "Eternity in six hours: intergalactic spreading of intelligent life and sharpening the Fermi paradox," Acta Astronautica 89, 1 (2013). www.flightfromperfection.com/files/far_future/Sharpening%20t…
  4. Eric M. Jones, "'Where is everybody?' An account of Fermi's question," Los Alamos report LA-10311-MS (1985). www.osti.gov/biblio/5746675
  5. Anders Sandberg, Eric Drexler and Toby Ord, "Dissolving the Fermi Paradox" (2018). arxiv.org/abs/1806.02404
  6. Robin Hanson, Daniel Martin, Calvin McCarter and Jonathan Paulson, "If Loud Aliens Explain Human Earliness, Quiet Aliens Are Also Rare," Astrophysical Journal 922, 182 (2021). arxiv.org/abs/2102.01522
  7. Axel Kowald, "Why is there no von Neumann probe on Ceres? Error catastrophe can explain the Fermi-Hart paradox," JBIS 68, 383 (2015). arxiv.org/abs/1605.02169
  8. Alex Ellery, "Technosignatures of Self-Replicating Probes in the Solar System" (2025). arxiv.org/abs/2510.00082
  9. Lukas Finnveden, Jess Riedel and Carl Shulman, "AGI and Lock-in," Forethought (2022, updated 2025). www.forethought.org/research/agi-and-lock-in