It’s time for a self-reproducing machine
I’ve wanted to build a self-reproducing machine since I was 17. It’s forty-five years later, and it has finally become feasible. (I’ve done a few other things along the way.) I’m going to describe one such device, and speculate as to its larger implications. It’s a pretty detailed design, which I had to come up with to convince myself that it is feasible. No doubt there are better designs than this.
The Autofac
Here’s a top-level description of the device I’m thinking of. It’s called an Autofac, which is what they were called in the earliest story about them. It looks like a little metal shed, about a meter cubed. It weighs about 50 kilograms. There’s a little gnome-sized door on each end. It’s full of robot arms and automated machine tools. It’s connected to electricity and by WiFi to a data center somewhere. It has a front door, where it accepts material, and a back door, where it outputs useful objects, and cans of neatly packaged waste. You can communicate with it, to tell it to make parts and assemble them into useful shapes. It can do all the metalworking operations available to a machinist with a good shop at their disposal. In return, it occasionally asks for help or clarification.
One particular thing it can be told to make is another one of itself. This is of course the case we’re all interested in. Here’s what that looks like. You feed a 60kg package of steel castings, electronics, and other parts, into the door at one end. It starts by building another shed, next to the other end. The two sheds are butted up next to each other, so the rain can’t get in. Once it’s enclosed, there is no visible progress for about a month, but it makes various metalworking noises. Then it announces that it’s done. The second shed is now another Autofac, and can be carried away to start the process elsewhere. There’s also a can full of metal scrap and used lubricant, which has to be disposed of responsibly. This process can be repeated a number of times, at least seven, to produce more offspring. Eventually the original Autofac wears out, but by then it has hundreds of descendants.
The software
The key part of the Autofac, the part that kept it from being built before, is the AI that runs it. Present-day VLMs (vision-language models) are capable of performing short-deadline manual tasks like folding laundry or simple tool use. But they are deficient at arithmetic, long term planning and precisely controlling operations. Fortunately we already have software for these three purposes.
First, of course, we have calculators for doing arithmetic. LLMs can be taught to use these. In the real world, machinists constantly use calculators. The Autofac will be no different.
Second, there is project planning software that lets a human break down an engineering project into tasks and subtasks, and accommodate changes of plan as things go wrong. We can provide the data structures of this software, initially constructed by humans, as a resource for the AI to use. The AI only has to choose the next task, accomplish it or fail, and either remove it from the queue or add a new task to fix the problem. There are thousands of tasks in the life of an Autofac; fortunately the AI doesn’t need to remember them all. The project planning software keeps track of what has been done and what needs to be done.
Third, there are programs that go from the design of a part to a sequence of machine tool movements that will make that part, and then controls the machine tool motors to do the job. These are called Computer Aided Manufacturing, or CAM. Using CAM relieves the AI of the lowest level responsibilities of controlling motor positions and monitoring position sensors. This software doesn’t do everything, of course, which is why being a machinist is still a skilled job. The AI needs to clamp parts in place for machining operations, notice when something goes wrong, change plans flexibly, and many other intelligent activities.
When the AI runs into a problem it doesn’t know how to handle, it can ask a human for help. This is where the language part of the vision-language model comes into play. The questions and answers from any one Autofac will need to be fed back to all Autofacs, so they don’t also ask the same question. As the number of Autofacs grows exponentially, the amount of human assistance to any one Autofac will have to drop exponentially. This requires updating the VLM on the fly; I’m not sure how that will work.
The VLM will require training data to specialize it into being an Autofac. There are thousands of books about metalworking, providing cheap training data. Even books more than a century old can be useful; metalworking hasn’t changed that much. There are also thousands of hours of video of machinists building things, while explaining what they’re doing and what they’re thinking. This is very useful, provided the VLM is smart enough to generalize from humans to a robot arm. Finally, we could teleoperate the Autofac, to carry out all the needed task. At the same time, the operator could narrate their thoughts. This would provide the highest-quality training data, but at a high price.
Because most of the software runs in a data center remote from the Autofac, it can be shared between Autofacs. This allows it to not have to think very hard when things are boring, and very hard when something goes wrong. Since a lot of machining is just watching the machines until they finish, this lets us economize on computer power during the boring bits.
Economics
The smallest technological system capable of physical self-reproduction is the entire economy. Obviously it’s not possible to duplicate the entire economy. But because we can take inputs from the economy, we can design a machine capable of converting these inputs into a copy of itself. The goal is to minimize and simplify inputs to the extent possible. If an Autofac is going to be economically reasonable, it has to capable of providing services that exceed the cost of its inputs. There have been self-reproducing machines since the 1950s, that require complex manufactured parts, and provide no particularly useful service. This is not that. The Autofac described above is a profitable device all by itself. There is a small but lively industry of custom parts manufacturing—send them a CAD design, and they’ll send you the parts by return mail. This is exactly what the Autofac is designed to do, when it’s not building another one of itself.
How long is the self-reproduction time?
It’s been observed many times that a machine shop is capable of manufacturing any part of itself, given steel, a machinist and enough time. It has just become feasible to embed the intelligence of a machinist into electronics. So now we can say “a machine shop is capable of manufacturing itself from steel and a copy of its electronics, given enough time.”
We can make an estimate of how much machinist time is required for a machine shop to duplicate itself. We can do that as follows. We know the weight and cost of all the machines. We can find the resource cost by assuming the machines are made out of mostly steel, and assume that all the remaining cost was the labor used in manufacture. Notice that this is an overestimate of the labor cost, because it ignores transportation, marketing, profit and all other costs. Also, the resource cost is relatively small—steel is cheap compared to the things we make from it. We can assume that all this labor is provided by our robot arms, at the same speed as a human. And from this, we can work out how long it takes for our machine to produce another one of itself.
The estimate comes out to about a year for a machinist to build another machine shop. I confirmed this estimate by asking a professional machinist, so that’s weak evidence that it’s not wildly off.
The cubic meter of an Autofac contains, roughly:
A robot, 30 cm tall, with two arms and a mobile base
A 3-axis CNC mill
A 2-axis CNC lathe
A tool grinder
A ball bearing making machine (really just an attachment for the grinder)
A motor winder attachment for the lathe
A thread rolling tool for the lathe
A heat treating oven
A manually operated rotary axis (for gears)
A wire drawing machine
A sheet rolling machine
An arbor press
A bunch of precision “hand” tools.
They’re all about 1⁄5 the scale of ordinary machine tools. Machine tools scale quite nicely. As long as motion velocity remains constant, making the machine smaller by a factor X makes it faster by X, but strain, stiffness and accuracy are the same. In this case, the factor is X=5. (Except that electric motors get bigger in proportion as you scale everything else down. But it’s not bad. Also the oven doesn’t scale right either. But that’s small for compared to the size of the whole machine shop, so even if it gets relatively bigger, that’s not too bad.)
The Autofac doesn’t have to sleep or rest, so that’s a factor of four speedup. But it’s clumsier than a human, so lets double the time. So the estimated self-reproduction time is (1/5)*(1/4)*2 of a year, or five weeks.
This may be an underestimate, if there are marvelous machines used in the assembly of machine tools, that I’ve forgotten to include, but which provide great productivity with little labor. I think I know enough about the production of machine tools that I’ve included all such machines: that’s why I include specialized machines for wire drawing, motor winding, sheet metal rolling, screw manufacture, and grinding ball bearings.
The original Autofac is only capable of building small parts, but it can also build a somewhat larger Autofac, which in turn can build one even larger. Autofacs can scale up to build whatever you want. It turns out that because the time is proportional to the scale, it’s optimal to do as much self-reproduction as you can at the smallest scale, only scaling up at the end, when you need to.
Next Steps
So at first the Autofac is a Minimum Viable Product. This is nifty, but the market for small custom parts is limited. What’s the next step in scaling? It turns out that much of the cost is in electricity (3 kW, continuously). To make our own electricity, we have the Autofac build a windmill generator. You could build one windmill per Autofac, but the power available from a windmill scales as the fifth power of the height, so it probably makes sense for a group of Autofacs to build one giant windmill to serve them all. (I haven’t worked out what tools and materials are needed to do this.)
