Prompt-to-Print: Manufacturing in an AI Age
VibeCADing is going to be a thing
Dear SoTA,
More than a decade ago, I wrote a book called “Makers”, which argued that 3D printing and open source innovation was going to create “A New Industrial Revolution” (my admittedly hyperbolic subtitle) that would bring about a return of advanced manufacturing to the West, with an accompanying boom in economic prosperity and entrepreneurship.
I was wrong. 3D printing got stuck between a hobbyist niche and an industrial boom-bust cycle and most of the decade passed with nary a dent in the dominance of traditional Chinese manufacturing. For many of the subsequent years, 3D printing remained a useful prototyping tool, but not much more than that. Big factories got bigger and hardware entrepreneurs continued to fly to Shenzhen to get stuff made.
But maybe I was just too early. The arrival of AI is reigniting talk of a new industrial revolution again and this time, as they say, it’s different.
What’s the connection between AI and manufacturing? That answer lies upstream of fabrication, in the earlier design stage that has traditionally been the domain of arcane computer-aided design (CAD) tools and the mechanical engineers that wield them. Just as AI has turned software programming into vibecoding, it’s starting to do the same with design: call it vibeCADing.
Right now it’s relatively easy to use AI-driven CAD tools to generate virtual 3D models (“text-to-mesh”) that can be used in videogames and movies, saving the time it would normally take a 3D modeller and animator to do by hand. This is not much harder than getting AI to generate images or videos of any sort. As long as you’re willing to accept something that’s kinda what you asked for and try to describe your changes in words to an AI that doesn’t actually know what it’s created, to say nothing of being able to edit any specific part of the object directly, it works. But to generate a design that can be functional in the real world is way harder.
For starters, real-world objects with real functions have to achieve certain weight, strength and mechanical properties, which requires knowing how the parts of the object work together and the various forces and loads each bears, often right down the micron level. It also requires knowing what the capabilities are of the machine making the object as well as those of the materials that machine will be using. This is called “design for manufacturability” (DFM) and is an entire profession for people, and even they struggle with it...
As an example, imagine that you want a gearbox for a humanoid robot joint. This gearbox translates fast-turning motor rotation on one side into slower rotation on the other side, but with much more torque, a process called “reduction”. The boundary conditions, which we’ll call “hard constraints”, are the physical dimensions of the motor, including its mounting holes and shaft, and the dimensions of the output shaft, which must fit into some other mechanism in the robot. There are probably some other requirements, which tend to be a little softer but still important: backlash, back driveability, wear tolerance, maybe weight, cost, size and frictional loss, along with manufacturability. Those might be considered dependent variables or “trades”; they are consequences of the design of the gearbox but maybe you’re willing to accept a bit more weight for better wear tolerance, and so forth.
Balancing across all these trades usually requires engineering degrees and years if not decades of experience. But surely this is exactly the sort of high-dimensional optimization that AI is ideal for, right? Everything that a mechanical engineer knows can, in principle, be taught to an AI; indeed, because mechanical engineers are good at writing things down and otherwise documenting their work, no active teaching may be necessary. Any company can just train or fine tune a model on their own repositories of files, much as they would on their internal code repositories.
Indeed, dozens of companies, from startups such as Zoo and Adam to CAD giants like Autodesk, are doing just this. Today you can describe what you want in plain English and the AI, trained on tens or hundreds of thousands of CAD files, can generate one for you (“text-to-CAD”). Even better, since the AI is equally adept at physics, materials, supply chain economics and manufacturing capabilities, it can ensure that the design is optimized across all those dimensions, something it would take a lifetime of experience, expensive and complex simulation tools, and often weeks for a human engineer or designer to do.
Back to the humanoid joint example. Here’s one such solution, which is called a planetary gear train:
Credit: Xuefeng Wang
For an engineer to design such a gearbox could entail weeks or months of design, prototyping and testing and then even longer to do the tooling to make it. Yet I was able to do the same in less than ten minutes with a single prompt (“Make me a planetary gearbox with a 10:1 reduction”) with Zoo and Adam, both of which helpfully identified and listed the key parameters that I could tweak (see screenshot), since I was pretty imprecise in my initial prompt. I didn’t even need to know all the elements of such a gearbox or think ahead of time of all my requirements – a simple high-level request generates something useful that I can review and tweak as needed; call it just-in-time requirements.
