Self-replicating probes continue to be a controversial subject, just as they were when Frank Tipler up the ante on Michael Hart by invoking them as a way of further tightening the tension of the Fermi Paradox. After all, Tipler had discovered an economic edge. Any civilization that wants to colonize the galaxy is going to expend vast resources, but if self-replication is available, that culture need only create the first probe, and let subsequent ones harvest resources as needed. Self-replication or not, the galaxy gets filled up in only a fraction of the current age of the Milky Way, but the economic stimulus provides yet another tightening of the Fermi knot. I hadn’t thought about all of this in connection with actual probe designs, but Peter Marinko’s article on the matter clearly touched a nerve, judging from the messages I’ve been getting about it. When Peter wrote recently with his thoughts on reader reactions, I asked him for permission to run it as a regular post rather than a comment, because I think this is a lively question and would like to see us continue to explore it.
by Peter Marinko
My previous post here, “A Metallurgist’s Doubts About Self-Replicating Probes,” argued that von Neumann probes are constrained less by physics than by process-chain closure and materials aging. The discussion that followed sharpened my thinking more than the original post did, and I want to begin by paying some of that debt.
On carbon: Adam Crowl supplied the numbers I should have had at hand: carbonaceous chondrites run some 3–5% organics, dormant comets are thought to be coated in hydrocarbon-rich asphalt, and Freitas’s study considered Titan — an organics-drenched world — alongside the Moon. Alex Tolley added the wider inventory: aromatics, CO, CO₂, CH₄, hydrogen-rich giant atmospheres, and tholins — the reddish-brown organic polymers, first named by Sagan, that form when ultraviolet light and charged particles work on simple molecules like methane and nitrogen, and that coat Titan, Pluto, and many cometary surfaces.
I accept the correction. Carbon is not scarce in the cosmos. But availability is not accessibility: 3–5% organics dispersed through a chondrite is a feedstock concentration problem, and concentration is exactly the step that has no gravity, no water, and no atmosphere to help it. What the correction really did was promote carbon from an afterthought to a criterion — as the reader will see below, three independent process chains now demand it.
On refractories: Alex Tolley and John both pointed to the same escape: induction heating with magnetic levitation, melting metal without touching a crucible at all. This is a genuinely good answer, and I concede it for the class of operations it covers — melting and casting nickel-iron, which asteroids supply free of charge. But it is not free, and the price is paid in the currency this study cares most about. A levitated melt has its entire surface exposed, and it radiates as T⁴. A crucible is not merely a container; it is insulation. Remove it and the induction coil must continuously replace radiative losses that a lined furnace would simply have prevented. In an exergy ledger, that is a permanent tax on every kilogram melted.
Nor does levitation cover reduction. Extracting metal from oxide requires a hot, chemically aggressive, contained environment, and containment is where linings live. Here I owe the discussion a nuance from my own field that I should have raised myself: melting is not the only route. Iron oxide can be reduced in the solid state, producing iron powder which is then pressed and sintered to finished shape — this is how Höganäs in Sweden has made metal powder for decades, and it is how tungsten-carbide drill inserts are produced. Powder metallurgy skips the melt entirely, which is a real advantage for a probe: no crucible, no tapping, no casting. But it does not escape the problem. Solid-state reduction needs long tunnel furnaces, a reducing atmosphere of hydrogen, carbon, or carbon monoxide, and sustained high temperature — and those furnaces need linings too. The refractory bootstrap survives every route I know how to draw.
Image: PG: I sometimes wonder what John von Neumann would say if he could see the length and depth of the debate over self-replicating interstellar probes. To my knowledge, he never considered self-replication in the context of star systems and certainly not colonizing an entire galaxy. We could use his insights today: What are we missing? Credit: Physics Today, although this old photo is widely available and I don’t know its origin.
Abelard Lindsey, writing from inside the industry, put the honest bound on it: difficult, but perhaps a decade or two of development. That is a fair estimate, and it belongs in the ledger as a research task rather than a wall.
On semiconductors: Lindsey and I are in violent agreement: this is the hard one, and he notes it must be solved on Earth as well before AI becomes cheap. “Building chips rather than carving them,” as he puts it, is the holy grail. I have no better idea, and I am not going to pretend otherwise.