Most of the input is steel. Scaling from the size of the US economy, an Autofac requires 92% steel by weight, 6% lubricant and cutting fluid, 2% everything else: copper wire, refractory clay, grinding wheels, abrasive powder, insulation, tungsten welding tips, rustproof paint, all the electronics. The steel can come in as rods, plates, and pieces cast into various shapes. The Autofac itself can roll the plates into sheet, and draw the rods into wires. So the next thing we need the Autofacs to build is a steel mill, to turn rock, air and water into cast steel. There’s a minimum size for a steel mill, which is pretty big, set by the rate of heat loss from a mass of molten steel. It doesn’t make sense to have a heat of steel smaller than a cubic meter (haven’t actually done any calculations here, just guessing.) That’s enough steel to build 100 Autofacs, and if an Autofac needs a reload once a month, and a steel mill can make five heats a day, one steel mill is enough to keep up with 15000 Autofacs.
Finding iron ore isn’t a problem: iron is the fourth most common element in the earth’s crust, and you have to make some effort to find rock that doesn’t contain iron. (As a child in Southern California, I used to amuse myself by dragging a magnet through the dirt and pulling it out fuzzy with iron ore.) If you heat up iron-rich rock to white heat, and then blow hydrogen through it, the iron gets reduced before most other elements, and forms a metallic sponge. After you’ve spongified most of the iron, you separate it, melt it in a crucible, add an appropriate amount of carbon, and pour it into molds. Presto, steel! The hydrogen you can get by electrolyzing water, and the carbon by extracting carbon dioxide from the air, and reducing it with hydrogen. I haven’t worked out how much electricity this all takes; I hope not too much.
Given sources for steel and electricity, all we need to supply is the Autofac “starter pack”, which weighs about a kilogram and contains all the non-steel parts described above. We need to drive down the cost for this as much as possible. We need to ride the learning curve for manufacturing the starter pack down, down, down.
The Baffin Island Plan
You probably aren’t familiar with Baffin Island, so let me describe it. It’s in the Canadian Arctic, between Hudson Bay and Greenland. It’s bigger than California, but is home to only 15,000 people. It produces iron ore, from one iron mine, some native handicrafts, and no other major exports. It’s mostly caribou-infested tundra, but there are substantial areas of barren rock. It is usually below freezing, constantly windy, and much of it is dark six months of the year. I love it; it’s perfect.
There are portions of Baffin Island of great natural beauty or local economic importance. We will avoid those portions. That still leaves 400,000 km^2, enough for over a billion Autofacs. (The limit is not actual land area; the limit is the windmill spacing for wind power. The actual Autofacs take up a relatively small amount of the area, leaving room for the caribou.)
Here’s the story. We put one Autofac on one of the less habitable parts of Baffin Island. At first, we have to provide electricity, steel, and Autofac starter packs. After one year, we have a thousand Autofacs, and they’re starting to build windmills and a steel mill. Now we only have to provide starter packs. Another two years and a billion starter packs later, we have a billion Autofacs. Now we can really get to work. About a third of the Earth’s industrial capacity is ours to do with as we will. We can
Fix carbon out of the atmosphere. If it can be turned into lubricant for the starter pack, so much the better.
Pump ocean water onto land, and let it freeze. Keep the resulting salty glaciers insulated when the weather is too warm; on cold nights roll back the insulation and freeze them harder. This lowers the ocean level and makes beach dwellers happy.
Manufacture anything that can be built out of steel, plus ships to take it to civilization.
Turn control of the Autofacs over to AI. I’m sure it will be fine. Nothing can possibly go wrong.
I think you’re substantially underestimating the difficulty here, and the proportion of effort which goes into the “starter pack” (aka vitamins) relative to steelworking.
If you’re interested in taking this further, I’d suggest:
getting involved in the RepRap project
initially focusing on just one of producing parts, assembling parts, or controlling machinery. If your system works when teleoperated, or can assemble but not produce a copy of itself, etc., that’s already a substantial breakthrough.
reading up on NASA’s studies on self-replicating machinery, e.g. Freitas 1981 or this paper from 1982 or later work like Chirikjian 2004.
While the “autofac” might not be feasible at this time, it’s interesting to consider how to minimize the required size of a civilization that’s productive enough to survive on eg Mars or an asteroid.
To be clear, I think the autofac concept—with external “vitamins” for electronics etc—is in fact technically feasible right now and if teleoperated has been for decades. It’s not economically competitive, but that’s a totally different target.
I thought I was pretty careful not to say how hard I estimated the difficulty to be, but just to be clear: I think it will be a large project and many years of effort. Can you point to a place where you got the opposite impression? Or was it my breezy style?
I too was surprised by how large the steel input was compared to the vitamins, and in turn how much of that was lubricant relative to everything else. I got these proportions by scaling off the US economy as a whole. Compared to how much steel we consume, the vitamins are really that small a fraction. It also seems like a reasonable fraction of non-steel parts in a bunch of machine tools in a metal box. Of course those vitamins are far more valuable per kilogram compared to the steel.
Thank you for your suggestions of how to demonstrate only part of the system; I’ve been trying to come up with a minimum viable product that is less difficult to get to. I’m an old guy, not in the full flower of my health, so I’m gonna let someone else build the startup company to do the whole job.
Thanks for the reference to Chirikjian 2004; I wasn’t aware of that one.
I’m just guessing from affect, yep, though I still think that “large project and many years of effort” typically describes considerably smaller challenges than my expectation for producing a complete autofac.
On the steel-vs-vitamins question, I’m thinking about “effort”—loose proxy measurements would be the sale value or production headcount rather than kilograms of output. Precisely because steel is easier to transform, it’s much less valuable to do so and thus I expect the billions-of-autofacs to be far less economically valuable than a quick estimate might show. Unless of course they start edging in on vitamin production, but then that’s the hard rest-of-the-industrial-economy problem...
Isn’t the US a bit too dependent on “Vitamin Xi” and “Vitamin T[aiwan]” (I amuse myself) for the electronics for these naive estimates based on the U.S. economy to work out too well? We’d have to find a way to carve out a part of the global economy that is basically self sufficient and very plausibly reproducible i.e. able to replace all critical aspects of itself once they run out their lifetimes without a substantial degradation of technology, and I can only imagine a claim of a self-reproducing subsection of (or just the entire) global economy to be very controversial due to predictions of a collapse or massive change to globalization, breakdown in U.S.-China trade relations, an invasion of Taiwan, worries of economic, societal and technological collapse, etc.
This was a very interesting post. A few scattered thoughts, as I try to take a step back and take a big-picture economic view of this idea:
What is an autofac? It is a vastly simplified economy, in the hopes that enough simplification will unlock various big gains (like gains from “automation”). Let’s interpolate between the existing global economy, and Feynman’s proposed 1-meter cube. It’s not true that “the smallest technological system capable of physical self-reproduction is the entire economy.”, since I can imagine many potential simplifications of the economy. Imagine a human economy with everything the same, but no pianos, piano manufacturers, piano instructors, etc… the world would be a little sadder without pianos, but eliminating everything piano-related would slightly simplify the economy and probably boost overall productivity. The dream of the Autofac involves many more such simplifications, of several types:
Eliminate luxuries (like pianos) and unnecessary complexity (do we really need 1000 types of car, instead of say, 5? The existence of so many different car manufacturers and car models is an artifact of capitalist competition and consumer preferences, not a physical necessity. Similarly, do we really need more than 5 different colors of paint / types of food / etc...).
Give up on internal production of certain highly complex products, like microchips, in order to further simplify the economy. Keep giving up on more and more categories of complex products until your remaining internal economy is simple enough that you can automate the entire thing. Hopefully, this remaining automated economy will still account for most of the mass/energy being manipulated, with only a small amount of imports (lubricants, electronics, etc) required.