Screenshot from Zoo Design Studio
What’s even better, the output of these CAD AIs are not just polygon meshes, as you might find in a videogame 3D model generator, but real CAD with each part having its own identity and parameters that you can change. And all without the usual manual CAD process of drawing, extruding, transforming and otherwise hand-crafting each individual part. Better yet, the CAD is not described in massive “feature trees” and history of each transformation, as in standard CAD, but in a human-readable programming language (using either KCL or OpenSCAD).
In short, designing mechanical parts with AI is now nearly as easy as writing using AI to generate code because it is code. Code-based CAD is AI-friendly in a way that drawing-based CAD never was. A generation of would-be mechanical engineers and designers are now empowered to realize their dreams without having to climb a career-long learning curve. Anyone can do it.
Six months ago, these were demos that rarely worked. Six months earlier they were pitch decks. A year before that they were a fantasy. This is what AI has done to engineering in the past two years. It is nothing short of remarkable.
But there’s still much more work to do to really compete with full-stack product design, fully realizing the vision of “prompt-to-print”. The Zoo and Adam example I gave above don’t (yet) have a physics simulator in the loop and don’t really know DFM. But in principle, they could. Experiments done by PhysicsX in London (full disclosure: funded in part by the stealth company I’ve been part of) have demonstrated the feasibility of doing the entire design process within the latent space of agentically orchestrated AI models trained on all the relevant physics of a product, ranging from the physics of its environment to the internal physics of its own structure.
Last year PhysicsX released a tiny glimpse of this with a toy demo that showed how AI might design an airplane, starting with a high-level function prompt such as “design a glider that can carry 10kg with a wingspan of less than 2m and a glide ratio of 40:1”.
The first step is for the AI to explore the search space of airframe shapes that will have the desired aerodynamic properties. This is mostly the physics of fluids (air is a fluid) so it uses an AI “surrogate” trained on complex fluid dynamics (CFD).
Then, once the ideal shape of the airframe, which in aerospace lingo is called the Outer Mold Line, or OML, is found, the AI switches to the internals, where the forces imposed by the air pressure on the outside of the aircraft must be distributed and borne by the interior structure. This uses another AI surrogate trained on Finite Element Analysis (FEA) physics. This model minimizes the material to just that needed to handle the forces, a process called topology optimization, also done entirely with machine learning rather than classic numerical computation so it happens in seconds, not minutes or hours.
But to do this optimization, it must know what kind of material the aircraft will be made out of. If it’s made up of carbon fibres in bundles rather than sheets, which are stronger lengthwise than they are width-wise (they have “anisotropic” properties), it’s not just a matter of specifying the material; you must also specify how it is deposited. For that, you need to not just describe the object, but also the steps necessary to make the object. If it’s a machine, such as a 3D printer, that’s making the object, these steps, or commands, are called toolpaths.
For an anisotropic material, this means laying the fibers along the load lines of the forces it will experience, and if those forces are 3D, then the fibers have to laid in equally 3D directions, which typically means a robot arm with six or seven degrees of freedom rather than a regular 3-axis 3D printer. So for the AI to do this, it needs to know what kind of robot arm will ultimately be doing the printing, with all of its dimensions and capabilities.
Think of this sort of DFM as being “machine in the loop”, as shown in the diagram below. This describes an agentic coordination and negotiation between all the different layers of physics and manufacturing considerations, with the only human input being the prompt at the start and then high-level thumbs-up/thumbs-down on the AI candidates at the end (a process called reinforcement learning with human feedback, or RLHF).
The reason to make DFM an intrinsic part of the process is that designing something that can’t be made is pointless, and limiting designs to things that can be made by any generic machine is suboptimal. Rather than generating CAD as an intermediate step, which is useful for humans to edit or for regular manufacturing to ingest, this sort of agentic flow can go straight from design to toolpaths that are automatically generated in the AI’s internal latent space. It’s no longer understandable by humans, but then again it doesn’t have to be. It’s machines talking directly to machines, cutting out the middleman.
Replicators, Finally
Finally, about those magical 3D printers that can make anything AI can design. Yes, we got a bit overexcited a decade ago about them and the industry did indeed crash in the intervening years. But since 2022, it’s been a whole different story, largely because of the entry of world-class Chinese consumer electronics companies such as Bambu, which is a spinout from DJI, the market-leading drone company based in Shenzhen.