On the mother-factory: Tolley’s most interesting move was to propose a way around closure rather than through it: a factory ship that travels star to star, spawning non-replicating probes from local resources, populating the galaxy more slowly but without ever needing full self-reproduction. I want to be clear that this is a good idea — and also that it concedes my central point. A mother-factory does not close the loop; it carries the fraction it cannot produce, and expands until that stock is exhausted. It converts an infinite-generation architecture into a finite one, and the number of generations it buys is precisely the vitamin inventory divided by the vitamins per copy. That is not a refutation of the closure argument. It is the closure argument, written as a mission design. The right question becomes: how many nodes does the stock buy? That number is computable, and Hephaistos is built to compute it.
On life: The deepest challenge came from Henry Cordova and Elisee Reclus, and it deserves more than a nod, because it attacks my framing rather than my arithmetic: robust self-replicating systems demonstrably exist, cover the Earth, and arose without design. If biology can do it, why not machines?
Here is my answer, and it is the reason I am not moving the goalposts. Life is not a counterexample to the closure problem; life is what closure looks like when you pay its actual price. A cell does not manufacture bearings to micron tolerance, does not need vacuum, does not require phase-pure silicon or reference metrology, and above all does not need to specify its output. It tolerates enormous error, discards most of its offspring, and lets selection curate the survivors — over billions of years, in a medium (liquid water, at moderate temperature, with an atmosphere and a gravity well) that supplies concentration and transport for free. Biology bought replication by abandoning precision, determinism, and speed, and by spending geological time as its currency.
A von Neumann probe cannot make that trade. It must arrive at a specified place, build a specified artifact to specified tolerances, and do so in decades. The moment we relax those requirements enough for a biological strategy to work — accept vast error, accept mostly-failed offspring, accept deep time — we no longer have an engineering project; we have seeded a biosphere and lost the ability to say what it will become. That may be a defensible thing to do. It is not the thing anyone is proposing when they invoke probes crossing the galaxy in a few hundred thousand years.
So I take the biological objection seriously, and my conclusion from it is not optimism but a sharper statement of the problem: self-replication is cheap if you can pay in error and time, and murderously expensive if you must pay in precision and schedule. The exergy ledger below is an attempt to price the second option honestly.
Where this goes next:
Skepticism is cheap. The honest next step, for a critic who spent a career in industrial process engineering, is to try to make the thing work — on paper, with real process chains and mass balances — and see exactly where it breaks.
That is what I am now attempting, in a study I am calling Project Hephaistos, after the god who forged automata for Olympus: a virtual self-replicating probe, audited line by line, where every assumption and every capitulation is logged in public. Two ledgers run through it. The Vitamin List records every component the probe cannot make for itself, with masses attached. The Exception Ledger records every problem I have deliberately set aside. I expect the second document to be the more valuable of the two.
In a future post I will set out the mission architecture, and the first trade study: whether it is better to send a probe slowly, with today’s technology, and let it fight fifty millennia of aging — or quickly, at a tenth of light speed, and let it fight the interstellar medium. Both, it turns out, are running the same race against the same opponent. Only the costume changes.
My thanks again to everyone who wrote in. Keep it coming — the ledger has room.
Hi Peter, now that I know what you are after, I’d like to make a few suggestions and comments. Designing a self replicating probe that replicates itself fully in one step is very hard. The probe would have to contain as many advanced tools as in an MIT Fab Lab, and even more.
But if you traded off time, the probe could build an advanced factory by building a few crude tools, and use those tools to build more more advanced tools. That’s what humans did – start with stones and bones, then build with bronze, then steel, then make machines that could make more accurate machines that could build small things like watch gears and big things like 10 ton presses.
As far as processing raw materials, if the probe is lucky, it will find an nickel-iron-cobalt asteroid. But even if it doesn’t it should be able to find an asteroid or comet with water. And once you have water, you can electrolyze it for hydrogen and oxygen.