Why make such a fuss about disentangling a “fully automatable” simplified subset of the economy from a distant “home base” that exports microchips and lubricant? I don’t think a self-sufficient autofac plan would ever make sense in the middle of, say, the city of Shenzhen in China, when you are already surrounded by an incredibly complex manufacturing ecosystem that can easily provide whatever inputs you need. I can think of two reasons why you might want to cleave an economy in half like this, rather than just cluster everything together in a normal Shenzhen-style agglomeration mishmash:
If you want to industrialize a distant, undeveloped location, and the cost of shipping goods there is very high, then it makes sense to focus on producing the heaviest materials locally and importing the smallest / most complex / most value-per-kg stuff.
If you can cut humans entirely out of the loop of the simplified half of the economy, then you don’t have to import or produce any of the stuff humans need (food, housing, healthcare, etc), which is a big efficiency win. This looks especially attractive if you want to industrialize a harsh, uninhabitable location (like Baffin Island, Antarctica, the Sahara desert, the bottom of the ocean, the moon, Mars, etc), where the costs of supporting humans are higher than normal.
Take an efficency hit in order to eliminate efficiency-boosting complexity. Perhaps instead of myriad types of alloy, we could get by with just a handful. Perhaps instead of myriad types of fastener, we could just use four standard sizes of screw. Perhaps instead of lots of specialized machines, we could make many of our tools “by hand” using generalized machine-shop tools.
But wait—I thought we were trying to maximize economic growth? Why give up things like carbide cutting tools in favor of inferior hardened-steel? Well, the hope is that if we simplify the economy enough, it will be possible to “automate” this simplified economy, and the benefits of this automation will make up for the efficiency losses.
Okay then, why does efficiency-imparing simplification help with automation? Couldn’t our autofac machine shop just as easily produce 10,000 types of fasteners, as 4 standard screws? Especially since the autofacs are making so many things “by hand?” Feynman seems very interested in an Autofac economy based almost entirely around steel—what’s the benefit of ditching useful materials like plastic, concrete, carbide, glass, rubber, etc? I see a few potential benefits to efficiency-imparing simplifications:
lt reduces the size/cost/complexity of the initial self-replicating system. (I think this motivation is misplaced, and we should be shooting for a much larger initial size than 1 meter cubed.)
It reduces the engineering effort needed to design the initial self-replicating system. (This motivation is reasonable, but it interacts in interesting ways with AI.)
By trying to minimize the use of inputs like rubber and plastic, we reduce our reliance on rare natural resources like rubber trees and oil wells, neither of which exist on Baffin Island, or the moon, or etc. (This motivation is reasonable, but it only applies to a few of the proposed simplifications.)
To me, it seems that the Autofac dream comes from a particular context—mid-20th-century visions of space exploration—that have unduly influenced Feynman’s current concept.
Why the emphasis on creating a very small, 1 meter cubed package size?? This is a great size for something that we are shipping to the moon on a Saturn V rocket, or landing on Mars via skycrane, or perhaps sending to a distant star system as a Von Neumann probe. But for colonizing Baffin Island or the Sahara Desert or anywhere else on earth, we can use giant containter ships to easily move a much larger amount of stuff. By increasing the minimum size of our self-replicating system, we can include lots of efficiency-boosting nice-to-haves (like different types of alloys, carbide cutting tools, lubricant factories, etc). Feynman imagines initially releasing 1000 one-meter-cubed autofacs (and then supporting them with a continual stream of inputs), but I think we should instead design a single, 1000x-size autofac (it doesn’t have to be one giant structure—rather a network of factories, resource-gathering drones, steel mills, power plants, etc), since that would allow for more efficiency-boosting complexity.
The remaining argument for 1000 one-meter-cubed autofacs is that it would be easier to design this much-smaller, much-simpler product. This is true! I’ll get back to this in a bit.
In general, I suspect that the ideal size of the autofac system should be proportional to the amount of transportation throughput you can support to Baffin Island / Mars / wherever. Design effort aside, it would be ideal to design the largest and most complex possible autofac which would fit into your transportation budget (eg, if you can afford five container ships to Baffin Island per year, then your autofac system should be large enough to fit into five container ships.
Cutting humans entirely out of the loop is very appealing for deep-space exploration, but less appealing for places like Baffin Island. As long as you are only relying on relatively unskilled labor (such that you aren’t worried about running out of humans to import, during the final stages of the industrialization of the island when millions and millions of windmills / steel mills / etc are going up), then importing a bunch of humans to handle a small percentage of high-value, hard-to-automate tasks, is probably worth it (even though it means you now have to provide housing, food, entertainment, law enforcement, etc).
As others have mentioned, this “compromise” vision seems similar to Tesla’s dreams of robotic factory workers (in large, container-ship-sized factories that still employ some human workers) and Spacex’s mars colonization plans (where you still have a few humans assembling a mostly-mechanical system of nuclear power plants and solar panels, habitable spaces, greenhouses for food, etc—but no 1-meter cubes to be seen, since Starship can carry 100 tons at a time to Mars).
But again, I admit that re-complexifying the economy by introducing humans, does greatly increase the design complexity and thus the design effort required at the beginning.
The one-meter-cubed autofac seems so pleasingly universal, like maybe once we’ve designed it, we could deploy it in all kinds of situations! But I think it is a lot less universal than it looks.
A Baffin-Island-Plan autofac wouldn’t fare well in the Sahara desert, where you’d want to manufacture solar panels (which rely more on chemistry and unique materials) instead of simple mechanical windmills that could be built almost entirely from steel. In the sahara, you’d also have less access to iron ore in the form of exposed rock; by contrast you’d have a lot of sillica that you could use to make glass. On the moon, you’d have no atmosphere at all for wind, and extreme temperatures + vacuum conditions would probably break a lot of the machine-shop tools (eg, liquid lubricants would freeze or sublimate). Etc.
The above point isn’t a fatal problem—just having one autofac system for deserts and another for tundra would cover plenty of use cases for industrializing the unused portions of the earth. But you’d also run into problems when you finished replicating and wanted to use all those Baffin Island autofacs to contribute back to the external, human economy. Probably it would be fine to just have the Baffin Island autofacs build wind turbines and export steel + electricity, while the desert autofacs build solar panels and export glass + electricity. But if you decided that you wanted your Baffin Island autofacs to start growing food, or manufacturing textiles, you would have a big problem. The autofacs would in some ways be more flexible than a human manufacturing economy (eg, because they are doing more things “by hand”, thus could switch to producing other types of steel products very quickly), but in other ways they would be much more rigid than a human manufacturing economy (if you want anything not based on steel, it might be pretty difficult for all the autofacs to reconfigure themselves).
Design effort & AI—if AI is good enough to replace machinists, won’t it be good enough to help design an autofac?
This post reminds me of Carl Shulman’s calculations (eg, on his recent 80,000 Hours podcast appearance) about the world economy’s doubling times, and how fast they could possibly get, based on analogies to biological systems.
Feynman says that, after many years, nowadays the dream of the Autofac is finally coming within reach, because AI is now good enough to operate robotics, navigate the world, use tools, and essentially replace the human machinist in a machine shop. This seems pretty likely come true, maybe in a few years.
But creating such a small, self-contained, simplified autofac seems like it is motivated by the desire to minimize the up-front design effort needed. If AI ever gets good enough to become a drop-in remote worker not just for machinsts, but also for engineers/designers, then design effort is no longer such a big obstacle, and many of the conclusions flip.
Consider how a paperclip-maximising superintelligence would colonize baffin island:
The first image that jumps to mind is one of automated factories tesselated across the terrain. I think this is correct insofar as there would be lots of repetition (the basic idea of industrial production is that you can get economies of scale, cheaply churning out many copies of the same product, when you optimize a factory for producing that product). But I don’t think these would be self-replicating factories.
A superintelligent AI could do lots of design work very quickly, and wouldn’t mind handling an extremely complex economy. I would expect the overall complexity of the global economy to go way up, and the minimum size of a self-replicating system to stay very large (ie, nearly the size of the entire planetary economy), and we just end up shipping lots of stuff to Baffin Island.
