Skip to content
Fully Doxxed

Every Machine Is About to Become Readable

— AI, Hardware, Manufacturing — 7 min read

Commercial equipment is expensive. Walk into a factory, a business, or a medical office and you can often find it from its pride-of-place location: pricey equipment that makes a business or product possible, or at the very least distinguishes the professional from the amateur. A specialty printer, a diagnostic instrument, a packing line, a robot arm. Maybe it costs tens of thousands of dollars, or hundreds of thousands of dollars. Sometimes it costs even more.

But why is it so expensive? The price of niche capital equipment is the cost of amortized R&D spread out and captured as margin in the low volume of machines produced.

Why don't competitors just copy expensive machines and drive down the costs? Is it patents? Contractual agreements? Physical locks? No, I don't think so. It's difficulty. A competitor who wants to copy an expensive piece of commercial equipment must re-derive all of the expensive engineering judgment baked into it. That re-derivation is costly and hard.

Or at least it was. New AI frontier models are changing the game. In the last few years, and especially in the last few generations, frontier models have gotten really good at skills that are changing what is possible when it comes to building real physical machines. And it has implications that will change how commercial equipment is designed, made, and purchased.

It started with software. In the last year, frontier models have gotten really good at reverse-engineering software into code with nearly identical features to the original. One developer used Codex to make a bug-for-bug remake of a twenty-year-old game.

Even more impressive is an engineer who used frontier models to create a working emulator of an Intellivision video game system, including simulated microchips and a bus, in just 36 hours.

Further examples include reverse-engineering a miniature hard drive, an IoT device, a proprietary radio system, an electronic sleep mask, and a home battery and inverter. In short, the software component of any hardware has very likely become readable recently.

And, with the recent release of Claude Mythos/Fable, we are seeing frontier AI that is really good at making CAD files. During the short window that Fable was publicly available, we saw examples of quasi-direct-drive robot joints, a model of a V8 engine, and even a humanoid robot.

Even before Fable, people were making 3D CAD files for things like a MagSafe stand. The utility of the designs has taken a big jump with models like GPT-5.5 and Claude Opus 4.7, which had large jumps in functionality, manufacturability, and assemblability compared with the previous generation of models, according to the MUSE benchmark. Increasingly, with the right information, frontier AI can make ever better CAD representations of parts.

But it's not enough to be able to make desired parts to do virtual disassembly of machines. You need part recognition for commercial parts, you need to break assemblies down into manufacturable parts, and you need to turn images into CAD files. Each of these is seeing progress:

Where disassembly is not possible, industrial CT is being used for scan-to-CAD.

In short, we have the building blocks to make a multimodal reverse-engineering system that can combine software capture, photographs, video, OCR, catalog search, manuals, existing CAD, measurements, and CT data to reverse-engineer a machine.

Now, this doesn't get us to 100% on all machines. It may still be difficult to determine important properties of key parts and processes, especially for parts that need specific materials and construction. Existing tools may make it difficult to determine things like alloys, heat treatment, grain structure, residual stress, detailed tolerances, and so forth. Where that matters, like semiconductors, advanced alloys, specialty chemicals, pharma, etc., machines may not be fully copyable.

You might think of every machine as consisting of two layers: a design layer, which includes the geometry, the control logic, and the embodied engineering judgment, and a production layer, which includes the actual fabrication, materials, tooling, QA, and supply chain. Readability collapses the cost of recovering the design layer, prying it loose from the physical layer and rendering physical machines much more like open source software: easier to copy, fork, and iterate on. So for machines where a lot of the value is in the design layer, like those built from off-the-shelf actuators, CNC, or printing, those machines will increasingly have software-like properties.

Readable machines have a few interesting consequences.

The first is that the frontier commoditizes almost immediately. Acquiring the best existing design is now cheap, and competitors can re-derive all of the embedded engineering judgment of the best machines at a fraction of what it cost to produce. Any high-value machine draws a crowd of new entrants, because starting at the frontier is no longer expensive. Still, what copying can't hand them is the ability to execute. With design no longer being the bottleneck, the scarce things become fabrication, materials, tooling, quality assurance, and supply chain. That's where value will accrue. The best producers will lean into these sources of strength to build moats.

The deeper shift is that the design layer starts obeying software economics: zero marginal cost, forkable, iterable. Once that's true, we will see strategies employed in software being deployed in hardware. A maker might "open source" its designs to capture value elsewhere in the stack. It might modularize the hardware into a platform others build on, in hopes of deepening lock-in. Or it might stop selling machines at all and start renting them, providing users with "subscription" hardware and perhaps even periodic upgrades to newer versions.

I think that last one is the most interesting. Software-as-a-Service started as a defense to the ease of copying enabled by the Internet, but resulted in incredible businesses that delivered more value for their customers at better prices. The same could happen here. Expensive equipment needs expensive, careful maintenance. When sold as a service, the risk of downtime shifts from the customer to the maker, who may be better placed to manage it. A maker selling machines may gold-plate each one to address customer fears about failure, driving up cost. A maker selling uptime may choose to provide spares or over-provision capacity to prevent downtime at lower cost. Suddenly, a defensive change in the business model ends up aligning the maker with the customer.

You might object that makers will simply patent their way out of this. They'll try, most likely, but I don't think it works. Patents may protect against someone producing an identical copy, but they can't stop knowledge from spreading. If anything, patents that work as intended speed the spread up by disclosing otherwise withheld information. It's the spread of knowledge, not the raw ability to clone a machine, that is driving all of the consequences above.

And the consequences may not only be found in the changing behavior of competitors, but of customers as well. When forking a machine is cheap, serving ever-smaller niches becomes sustainable. If an off-the-shelf machine only half-fits a use case, the customer might fork it. The advantage has shifted from people who are experts at making the machine to people who are experts in the demand for whatever it makes. This could make every customer a potential competitor. Someone may set out to solve their own narrow problem, perhaps a tweak for one odd use case, and get pulled up the stack into building hardware for everyone else who shares it.

Readable machines are coming, and soon. The best design is becoming something anyone can read off the machine in an afternoon. That changes how machines get made, who makes them, and how they get used. I'm excited to see what happens.

© 2026 by Fully Doxxed. All rights reserved.