And once you have hydrogen, you can ionize it and even make high energy ions like in an ion propulsion engine, then use the hydrogen ions to reduce most of a rock into constituent elements. For instance, you may not be able to use aluminum to build a machine with the same quality as using steel, but it can be useful enough to go to the next step.
So I believe things can be built up slowly, step by step. If a factory can grow itself at 1% per day (a biological plant grows by 5% to 30% dry mass per day), it would take only 70 days to double in size (mass). First, build a factory floor. Then build a hammer. Then a lathe, etc.
As far as making insulated wire, it may be easier to start with something like a printed circuit board. In place of plastic insulation, or epoxy, start with metal wires embedded in sintered ceramic powder until you have the capability to build valves for a chemical processing plant.
Anyway, have fun!
@ Randy Chung. “first build a factory floor”. Apparently without even a hammer since the hammer is mentioned as the next item? No, this argument is not convincing at all.
Hi Randy! Thank you for yr comment!
This is the strongest form of the optimistic case, and I want to engage it seriously rather than wave it away — it is essentially the argument Alex Ellery has made in the literature, and it deserves respect.
You are right that humanity climbed exactly this ladder: stone, bronze, iron, steel, machines that build better machines. The question Hephaistos is built to answer is not whether that ladder exists but how many rungs it has and what each rung costs — in mass carried, in energy, and above all in time. Two cautions from the practitioner’s side. First, the ladder is not free of imports: every rung on the human path relied on the previous civilization’s accumulated knowledge, infrastructure, and — crucially — a whole planet’s pre-concentrated ores, water, and air. The probe has none of that hinterland. Second, growth rates borrowed from biology (1% a day, doubling in 70 days) assume the hardest thing is already solved: a system that can already make every part of itself from ambient feedstock. A plant does not have to smelt its own chloroplasts from ore. The doubling time is not the problem; the closure fraction is. If the factory can copy 99% of itself and must import the last 1%, it does not double every 70 days — it stops when the imported stock runs out.
So I am not arguing the staged bootstrap is impossible. I am arguing it has a bill, and that nobody has yet written the bill down. That is precisely what I am trying to do — rung by rung, with the vitamins that each rung cannot make listed honestly
A good friend of mine, Andrew Hall Cutler, did a lot of materials processing research while at University of Arizona in the early 90’s. He also did SBIR work before then. He and I met in L5-Society. He commented that there was very little real work being done in the actual process technology necessary to effect cost-effective space settlement. The good metallurgist who authored the previous posting brought back a lot of memories from that time. I remember Andy showing me an English translation of a Russian book detailing materials process work done on Mir and the previous smaller stations that the Soviets launched and operated in the 1980’s. He said it was the only real work that had been done at the time. No one, and I mean no one, was doing any of this work in the 1990’s (other than Andy, of course).
Years later, I worked with a guy (Hamada-san) who was developing spray deposition technology for flat panel displays. He spent some years trying to develop a replacement material for ITO for the transparent upper electrode. He was not successful. At the time (around 2010) no one, and I mean no one, was trying to do the same. This would not only get around the resource limitations of Indium (the “I” in ITO) but would also lead to a web process (roll to roll) production of solar cells without the need for vacuum process. This took us in contact with guys doing printed electronics start-ups. Space will not happen until it becomes cost-effective (meaning self-financing).
Find and read Freeman Dyson’s “Pilgrims, Saints, and Spacemen”.
The point is that very few people are doing the real work necessary to get us into space in a large way. Forget about self-replicating probes. I’m talking about the process technology to realize planetary surface colonies (Mars, Calisto, etc.) as well as the O’Neill cylinders. There is no pathway to space without doing this work.
Hi Aberlard!
Thank you — and yes. The point that so little real process-technology work has been done is the one I most want this series to make land. I will look up Cutler’s work and the Dyson essay you mention. Your ITO example is exactly on point: a single materials substitution problem, unsolved for decades, with an entire industry riding on it. Multiply that by every specialized input in the closure list and the scale of the omission becomes clear.
As I understand it, the Space Studies Institute (SSI) has been funding and coordinating this kind of practical research for decades as well as hosting conferences on space related manufacturing. I’m skeptical of claims that “no one” was or is doing the “real work”.