If we say that the superintelligence has to start “from scratch” on Baffin Island alone, with only a limited budget for imports, then I’d expect it to start with something that looks like the 1-meter-cubed autofac, but then continually scale up in the size and complexity of its creations over time, rather than creating thousands of identical copies of one optimized design.
A superintelligence-run economy might actually feature much less repetition than a human industrial economy, since the AI can juggle constant design changes and iteration and customization for local conditions, rather than needing to standardize things (as humans do to allow interoperability and reduce the costs of communicating between different humans).
Okay, I will try to sum up these scattered thoughts...
I think that the ideal autofac design for a given situation will vary a lot based on factors like:
what resources are locally available (wind vs solar, etc)
how expensive it is to support humans as part of the design, vs going fully automated
how much it costs to ship things to the location (the more you can ship, the bigger and more complex your autofac should be, other things equal)
the ultimate scale you’re aspiring to industrialize, relative to the size of your initial shipments (if you want to generate maximum energy on 1 acre of land using a 100 tons of payload, you should probably just import a 100-ton nuclear reactor and call it a day, rather than waste a bunch of money trying to design a factory to build a factory to build a nuclear reactor. Wheras if you are trying to generate power over the entire surface of Mars with a 100 ton payload, it is much more important to first create a self-replicating industrial base before you eventually turn towards creating lots of power plants.
how much it costs to design a given autofac system
a larger, more complex autofac will cost more to design, but will be more efficient
a more-completely-self-sufficient system (eg, including lubricant factories, or eliminating the need for humans on Baffin Island) will cost more to design, but will save on shipping costs later
if you can use lots of already-existing designs, that will lower design costs (but it will increase complexity elsewhere, since now you have to manufacture all the 10,000 types of fasteners and alloys and etc used by today’s random equipment designs)
advanced AI might be able to help greatly reduce design costs
A fully-automated “autofac” design wins over a more traditional human-led industrialization effort, where the upfront costs of designing the mostly-self-sufficient autofac system manage to pay for themsleves by lowering the recurring costs of importing stuff, paying employees, etc, of a human-led industrialization effort.
Whoever decides to start the cool autofac startup, should probably spend a bunch of time considering these big-picture economic tradeoffs, trying to figure out what environment (baffin island, the sahara desert, the oceans, the moon, Mars, alpha centauri, etc) offers the most upside from an autofac-style approach (Baffin-Island-like tundra might indeed be the best and most practical spot), and what tradeoffs to make in terms of autofac size/complexity, how much and where to incorporate humans into the system, and etc.
I would personally love to get a better sense of where the efficiencies are really coming from, that help an autofac strategy win vs a human-led industrialization strategy. Contrast the autofac plan with a more traditional effort to have workers build roads and a few factories and erect windmills all over Baffin Island to export electricity—where are the autofac wins coming from? The autofac would seem to have some big disadvantages, like that its windmill blades will be made of heavy steel instead of efficient fiberglass. Are the gains mostly from the fact that we’re not paying as many worker salaries? Or is it mostly from the fact that we’re producing all our heavy materials on-site rather than having to ship them in? Or somewhere else?
Wow, I think that comment is as long as my original essay. Lots of good points. Let me take them one by one.
The real motivation for the efficiency-impairing simplifications is none of size, cost or complexity. It is to reduce replication time. We need an Autofac efficient enough that what it produces is higher value than what it consumes. We don’t want to reproduce Soviet industry, much of which processed expensive resources into lousy products worth less than the inputs. Having achieved this minimum, however, the goal is to allow the shortest possible time of replication. This allows for the most rapid production of the millions of tons of machinery needed to produce massive effects.
Consider that the Autofac, 50 kg in a 1 m^3, is modeled on a regular machine shop, with the machinist replaced by a robot. The machine shop is 6250 kg in 125 m^3. I just scale it down by a factor of 5, and thereby reduce the duplication time by a factor of 5. So it duplicates in 5 weeks instead of 25 weeks. Suppose we start the Autofac versus the robot machine shop at the same time. After a year, there are 1000 Autofacs versus 4 machine shops; or in terms of mass, 50,000 kg of Autofac and 25,000 kg of machine shop. After two years, 50,000,000 kg of Autofac versus 100,000 kg of machine shop. After 3 years, it’s even more extreme. At any time, we can turn the Autofacs from making themselves to making what we need, or to making the tools to make what we need. The Autofac wins by orders of magnitude even if it’s teeny and inefficient, because of sheer speed.
That’s why I picked a one meter cube. I would have picked a smaller cube, that reproduced faster, but that would scale various production processes beyond reasonable limits. I didn’t want to venture beyond ordinary machining into weird techniques only watchmakers use.
This is certainly a consideration. Given the phenomenal reproductive capacity of the Autofac, there’s an enormous return to finishing design as quickly as possible and getting something out there.
Let me tell you some personal history. I happened upon the concept of self-reproducing machines as a child or teenager, in an old Scientific American from the fifties. This was in the 1970s. That article suggested building a self-reproducing factory boat, that would extract resources from the sea, and soon fill up the oceans and pile up on beaches. It wasn’t a very serious article. Then I went to MIT, in 1979. Self-reproducing machines were in the air—Eric Drexler was theorizing about mechanical bacteria, and NASA was paying people to think about what eventually became the 1981 lunar factory design study. I thought that sending a self-reproducing factory to the asteroid belt was the obvious right thing, and thought about it, in my baby-engineer fantasy way. But I could tell I was ahead of my time, so I turned my attention to supercomputers and robots and AI and other stuff for a few decades.
A few years ago I picked up the idea of self-reproducing boats again. I imagined a windmill on deck for power, and condensing Seacrete and magnesium from the water for materials. There was a machine shop below decks, building all the parts. But I couldn’t make the energy economy work out, even given the endless gales of the Southern Ocean. So I asked myself, what about just the machine shop part? Then I realized the reproduction time was the overriding consideration. How can I figure out the reproduction time? Well, I could estimate the time to do it with a regular human machine shop, and I remembered Eric Drexler’s scaling laws. And wow, five weeks?! That’s short enough to be a really big deal! So, a certain amount of calculation and spreadsheets later, here we are, the Autofac.
I considered varied environments for situating the Autofac:
a laboratory in Boston. Good for development, but doesn’t allow rapid growth.
a field near a railroad and power line in the Midwest. Good for the resource inputs, but the neighbors might reasonably complain when the steel mill starts belching flame, or the Autofacs pile up sky-high.
Baffin Island. Advantages described above.
Antarctic Icecap. Bigger than Baffin, but useful activities are illegal. Shortage of all elements except carbon, oxygen, nitrogen and hydrogen.
The Moon. Even bigger. Ironically, shortage of carbon, nitrogen and hydrogen. No wind, so the Autofac has to include solar cell manufacture from the git-go. There will be lots of problems understanding vacuum manufacturing. Obvious first step toward Dyson Sphere.
Carbonaceous asteroids. Obvious second step toward Dyson Sphere.
So, I decided to propose an intermediate environment. Obviously, it was rooted in the mid-20th-century visions of space exploration. But that didn’t set the size, or the use of Baffin Island, or anything else really. We’ll build a Dyson Sphere eventually, but I don’t feel the need to do it personally.
More to come.
That’s what’s doing the work here.
We can’t automate machining because an AI that can control a robot arm to do typical machinist things (EG:changing cutting tool inserts, removing stringy steel-wool-like tangles of chips, etc.) doesn’t exist or is not deployed.
If you have a robot arm + software solution that can do that it would massively drop operational costs which would lead to exponential growth.
The core problem is that currently we need the humans there.
To give concrete examples, a previous company where I worked had been trying to fully automate production for more than a decade. They had robotic machining cells with machine tools, parts cleaners and coordinate measuring machines to measure finished parts. Normal production was mostly automated in the sense that hands off production runs of 6+ hours were common, though particular cells might be needy and require frequent attention.
