February 17th, 2025
Imagine you have a box with a microphone. You speak into the microphone, describing an object you desire. The box works for a while, and eventually materializes it for you. You've just spoken an object into existence. Literally. We're about to cross this wild line in reality.
This is already true with software. You can describe a product to ChatGPT or Replit, and it will do a decent job at generating it for you. It has a long ways to come, but it's already functional. Companies are hiring less junior engineers already.
How would generating physical objects actually work?
Well, all that's really needed is a CAD file. Once you have a CAD file, it's just a matter of figuring out the best way to produce the object. Is it a metal piece that needs CNC milling or casting? A plastic piece that needs to be 3D printed? A wooden piece that needs to be carved and polished?
We’ve been translating mechanical drawings into physical objects for centuries, evolving from 2D technical sketches to modern parametric CAD tools over the last few decades. We’re experts at production; the real challenge is generating accurate, fully parameterized CAD designs from scratch.
How close are we already?
Generating physical objects is simply the next evolution of digital ones. How far down this evolutionary path have we gotten already?
Text-to-Text
The AI revolution started out humbly by effectively re-writing text. You’d take a conversation transcript, pop it into an AI tool, and get a nicely summarized version. Or you’d have a rough concept for an article and, with a few prompts, AI would re-organize your messy first draft into something more structured (albeit poorly written). Pretty neat, but only the beginning.
Over time, text-to-text functions became more advanced. It's now better at writing and thinking than most junior employees.
Text-to-Code
Then we started to see text-to-code. Instead of manually writing lines of code, you can just describe what you want. AI can actually build a working digital product from that description. You are effectively writing software into existence.
At first you needed to know how to code to work with this. It was an accelerant for software engineers. It's now quickly becoming a revolutionary no-code tool. Text-to-code gives you a new skillset. It turns a non-technical person into a software developer.
Speech-to-Text
Now, combine those: speech-to-text meets text-to-code. You can literally talk your way through software creation. No need to even type. With your voice alone, you can get from an idea (“Let’s build an app that tracks how many glasses of water I drink”) to a functional piece of software.
This has been dubbed "vibe coding", and love it or hate it, it's here. Right now it's still shoddy, but rapidly improving. Vibe coders are nowhere near the level of experienced developers (who are also using AI btw), but the gap is closing faster and faster.
Speech/Voice-to-Image
Generating images from text prompts has also been around for a little while now, but only recently has it gotten really good. And it needs to get really, really good before AI CAD tools are a viable solution. But it's the obvious stepping stone.
Lets say you want to generate a custom stick shift knob for your manual car. You can already generate a pretty realistic version just on ChatGPT.
This is already quite close to a CAD file. It basically just needs to be stripped down into a blueprint CAD file format with exact measurements.
Bridging the Text-to-CAD Gaps
We’re crossing this line as we speak. Several early-stage companies, like zoo.dev, are working on generative CAD from natural language prompts. They're still in the early phases and not everyone is a believer. Predictably, many mechanical engineers do not think text-to-CAD is possible.
We’re already seeing early glimpses of generative design in software like Autodesk Fusion 360 or nTopology, which can optimize geometry for weight reduction and strength. Text-to-CAD is the next logical step—just in a more intuitive, language-driven way.
5-10 years ago, it also seemed impossible that AI would be replacing software engineers and artists. Both of those industries also seemed irreplaceable. We learned they were actually the first to go.
There are some manufacturing obstacles as well. Once you have a CAD file, you can go straight to 3D printing or even automated CNC milling. These machines do the heavy lifting, but humans still step in for final checks, finishing touches, or specialized tasks.
Some of these can be automated with existing technologies. Other niche tasks may need to wait for humanoid robots that can replicate the customized actions needed from a human.
For instance in 3D printing, part orientation drastically affects strength and surface finish. A robust AI solution would account for layer directions, support structures, and post-processing needs. These things a seasoned engineer or machinist would do by habit. Soon they won't need to at all.
How Close Are We?
Text-to-CAD Is Here (or Almost)
If you talk to mechanical engineers or industrial designers, many will say text-to-CAD flat out isn’t possible on a professional level. But don't forget, two years ago most AI experts would have laughed at the idea of writing entire software products with a few sentences of instructions. Now it’s old news.
There is also an instinct from within the industry to reject something as powerful as AI. Designers don't want to believe a computer can out-design them. Writers don't want to believe their craft is not uniquely human. It's natural to reject an existential threat (unless you're a SWE apparently), but that doesn't mean it's not coming.
Text-to-CAD has major challenges—like factoring in tolerances, mechanical constraints, and material properties (e.g., yield strength or thermal expansion). These aspects are critical for real-world reliability, especially when safety factors are involved.
But these are the exact kinds of problems that modern AI seems to be getting better and better at. Give it a year or two, and we’ll likely see a tool that can crank out decently accurate CAD designs from sketches, photos, or even just a spoken prompt.
We're likely to see a quality jump like we have with video generation.
Give it a year or two, and we’ll likely see a tool that not only cranks out CAD designs from sketches or spoken prompts, but also runs quick FEA or stress simulations in the background to ensure structural integrity.”
Production and the Human Factor
The next step is production. Even with advanced 3D printing and CNC, humans will remain integral for:
- Quality Control: Checking that the final product meets safety and design specs.
- Material Selection and Procurement: AI can help, but real-world availability and cost still require human judgment.
- Aesthetic Touches: Sometimes you want that artisanal flair or unique finishing technique.
- Complex Assemblies: For multi-part products, humans may still oversee the puzzle pieces coming together—at least for now.
But we’re quickly marching toward production lines that handle most tasks independently. Automated production was well underway long before the AI revolution. AI could organize designs, schedule processes, and dispatch robots to pick, place, weld, or finish.
What Does a World Look Like Where Objects Can Be Spoken into Existence?
We already have a pretty clear idea of the disruption generative CAD will cause, because we are currently watching it happen in the software industry.
- Democratized Manufacturing
- Costs Plummet: The barriers to making physical products drop dramatically. You won’t need a massive factory or big upfront capital.
- End of Mass Production as We Know It: Why settle for off-the-shelf when custom is just as cheap and easy? Economies of scale become economies of one. That means personalized everything—from your coffee mug to your car chassis.
- Hyper-Customization
- Picture every single component in your house—hinges, drawer knobs, furniture joints—custom-made to match your style or ergonomic preferences.
- Building your own car with parts that are designed to your exact specs? It’s a lot easier to imagine when you just have to say what you want and let the AI handle the design.
- Inventiveness Explodes
- AI-driven design can find shapes and solutions we might never conceive of on our own. It will simulate new structural forms, optimize for weight and strength, and come up with materials or methods that defy our traditional approach to manufacturing.
- We’ll see new methods of assembly, new ways to integrate technology into everyday objects, and brand-new products that no one has dreamed up yet.
There will be many growing pains. This wave of disruption will be uncomfortable for many established players. Entire industries that rely on large production runs might need to reinvent themselves. I'm guessing that as soon as text-to-CAD becomes viable, we will see extreme pushback from manufacturing industries. Intellectual property laws might also get murky (“Who designed this product, me or the AI?”). Similarly we'll need to figure out quality, safety, and regulatory hurdles.
Watch this space. I’m personally excited to see how this unfolds and plan to dive deep into these new AI-driven CAD tools. The future is coming fast and you might be able to just ask for it by name. We're on the cusp of an Indie Hardware Revolution. What will you build if it succeeds?