‘A levitated melt has its entire surface exposed, and it radiates as T⁴. A crucible is not merely a container; it is insulation. Remove it and the induction coil must continuously replace radiative losses that a lined furnace would simply have prevented. In an exergy ledger, that is a permanent tax on every kilogram melted.’
In weightlessness if you melt a mixture of iron and regolith the iron should still go to the centre, iron likes iron, so you will get insulation from the oxides on the surface. Also you can get hydrogen on asteroids from solar wind impingement for reduction processes and asteroids are far from waterless.
Hi Michael and Alex!
Thank you both — this is exactly the kind of pushback that earns its keep, and it forces me to be precise about why the crucible is hard to escape.
Induction melting of iron-bearing regolith might serve for small quantities. But two problems scale badly. First, it is doubtful the regolith contains much metallic iron; far more likely it is present as oxides, which means reduction must happen before there is anything to levitate — and reduction is the step that wants containment. Second, suppose against expectation that native iron is present in low concentration. Then the question becomes how solidification under low gravity is arranged, because the iron and the molten silicate must never mix during freezing.
This is not a fine point; it is the center of the craft. On Earth, metallurgists work relentlessly to keep oxide inclusions out of cast steel, because non-metallic inclusions — even at micrometer scale — wreck mechanical properties. Think of a bearing under constant load: a single micron-scale inclusion is a crack initiation site, and the bearing fails. Alex’s reflective-sphere suggestion is ingenious, and containerless electromagnetic processing is real (we do it terrestrially for small samples). But in the scenario proposed, the risk is precisely that the steel and the silicate do not separate in weak gravity — they freeze together, and the steel is worthless for any use. Add that steel shrinks about 7% on solidification, which must be managed in every downstream forming step, and it becomes hard to see this scaling to the quantities and qualities a self-replicating factory needs.
There is also the matter of grade, and here I want to concede a point before making one. If the regolith is iron-poor — as it probably is — then winning a little metal means heating the entire bulk to melting temperature, overburden and all, because you cannot selectively heat the few percent you actually want. The energy per kilogram of usable metal is therefore large, most of it spent on material you will immediately discard. In a terrestrial economy that would be decisive. In space it may not be: sunlight is abundant, and as Abelard Lindsey noted earlier in this thread, energy is probably not the scarce resource. So I will not rest the argument on the power bill. I rest it on what the heat does not fix — the oxide reduction that must precede melting, the inclusions that ruin the product, the separation that low gravity does not guarantee. Cheap energy heats the rock; it does not make the steel clean.
And here is the recursive point I keep returning to: the induction coil itself needs insulated electrical conductors. So do the many pumps, valves, and machines around it. Containerless melting does not avoid the insulation problem — it presupposes it is already solved.
To Michael (xenon as thermal barrier)
Interesting thought, but I don’t think an inert gas rescues the situation. A gas layer conducts and convects heat rather than blocking it, and in any case it does nothing about the two real problems above: the oxide-reduction step that precedes melting, and the inclusion/separation problem during freezing. It also introduces a consumable (the xenon) that must be contained and replenished — another line in the ledger, not a saving.
Gravity can be emulated by rotation – spin the factory, cetrifugal force will yield the artificial gtavity you need.
Energy is abundant near the star, but further from it, in some distant asteroid belt it may be worse. If you harvest far from star, there is much less energy. In solar system, asteroids are somewhat abundant beyound mars orbit, with much lesss availavle solar power.
It may be better strategy to use some Mercury-equivalent if available in target solar system rather than scavenging crumbles from asteroids.
‘But in the scenario proposed, the risk is precisely that the steel and the silicate do not separate in weak gravity — they freeze together, and the steel is worthless for any use. Add that steel shrinks about 7% on solidification, which must be managed in every downstream forming step, and it becomes hard to see this scaling to the quantities and qualities a self-replicating factory needs’
When a materials like iron and silicon freeze it actually pushes out impurities. Even sulphur which is an annoying impurity has its uses, it improves its machineability.