What humans had to do:
Changing cutting tools isn’t automated. An operator goes in with a screwdriver and box of carbide inserts 1-2x per shift.
The operators do a 30-60 minute setup to get part dimensions on size after changing inserts.
Rough machining can zero tools outside the machine since their tolerances are larger but someone is sitting there with a T-handle wrench so the 5-10 inserts on an indexable end mill have fresh cutting edges.
Intermittent problems operators handled:
A part isn’t perfectly clean when measuring. Bad measurement leads to bad adjustment and 1-2 parts are scrap
chips are too stringy and tangle up, clear the tangle every 15 mins
chips are getting between a part and fixture and messing up alignment, clean intermittently and pray.
That’s ignoring stupider stuff like:
Our measurement data processing software just ate the data for a production run so we have to stop production and remeasure 100 parts.
someone accidentally deleted some files so (same)
We don’t have the CAM done for this part scheduled to be produced so … find something else we can run or sit idle.
The employee running a CAM software workflow didn’t double-check the tool-paths and there’s a collision.
And that’s before you get to maintenance issues and (arguably) design defects in the machines themselves leading to frequent breakdowns.
The vision was that a completely automated system would respond to customer orders and schedule parts to be produced with AGVs carrying parts between operations. In practice the AGVs sta there for 10+ years because even the simple things proved nearly impossible to automate completely.
Conclusion
Despite all the problems, the automated cells were more productive than manual production (robots are consistent and don’t need breaks) and the company was making a lot of money. Not great automation is much much better than using manual operators.
It’s hard to grasp how much everything just barely works until you’ve spent a year somewhere that is trying to automate. Barely works is the standard in industry AFAIK so humans are still highly necessary.
It is, in theory, possible to automate almost everything and to build reliable machines and automation. The problem is O-ring theory of economic development. If tomorrow median IQ jumps +2SD automation would rapidly start to just work as actually good solutions are put in place. As is, organizations have to fight against institutional knowledge loss (employees leaving) just to maintain competence.
Yes, absolutely! A fine description of the current state of the art. I upvoted your post by 6 points (didn’t know I could do that!).
I’m imagining doing everything the machinist has to do with a mobile pair of robot arms. I can imagine a robot doing everything you listed in your first list of problems. Your “stupider stuff” is all software problems, so will be fixed once, centrally, and for good on the Autofac. The developers can debug their software as it fails, which is not a luxury machinists enjoy.
Call a problem that requires human input a “tough” problem. We can feed the solutions to any tough problems back into the model, using fine-tuning or putting it in the prompt. So ideally, any tough problem will have to be solved once. Or a small number of times, if the VLM is bad at generalizing. The longer we run the Autofacs, the more tough problems we hit, resolve, and never see again. With an exponentially increasing number of Autofacs, we might have to solve an exponentially increasing number of tough problems. This is infeasible and will destroy the scheme. We have to hope that the tough problems per hour per Autofac drops faster than the number of Autofacs increases. It’s a hope and only a hope—I can’t prove it’s the case.
What’s your feeling about the distribution of tough problems?
TLDR:autofac requires solving “make (almost) arbitrary metal parts” problem but that won’t close the loop. Hard problem is building automated/robust re-implementation of some of the economy requiring engineering effort not trial and error. Bottleneck is that including for autofac. Need STEM AI (Engineering AI mostly). Once that happens, economy gets taken over and grows rapidly as things start to actually work.
To expand on that:
“make (almost) arbitrary metal parts”
can generate a lot of economic value
requires essentially giant github repo of:
hardware designs:machine tools, robots, electronics
software:for machines/electronics, non-ai automation
better CAD/CAM (this is still mostly unsolved (CF:white collar CAD/CAM workers))
AI for tricky robotics stuff
estimate:100-1000 engineer years of labor from 99th percentile engineers.
Median engineers are counterproductive as demonstrated by current automation efforts not working.
EG:we had two robots develop “nerve damage” type repetitive strain injury because badly routed wires flexed too much. If designers aren’t careful/mindful of all design details things won’t be reliable.
This extends to sub-components.
Closing the loop needs much more than just “make (almost) arbitrary metal parts”. “build a steel mill and wire drawing equipment”, is just the start. There are too many vitamins needed representing unimplemented processes
A minimalist industrial core needs things like:
PCB/electronics manufacturing (including components)
IC manufacturing is its own mess
a lot of chemistry for plastics/lubricants
raw materials production (rock --> metal) and associated equipment
wire
Those in turn imply other things like:
refractory materials for furnaces
corrosion resistant coatings (nickel?, chromium?)
non-traditional machining (try making a wire drawing die with a milling machine/lathe)
ECM/EDM is unavoidable for many things
Things just snowball from there.
Efficiency improvements like carbide+coatings for cutting tools are also economically justified.
All of this is possible to design/build into an even bigger self-reproducing automated system but requires more engineer-hours put into a truly enormous git repo.
STEM AI development (“E” emphasis) is the enabler.
Addendum: simplifying the machine tools and robots
Simplifications can be made to cut down on vitamin cost of machine tools. Hydraulics really helps IMO:
servohydraulics for most motion (EG:linear machine tool axes, robots) to cut down on motor sizes and simplify manufacturing
efficiency is worse, but saves enormously on power electronics and manufacturing complexity.
Similar principles to hydraulic power steering used in cars.
Boston Dynamics ATLAS Robot uses rotary equivalent for joints (I think).
https://www.researchgate.net/figure/Comparison-of-the-Hy-Mo-actuator-with-traditional-hydraulic-actuator-52_fig5_346755993
hydrostatic bearings for rotary/linear motion in machine tools.
No hardened metal parts like in ball/roller bearings
Feeding fluid without flexible hoses across linear axes via structure similar to a double ended hydraulic cylinder. Just need two sliding rod seals and a hollow rod. Hole in the middle of the center rod lets fluid into the space between the rod seals.
Both bearings and motion can run off a single shared high pressure oil supply. Treat it like electricity/compressed-air and use one big pump for lots of machines/robots.
End result: machine tools with big spindle motors and small control motors for all axes. Robots use rotary equivalent. Massive reduction in per-axis power electronics, no ballscrews, no robot joint gears.
For Linear/rotary position encoders, calibrated capacitive encoders (same as used in digital calipers) are simple and needs just PCB manufacturing. Optical barcode based systems are also attractive but require an optical mouse worth of electronics/optics per axis, and maybe glass optics too.
This kind of thing is why I’m paying a lot of attention to Tesla these days. They seem like the likeliest candidate to close the loop first. Obsessed with robotics, vertical integration, and manufacturing automation. Emitting slogans like “the machine that makes the machine” and “the factory is the product”.
Yes. The alternate approach to achieving a self-reproducing machine is to build a humanoid robot that can be dropped into existing factories, then gradually replace the workers that build it with robots. That path may well be the one that succeeds. Either path delivers an enormous expansion of industrial capabilities.
Ah, hmm I don’t know if this is really true, the humanoid robots aren’t going to be replacing the humans doing debugging processes, which I feel is where a lot of the bottlenecks are, because you don’t have many humans who can do that.
So on reflection, closing the loop might require deliberately engineering things so that bugs don’t occur, or so that they’re solved in cruder ways that pre-AGI ML can do, more in the direction of ‘demolish and recycle the entire factory when it wears out’, but hopefully not that far, and that probably isn’t cost-effective. Unless it is??
Its part of the space colony philosophy Tesla + SpaceX want to achieve that. We won’t get an idea of how hard/easy it is until there are >100 people permanently living outside earth trying to make it happen (With >>10K helpers on earth trying to make it happen)
I do not intend to be rude by saying this, but I firmly believe you vastly overestimate how capable modern VLMs are and how capable LLMs are at performing tasks in a list, breaking down tasks into sub-tasks, and knowing when they’ve completed a task. AutoGPT and equivalents have not gotten significantly more capable since they first arose a year or two ago, despite the ability for new LLMs to call functions (which they have always been able to do with the slightest in-context reasoning), and it is unlikely they will ever get better until a more linear, reward loop, agentic focused learning pipeline is developed for them and significant amount of resources are dedicated to the training of new models with a higher causal comprehension.