Today’s Bambus, which can cost as little as $300, are around six times faster than even professional 3D printers of a decade ago and “just work”, capitalizing on the Apple-style design, reliability and ease-of-use that Apple seeded in the Shenzhen ecosystem with decades of outsourced manufacturing.
The result is that 3D printing (known as “additive” manufacturing, because the machines lay down material where there was none, as opposed to “subtractive” technology where CNC machines cut away material from blocks of raw material) is increasingly moving beyond prototyping to become a viable manufacturing process for plastic parts that don’t require high strength or polished surface finish. The breakeven compared to injection molding, which requires expensive tooling before you can make the first part, now often exceeds 10,000 units (see illustrative diagram). And that doesn’t even factor in flexibility (every part can be different, and designs can be improved over a run) and speed (parts can be generated in minutes, compared to a month or more to get an injection mold made).
Representative figure for typical parts. Actual numbers depend on part complexity and size.
Although this improvement is most dramatic in the 3D printing that uses plastics of various sorts, a process called Fused Deposition Modelling (FDM), printing with metals, continuous carbon or fiberglass fibre strands and even ceramic paste is also advancing quickly in the size of what they can make, along with speed, quality and cost.
Today, many of the drones used in Ukraine are 3D-printed, as are SpaceX rocket motors and even metal parts of the Apple iWatch and iPhone Air. Modern “print farms” can have as many as 10,000 3D printers working 24/7 in massive warehouses. Although Bambu only launched in 2022, it has already sold an estimated 3 million printers, as many as the entire industry sold in the previous decade.
An industrial revolution that looks more like nature
Combine these two trends and you really do have the ingredients of a revolution: next-gen AI designs products that next-gen additives make.
But let’s go further. What would happen if you completely reimagined manufacturing with AI? Not just a little tweaking here and there – slightly better supply chain management or slightly easier CAD – but really ask the question: could we take humans out of the loop entirely and just let AI make things on its own? (note to self: make sure we don’t accidentally mine the planet to maximally make paperclips) If so, how would it do it differently?
We don’t need to completely imagine this. We already have a global superintelligence that makes things, called Nature. There is no Grand Designer – evolution just does the usual exploration/exploitation design variation with survival and reproduction as the fitness function. The “factory” is the cell and objects large and small grow to extraordinary complexity through the process of cells following relatively simple rules and gradients. Living things don’t have engineering precision, but they don’t need it, either. They just work, in a loosey goosey sort of way.
Now think about how an AI would design things, especially if it could be paired with a magical machine, perhaps even augmented by biology, that could make anything it designed. Or, more practically, a slightly less magical fabricator that could still make a broad class of things, and the AI would just design within that class. It could also be a little loosey goosey. No need for engineering precision in order to use standard fasteners or off-the-shelf parts. No need for human-readable steps or descriptions. It would just come out of the fabricator working, fit for purpose if not traditional manufacturing processes.
This gets into sci-fi territory pretty quickly, but if we already know we can engineer biology with AI, without waiting for evolution. On Earth we might want to use the factories and supply chains we already have, but how about on Mars and beyond, where Amazon doesn’t deliver? Maybe AI-designed microbes, which the Astera Institute’s Pioneer Labs are designing, will speedrun the industrial revolution on Mars, terraforming the planet for eventual human habitation in the process.
In 1966, nearly ten years after his death, John von Neumann’s “Theory of Self-Reproducing Automata” was published, which popularized the idea of what are now called “von Neumann machines”. These are micro-factories that can make more copies of themselves, along with almost anything else. His idea was that if we are going to spread to the stars, it probably won’t be accompanied by our human bodies, but rather in the form of such self-replicating robots that can harvest and use minerals and other feedstock from wherever they land to create machine civilizations that retain some essence of human intelligence.
Von Neumann died before he could witness AI or, for that matter, 3D printers. But had he lived, we would have recognized the seeds of his idea in what we’re inventing now.
Yours,
Chris Anderson
Chris Anderson is a senior advisor with Renaissance Philanthropy and PillarVC. He previously ran an AI and Advanced Manufacturing lab that was a spinout from Google co-founder Larry Page’s airtaxi company Kittyhawk, where Anderson was CTO. Prior to that he was CEO of 3D Robotics, the drone pioneer, and before that the longtime editor of Wired Magazine. His background is in science, starting with computational physics at Los Alamos National Lab and work at Nature, Science and The Economist. He splits his time between San Francisco and London. He’s on X at @chr1sa.
Write to the Society for Technological Advancement on letters@ilikethefuture.com.