A nice article about useful product manufacture in space.
https://home.ifa.hawaii.edu/users/meech/a281/handouts/permanent_chap4.pdf
In micro-g space, there are only radiative losses. What stops one from suspending the metal ball in a perfectly reflective sphere? The radiative flux may be high, but we can contain plasmas in fusion reactors at far higher temperatures. The suspension may need nothing more than inertia to keep the metal in one place, with the reflective shell sufficiently separated to prevent any errant molten metal from reaching the sphere’s surface. Would this not solve the need for a refractory container to hold the molten metal? (Handling the molten sphere to use it is another issue.)
You could use xenon gas, it is more thermally resistant than refractory materials by a significant margin. Not sure how it would behave in zero g though.
Oh darn. Same old, same old Fermi Paradox: We have not detected any machines here that weren’t made by us…
“Life is not a counterexample to the closure problem; life is what closure looks like when you pay its actual price.”
This raises the fascinating question of how you could tell the difference between a planet with a naturally evolved biosphere (like Earth seems to be) vs a planet deliberately seeded with engineered life forms.
Hi Ivan! Thank you for yr comment!
Your question — how to tell an evolved biosphere from a seeded one — sent me to a number I can’t stop turning over. In 2024 a Harvard–Google team (the H01 study, Science) mapped a single cubic millimetre of human cortex, about half a grain of rice. That speck holds roughly 57,000 cells and 150 million synapses, and merely imaging it took 1.4 petabytes. A whole brain is on the order of a million such specks; extrapolated, the raw map runs to something like a zettabyte.
I want to be careful with that figure: the 1.4 PB is electron-microscope imagery, an inefficient encoding, not the compressed information content of the wiring, which is surely far smaller. But even discounted by orders of magnitude, it says something a metallurgist feels in the gut — the specification of a biological structure, the blueprint of what goes where, is astronomically deep.
This is where your question bites. A seeded biosphere would not escape that cost; it would inherit it, pre-paid, in the DNA of whatever was sown. And that is precisely the trick a machine cannot borrow. Biology stores its blueprint in a few gigabytes of DNA and lets development — and a billion years of selection, and, in our case, many thousands of cooperating brains over the long climb of human culture — unfold the rest for free. A fertilised cell becomes a brain not because it carries the petabytes, but because environment and deep time supply them. The machine has no such unfolding: it must carry or reconstruct its full specification explicitly. That is a third axis of the closure problem, alongside materials and energy, and it is the one the standard arithmetic ignores completely.
I’ll stop short of saying such machines can never rival this — I’ve spent these posts insisting the constraint is thermodynamic and economic rather than forbidden, and I won’t quietly change my terms now. But the honest reading of H01 is that the information cost of closure, for anything approaching biological sophistication, is so large it deserves its own ledger. The probe, mercifully, need not grow a brain from ore; it carries its computer as a vitamin. What H01 shows is what that vitamin is really worth.
I would think a larger controlling robot would be better sent to each target star which then controls the many minions. I just can’t see small devices having much brain power to make large scale decisions unless they work in a hive mind fashion.
Fossils.
A fossil record would document the evolution of the local biota. If they had been introduced artificially at any time during the planet’s history there would be discontinuities and abrupt interruptions in the fossil record. No matter how carefully the bio-engineers tried to cover their tracks, this test would settle it once and for all.
Of course, if the local wildlife was such that it left no fossils, or if the planet’s geology did not preserve fossils, this test might not work. It also reminds us that local conditions do not necessarily duplicate conditions here on earth. Every planet will be different, perhaps very different. Our analyses here often tend to forget that.
Yeah finding a planet where the fossil record is only a few millions years old would be a slam dunk. Or finding a mega discontinuity in the fossil record.
Another sign I would expect to distinguish between a terraforming project and a replication site for von Neuman probes, would be to find organisms whose purpose would seem to be jump starting industrial processes. Such as organisms that mimic hydro geological processes or organisms that “ferment” industrially useful materials from the environment. Imagine organisms that concentrate useful metals in stromatalite or coral structures.