I think you’re underestimating the complexity of inputs to a modern machine shop. For example, they need tool bits that can drill/mill steel; these don’t last forever and are a major cost.
Nowadays we use carbide bits, but we used to use steel bits to cut steel. It’s called high speed steel. It differs from regular steel by being a different crystal structure, that is harder (and more brittle). It used to be perfectly common to cut a steel-cutting tool out of steel, then apply heat treatment to induce the harder crystal structure, and use the hardened tool to cut the softer original steel. It’s one of the reasons I specified steel instead of aluminum or brass.
The machine shop can use a tool until it wears down too much, then un-harden it (a different heat treatment), cut it back to have a sharp edge again, and then re-harden. Steel really is amazing stuff.
I’ve looked into machine tool techniques pretty closely, and I believe I can make them with only 2% by weight that’s not steel or lubricant. In a lot of ways, it’s going back to the designs they used a hundred years ago, before they had good plastics or alloys. For example, the only place you HAVE to use plastic is as a flexible wire insulation.
I welcome your suggestions as to inputs I may have overlooked.
So...the entire industrial economy is basically an autofac. What you’re trying to do is simplify it, replacing eg plastics with more steel. But you seem to be expecting that to reduce costs, and the reason people use eg injection-molded plastics instead of steel enclosures is because it’s cheaper. Using carbide bits instead of “high-speed steel” (which requires uncommon metals, btw) is worthwhile, and by replacing things with alternatives that smart specialists have decided are not as good, you’re reducing the overall “replication factor/capability” and input-output efficiency relative to “the entire current industrial economy” as a competing design. The same goes for occasional human intervention for eg maintenance and lubrication—people have decided it’s more efficient overall, despite humans being expensive. It doesn’t make sense to take a design (for an entire economy, or anything else), add a bunch of arbitrary restrictions and simplifications, and expect it to be better in a way that reduces the costs of its products.
The simple autofac plan does oversimplify things in my perspective. I think there is sufficient value in specialization that only some of the autofac economy will be the autofac as described.
Obviously, there would need to be a lot of scaling before it would make sense to internally produce computer chips.
Other specialized robots and subunits do make sense. For example:
Automated diggers to collect the iron ore, and excavate sheltered underground spaces for the autofacs to operate in.
Hydrocarbon facilities which take in seawater, rock dust, waste ash from previous algae batches, and electricity, and use this to power lights and robots to grow algae. Then the algae can be processed into hydrocarbon feedstock for plastics, lubricants, and rust-resistant coatings.
Cranes for assembling wind turbines.
Mentioned elsewhere in the comments, large iron smelter facilities.
General weather-resistant maintenance bots, basically tractors robot arms and flatbed trailers that they tow for transport. They would install and maintain the powerlines and such.
Flywheels for maintaining stability of the primarily wind-based power grid.
Etc...
This still allows for the bulk of the autofac economy to be made up of autofacs that spend a fraction of their capacity producing/repairing/recycling these more specialized forms.
More than mere scaling, this would require equipment orders of magnitude more precise and the necessary ultra-clean environment and all the minutiae those entail. Microchip manufacturing is Hard.
.
1950s era computers likely couldn’t handle the complex AI tasks imagined here (doing image recognition; navigating rough Baffin Island terrain, finishing parts with hand tools, etc) without taking up much more than 1 meter cubed.
.
idk, you still have to fit video cameras and complex robotic arms and wifi equipment into that 1m^3 box, even if you are doing all the AI inference somewhere else! I have a much longer comment replying to the top-level post, where I try to analyze the concept of an autofac and what an optimized autofac design would really look like. Imagining a 100% self-contained design is a pretty cool intellectual exercise, but it’s hard to imagine a situation where it doesn’t make sense to import the most complex components from somewhere else (at least initially, until you can make computers that don’t take up 90% of your manufacturing output).
.
Feynman is imagining lots of components being made with “hand tools”, in order to cut down on the amount of specialized machinery we need. So you’d want sophisticated manipulators to use the tools, move the components, clean up bits of waste, etc. Plus of course for gathering raw resources and navigating Canadian tundra. And you’d need video cameras for the system to look at what it’s doing (otherwise you’d only have feed-forward controls in many situations, which would probably cause lots of cascading errors).
I don’t know how big a rasberry pi would be if it had to be hand-assembled from transistors big enough to pick up individually. So maybe it’s doable!
.
I was actually thinking of a pair of humanlike arms with many degrees of freedom, and one or more cameras looking at things. You can have dozens of single datum sensors, or one camera. It’s much cheaper. Similarly, once you have some robot arms, there’s no gain in including many single use motors. For example, when I include an arbor press, I don’t mean a motorized press. I mean a big lever that you grab with the robot arm and pull down, to press in a shaft or shape a screw head.
There are two CNC machine tools, to automate some part shaping while the robot does something else.
Mere scaling? Scaling is doing a lot here. Like, an economy the size of the UK’s or something. I do agree that this would require its own set of chip-fab specialized autofacs. Related: I enjoyed Breaking Tap’s recent video on working towards DIY chip fab. https://youtu.be/RuVS7MsQk4Y?si=EwQt9e_7BB-KVKAy
I see, I suppose I interpreted ‘scaling’ a bit less generally. In that case I agree.
Also I just noticed you mentioned flywheels, which are one of my favorite pieces of technology. I long for someone to make a phone with a flywheel battery as a meme/gag gift.
You could go some way with 1980s-level integrated circuits for all the onboard electronics. The manufacturing requirements are much more tolerable. But even 1980s semiconductors require a couple of dozen chemically exotic and ultra pure feedstocks. The Autofacs would have to build a complex chemical industry before they could start building chips.
How many different alloys are you expecting to have in stock in there?
HSS is apparently pretty finicky to work with. As I understand it most hobbyists content themselves with O1 or A1 or whatever, which wear a lot faster. But it’s true they’ll cut themselves.
There’s probably a reason, above and beyond cost, why the “bodies” of machines tend to be cast iron rather than tool steel. And for the truly staggering number of standardized alloys that are out there in general.
Well, I seem to be talking to someone who knows more about alloys than I do. How many alloys do you think I need? I figure there’s a need for Neodymium Iron Boron, for motor cores, Cast Iron in the form of near-net-shape castings for machine frames, and some kind of hardenable tool steel for everything else. But I’m uncertain about the “everything else”.
I don’t think the “staggering number of standardized alloys” needs to alarm us. There are also a staggering number of standardized fasteners out there, but I think 4 sizes of machine screws will suffice for the Autofac. We don’t need the ultimate in specialized efficiency that all those alloys give us.
Maybe. But what I know tends to be very patchy, depending on what rabbit holes I happen to have gone down at various times.
I hadn’t thought about magnetics at all, or anything exotic. I was just talking about basic steel.
Unless I’m mixed up, NdFeB is for permanent magnets. You might not need any permanent magnets. If you do, I believe also you need a big solenoid, possibly in an oven, to magnetize them. Said solenoid needs a metric butt-ton of current when it’s on, by the way, although it probably doesn’t have to be on for long.
Inductor and electromagnet cores, including for motors, are made out of “electrical steel”, which is typically cut to shape in thin plates, then laminated with some kind of lacquer or something for insulation against eddy currents. You can also use sintered ferrite powders, which come in a bewildering array of formulations, but if you’re just worried about motors, you’d probably only really need one or two.
Those plates are an example of a generalized issue, by the way. I think those plates are probably normally hot die cut in a roll process. In fact, I suspect they’re normally made in a plant that can immediately drop the offcuts, probably still hot to save energy on reheating them, into equipment that rerolls them back into more stock. Or maybe they even roll them out in their final shapes directly from a melt somehow.