Another suspicious sign could be to find organisms with DNA equivalents that have say six bases rather than our 4. This seems to have a number of advantages, but may be harder to kick start naturally (i.e our 4 base system is easier to self organise).
https://foreveron.com/podcast/episode-031/
Along this line of thinking, wouldn’t it be more efficient if an ETI sent out a number of large fully automated “mother” ships in various directions, with all the materials necessary to then fabricate (if necessary) or simply auto-provision smaller ships containing robotic/synthetic AI agents, materials (raw and pre-fabbed) with all the necessary equipment and industrial tools to establish an outpost constructed by the synthetics agents. This process would happen as the mother ships identified planets of interest as they proceeded along their paths of stellar visitations. Would this not be a more economically sensible approach if not a simpler engineering approach that is more likely to succeed as opposed to devising probes that are responsible for creating a fabrication industry from scratch needed to build an outpost from raw materials?
What is a self replicating probe — is there a strict definition or is it more open ended? Is it something very small & nuclear or could it be something large, highly stratified while being automated … literally an industrial park in a “probe”?
After all, we know that one type of replicators exist: the life itself. They can be as simple as bacterium and as advanced as vertebrates, but there is some sense in which they are near-100% closed loops. And cellular machinery is pretty impressive “nanotech”. So, there is nothing fundamental which forbids replicators. With this in mind, it is much more difficult to imagine why other types of replicators, including mainly inorganic and living in space and micro-g, cannot exist. But there is really long way from asteroid mining and space fabs to such things, and we are only considering the first step now. Space and zero-g is so much different that it’s difficult to conceive truly space-native industrial processes, we don’t have much in our previous experience about it. But surely it has it’s own advantages. We think about asteroid mining in terms of liquid- or air-assisted gravitational separation. Maybe a full-fledged interplanetary civilization just roasts asteroids whole with solar concentrators, distilling them progressively into volatiles and refractories and harvesting vapors and condensed dust until the last drop of platinoids evaporates. But likely they do it entirely differently.
@torque_xtr
If the asteroid could be heated so that the local material became a plasma, the ionized material could be electrically contained and sorted (via some mass spec method) to separate the elements into their respective “buckets”.
As you say, we don’t know the details of how an advanced technological civilization would manage the resources of asteroids, but there should be conceptual ideas on how it could be done, in principle.
I wish your Project Hephaistos idea success — despite and because of the odds against it. All the world is a sword, and the very tip of that sword is the act of restricting, banning, surveillancing, and monopolizing chip technology. Everything about the ‘fabs’ is “export controlled”, which for ordinary people means reserved to their betters. If someone came up with a way that ordinary people could make their own general-use computers, beholden to no one, that might at any moment be used to design nuclear weapons, receive anonymous leaks of the Leader’s tax returns, or create videos of celebrities without their clothes, their governments would experience a terror to rival the Pure Light of the Void. And of course the self-replicating probes would become a doomsday weapon if ever really made, needing only some time to bring their kinetic and chemical presents back home. Yet humans themselves are a doomsday weapon: any population so changed that it is honestly not a potential threat to the world would in fact not be sentient at all.
While we don’t have self-replicating, intelligent machines (yet), we have the next best thing – cheap robots (drones) with self-improving intelligence, churned out by factories and home 3-D printing. A decade or so ago, I went to an AI meeting where a naval researcher was demonstrating the “power” of AI in a battlefield scenario. Lots of “collateral damage” and “kills” of their own soldiers. Currently, the situation has changed, and we are nearing Black Mirror scenarios for these [semi] autonomous machines. Significantly, the US Dept. of War doesn’t want any restrictions on the capabilities of AI and presumably embodied AI.
We seem to be headed in P. K. Dick’s Second Variety world (filmed as Screamers) where the machines can both replicate and evolve. We will get your scenario;
without creating interstellar traveling self-replicating probes, because their numbers will be manufactured en masse on Earth within the foreseeable future. We can see drone capabilities rapidly evolving in a real war situation in Eastern Europe, with teh supposedly leading technology nation looking to buy the technology and apply greater resources to improve its capabilities. It should also be of note that China is showing the capability of producing very “intelligent” AI that is smaller and cheaper than US “Frontier Models” and that corps. are looking to deploy it on “Edge devices” like smartphones and tiny hardware like the Raspberry Pi. Small, killer robots with facial recognition are coming “real soon now.”