You could mill every single plate in a motor core out of sheet stock on a milling machine… but it would take eternity, go through a ton of tooling, and generate a lot of waste (probably in the form of oily swarf mixed in with oily swarf of every other thing you process in the shop).
There are lots of processes like that, where stuff that you could “hand make” with the “mother machines” isn’t made that way in practice, because specialized machines, often colocated with other specialized machines in large specialized plants, are qualitatively more efficient in terms of time, energy, waste, consumables, you name it. Stuff that’s hot is kept hot until it needs to be cool (and often you try to cool it by putting as much as possible of the heat back into an input stream). Steps are colocated to avoid reheats. Waste products are recycled or used for something else, and the plant for “something else” is often also colocated.
It’s really hard to compete with that kind of efficiency. Most of the individual specialized machines are a lot more than a cubic meter, too. You already mentioned that temperature-sensitive processes tend to have optimal sizes, which are often really big.
Can you afford to use 10 times the energy and produce 10 times the waste of “traditional” processes? If not, you may need a lot of specialized equipment, more than you could fit in a reasonable-sized self-replicating module.
All castings are imported, right?
By the way, you need nichrome or Kanthal or something like that for the heating elements in your furnace. Which isn’t really different from the copper wire you use, but it’s another item.
Here I break down. I suspect, but do not know, that if you only think in terms of making usable parts, you could at least get away only with “mild steel”, “alloy steel”, “tool steel”, and perhaps “spring steel”. Or maybe with only three or even two of those. I could be wrong, though, because there are tons of weird issues when you start to think about the actual stresses a part will experience and the environment it’ll be in.
If you do want to reduce the number of alloys to the absolute minimum, you probably also have to be able to be very sophisticated about your heat treating. I’d be pretty shocked, for instance, if a high-quality bearing ball is actually in the same condition all the way through. You’d want to be able to case-harden things and carburize things and do other stuff I don’t even know about. And, by the way, where are you quenching the stuff?
Even if you can use ingenuity to only absolutely need a relatively small number of alloys, on a similar theme to what I said above, there’s efficiency to worry about. The reason there are so many standard alloys isn’t necessarily that you can’t substitute X for Y, but that X costs three or four or ten times as much as Y for the specific application that Y is optimized for. Costs come from the ingredients, from their purification, from their processing when the alloy is formulated, and from post-processing (how hard is the stuff on your tooling, how much wear and tear does it put on the heating elements in your furnace, how much energy do you use, how much coolant do you go through, etc).
As I describe in my first reply to Jackson Wagner above, I can tolerate some inefficiency, as long as I stay above Soviet-style negative productivity. The goal is minimum reproduction time. Once I’ve scaled up, I can build a rolling mill if needed.
As you point out, that would be madness. I’ve got a sheet rolling machine listed, so I assume I can take plate and cold-roll it into sheet. Or heat the plate and hot-roll it if need be. The sheets are only a meter long and a few centimeters wide, so the rolling machine fits inside. They function like shingles for building the outside enclosure, and for various machine guards internally, so they don’t have to be big.
I’m quenching in a jar of used lubricant. Or fresh oil, if need be. 6% of the input is oil.
I’m a little reluctant to introduce this kind of evidence, but I’ve seen lots of machinist videos where they say “I pulled this out of the scrap bin, not sure what it is, but lets use it for this mandrel” (or whatever). And then it works fine. I am happy to believe that different alloys differ by tens of percent in their characteristics, and that getting the right alloy is an important occupation for real engineers. I just don’t think that many thousands of them all vary by “three or four or ten times.” I think I can get away with six or so.
I agree with jbash that the value of larger, more specialized structures is sufficient to justify the autofac economy being made up of units besides just the autofacs.
That doesn’t seem like a blocker to the general concept. You just need a slightly more complicated plan. With the rate at which AI is progressing these days, it doesn’t seem to me that plan complication is going to be a barrier.
So the real questions that stand out to me are:
What would the seed cost be to get the initial autofac economy to the scale that it’s exports would more than pay for its imports?
Of the various potential specialized facilities, which ones make economic sense at which scales?
What sort of timeframe does it make sense to operate at sub-profitable scale for, while building things which have internal value that outweighs the cost of interest incurred by operating at a loss?
How much more powerful than current AI would the VLM control system need to be for the initial stages?
Would the development pace of AI keep pace with the increasing complexity as the autofac economy ramps up? (my guess is yes)
What things are worth importing versus making do with self-built?
Computer chips, certainly. Cutting lasers? Probably at least some. The initial parts for the first set of autofacs. Specialty metals and alloys. It’s more a question of how much to import, and the costs versus export value.
Macroscopic self-replicators are extremely powerful, and provide much of the power of nanotech without relying on nanotech. Seems like they might be worth mentioning more often as a rhetorical tool against those who dismiss anyone who mentions nanotechnology.
Back when I read about people claiming a RepRap can reproduce itself, I felt like the claim implied it would build the electronics of the new RepRap from scratch as well and was confused since obviously a 3D printer can’t double as a chip fab. The gold standard for a self-replicating machine for me is something like plants, which can turn high-entropy raw materials like soil and ores into itself given a source of energy. I guess you could talk about autotrophic self-reproducing machines that can do their thing given a barren planet and sunlight, and heterotrophic self-reproducing machines that have selling machined components over the internet and using the income to buy CPU chips and hire workers to assemble the skeleton of a new automated workshop as a valid strategy.
There is also Project Quine, which is a newer attempt to build a self-replicating 3D printer
I would love to see someone actually do this.
It might get people take the idea of self-replicating AI seriously
The abundance provided by something like this might convince people that they actually don’t need to AGI to solve all of the worlds issues
#2 is why I’m coming up with this scheme despite my substantial p(doom). I think we can do something like this with subhuman (hence non-dangerous) levels of AI. Material abundance is one of the things we expect from the Singularity. This provides abundance without superhuman AI, reducing the impetus toward it.
Just out of curiousity, how many of those tools have you personally run? How many have you built and/or maintained?
I’ve operated a lathe and a mill (both entirely manual), various handheld power tools, a robot arm, robot eyes, autonomous mobile robots, and a data center. For the rest, I’ve read books and watched videos.
I’ve built and/or maintained various kinds of robots.
I have no experience with cutting-edge VLMs.
How much of a loss of precision would we expect in one generation of autofacs?
As a concrete example, let’s say one of the components of an autofac is a 0.03125 inch (±0.1 thousandths) CNC drill bit. Can your autofac make another such drill bit out of the same material and at the same level of precision?
If not, maybe we have to ship in the drill bits as well. But there are a large number of things like this, and at some point you’ve got a box that can assemble copies of itself from prefabricated parts, but uses a pretty standard supply chain to obtain those prefabricated parts. Which, to be clear, would still be pretty cool.
There are standard ways to make more precise tools from less precise tools. The methods were invented 1750-1840 to allow upgrading handmade metal tools to the precision of thousandths of inches we enjoy today. We just have to apply such methods a little bit at every generation to keep the level of precision constant.
Don’t those methods tend to rely on lapping and optics? You don’t seem to have any equipment for those.
They depend on lapping, which can be done “manually” by the robot arm. I forgot to list “abrasive powder” in my list of vitamin ingredients. Fixed now.
The fancier optical techniques provide precision on the order of a wavelength, which is far in excess of our needs. All we need is eyeball-class optical techniques like looking along an edge to make sure it’s straghtish, or pressing a part against a surface plate and seeing if light passes under it.
I really like this video on the subject of how humanity first developed precision machines: https://youtu.be/gNRnrn5DE58?si=NMk1BAsU6_XMuLbL
See also: https://youtu.be/djB9oK6pkbA?si=ObFPCZkR-Rx9AC2G
This has parallels with how the factory-building game Factorio presents things. The thing that makes Factorio fun[1] is how it abstracts away those pesky prohibitively complex nuances of manufacturing & automation so that everything can feasibly be automated quickly and scaled ad infinitum. For example:
The conveyor belts run on magic (they don’t require any power, which isn’t really explained considering every other electrical thing in the game requires pseudo-realistic levels of electrical input.)
The assembling machines (essentially Autofacs) don’t require any retooling/tuning/cleaning/etc to switch between completely different recipes seamlessly)
No manufacturing equipment ever wears out or needs physical maintenance.
Inserters (robot hands that handle objects to and from conveyor belts and the various structures) detect objects flawlessly and reliably and any kind can handle any type of object (and furthermore object sizes are abstracted such that all take the same amount of space on a conveyor belt)
Electrical usage is abstracted so that you only need to keep generated power >= usage or things will start to slow down/get rolling blackouts. You don’t need to worry about pesky throughput tolerances on cables, one small wooden power pole and its wire can handle just as much as a giant cross-country steel behemoth.
& thousands of other small details abstracted away for simplicity.
Overall I hope we are able to progress to Autofacs in real life, I just don’t see it being nearly as straightforward as any of us would prefer. Not that I want to discourage anyone from making the attempt! I just hope that they know what they are getting into.
Really really fun for engineering-minded people like myself.
Fun-hazard level — If you’ve never tried it before it might be wise not to unless you have incredible self discipline or several days of free time in the near future; It can be addicting.
I’ve been avoiding Factorio. I watched a couple of videos of people playing it, and it was obviously the most interesting game in the world, and if I tried it my entire life would get sucked in. So I did the stoic thing, and simply didn’t allow myself to be tempted.
You bring up a good point that the amount of picky details that need to be dealt with is huge. I basically think that this whole plan is infeasible unless the controlling AI is near or above AGI. So, by the time such a project could be launched, there will be a lot of other disruptive things also happening in the world.
The fact that this could become reality inless than 5 years from now is both exciting and terrifying. Disruptive change indeed!
The swept area of a wind turbine scales as the second power of the height (assuming constant aspect ratios), and the velocity of wind increases with ~1/7 power with height. Since the power goes with the third power of the velocity, that means overall power ~height^2.4. The problem is that the amount of material required scales roughly with the 3rd power of the height. This would be exactly the case with constant aspect ratios. The actual case and the scale up of wind turbines over the last few decades has not scaled that fast, partly because of higher strength materials and partly because of optimization. Anyway, I agree there are economies of scale from micro wind turbines, but they aren’t that large from a material perspective (mostly driven by labour savings).
Now I know more! Thanks.
That would suggest that an equal mass of tiny wind turbines would be more efficient. But I see really big turbines all over the midwest. What’s the explanation?
As I mentioned, the mass scaling was lower than the 3rd power (also because the designs went from fixed to variable RPM and blade pitch, which reduces loading), so if it were lower than 2.4, that would mean larger wind turbines would use slightly lower mass per energy produced. But the main reason for large turbines is lower construction and maintenance labour per energy produced (this is especially true for offshore turbines where maintenance is very expensive).
If I did not see a section in your bio about being an engineer who has worked in multiple relevant areas, I would dismiss this post as a fantasy from someone who does not appreciate how hard building stuff is; a “big picture guy” who does not realise that imagining the robot is dramatically easier than designing and building one which works.
Given that you know you are not the first person to imagine this kind of machine, or even the first with a rough plan to build one, why do you think that your plan has a greater chance of success than other individuals or groups which have tried before you? Is there something specific that you bring to the table that means you will avoid the challenges or be more suited to tackle them?
I think you might do better starting out with creating a machine which can assemble a copy of itself, from pre-built off-the-shelf parts. A robot arm and camera attachment, which is capable of recognising the pre-made parts of itself and fitting them together autonomously would be very challenging to make and would be a good proof of concept for the larger project.
If you have this system working, then your next step would be creating the same machine but including a 3d print bed (which it is also capable of assembling) or small scale milling machine to build a few of the parts, and continue by adding more and more manufacturing capabilities, so you have to supply fewer parts with each iteration of the design. I remember assembling my 3d printer a few years ago, and there were quite a lot of steps which would be major practical challenges to a robot a similar size to the printer, even in just assembling the pre-made parts.
I’m glad you wrote this, it adds some interesting context that was unfamiliar to me for this market I opened around a week ago: https://manifold.markets/dogway/which-is-the-earliest-year-well-hav#wji33pv4fcj
I was entertaining the possibility of a powder or fluid-based metal as an input to a 3D printer which works today for fabricating metal components and seems likely to improve significantly with time. I was considering this avenue to be the most likely way that the threshold of full fidelity-preserving self-reproduction is passed, but I have no expertise in this area. It seems like the 3D printing path, if viable, would alleviate the need for some of the tools in the Autofac setup and might reduce reproduction time considerably.
Yeah, I looked at various forms of printing from powder as a productive system. The problem is that the powder is very expensive, more expensive than most of the parts that can be produced from it. And it can’t produce some parts— like ball bearings or cast iron— so you need tools to make those. And by the time you add those in, it turns out you don’t need powder metallurgy.
“most”, but not all. How does the Autofac generate the control system for the next Autofac? Doesn’t this require a chip fab, or are we just hand waving away the need for more processors and just saying we will import them from outside like we do the raw materials?
Yes, that’s totally part of the starter pack. All the electronics are imported—CPUs, radios, cameras, lights, voltage converters, wire harnesses, motor controllers...
I don’t know how to plan the split between the part of the thinking that is done inside the Autofac and the part that is done in the data center.
I’m not quite sure how much of an AI is needed here. Current 3d printing uses no AI and barely a feedback loop. It just mechanistically does a long sequence of preprogrammed actions.
Some notes on self-replicating machines: Complexity/precision: The dexterity required to move a few wires into a crude machine is far in excess of the dexterity of that crude machine. Generally, designing something which can produce itself is complex for that reason, the relationship between complexity and ability to create complex things is nonlinear, difficult to affect in useful ways, and hard to measure without creating real test objects. Very complex objects(like organisms) can assemble ‘copies’ of themselves, through complex and error prone processes.
Have you heard of Neil Gershenfeld?
Yes. I know him. We met years ago when I was a grad student at the Media Lab. I haven’t followed his work on self-reproduction in detail, but from what I’ve seen he is not aiming at economically self-sufficient devices, while I am. So I’m not too impressed.
Going from this to a 1M cube is a big jump. This comment is the basis for the enthusiasm for space colonies etc. E.g. SpaceX says 1 million people are needed, vs 1M cube, huge uncertainty in the scale of the difference. To me, almost all the difficulty is in the inputs, especially electronics.
I think you’re missing a few parts. The Autofac (as specified) cannot reproduce the chips and circuit boards required for the AI, the cameras’ lenses and sensors, or the robot’s sensors and motor controllers. I don’t think this is an insurmountable hurdle: a low-tech (not cutting-edge) set of chips and discrete components would serve well enough for a stationary computer. Similarly, high-res sensors are not required. (Take it slow and replace physical resolution with temporal resolution and multiple samples.)
Second, the reproduced Autofacs should be built on movable platforms so different groups can get their own. (Someone comes with a truck and a few forklifts, lifts the platform onto the truck, and drives the Autofac to the new location.)
If you look at what I wrote, you will see that I covered both of these.
I really like this idea, especially the part about doing it on Baffin Island. A few questions/comments/concerns
During the winter, the polar ice cap expands to the point that Baffin Island is surrounded by ice. This makes shipping things to and from the island difficult for a large part of the year. I also imagine most people don’t want to be there during the winter to check up on things. Do you imagine things progressing more slowly during the winter because of this?
Looking at the climate data for Baffin island and comparing mean daily maximum in July and mean daily minimum in January, it looks like there’s a range of about 40 Celsius, which seems significant (and over a range pretty different from what most engineers are used to building for). Do you expect this will interfere with the equipment? Will the Autofacs need some kind of temperature control?
Do you have any ideas for how to deal with defective/broken AutoFacs? My first thought is that you could automatically disassemble them, throw away the defective parts and use the working parts to build new Autofacs. There’s probably something more clever.
Will the AutoFacs be able to clean themselves or fix the other normal, small things that worsen performance as machinery operates? If so, how?