This video by Michael Lydick titled I’m an Engineer. This ChatGPT Test Scared Me (opens in a new tab) popped up in my youtube feed this morning. He opens it by stating that chats with language models for CAD and 3D printing are where software is two years ago. The comments, by working mechanical and civil engineers, mostly confirm the capabilities. It is now possible to replicate 3D objects with relatively little knowledge of modeling, and as Michael shows, a camera, a measurement tool, a regular chat web interface and a 3D printer are enough to re-create fairly complicated objects. With some feedback.
I don’t think you need no engineering knowledge. There clearly is knowledge and experience he applies in the video to write his prompts and deciding what steps to take. For instance, I wouldn’t normally use words like ’tapering’. Nor do I have a micrometer. Yet. But I am fairly confident I can follow the steps in his video, buy a micrometer, and start re-creating objects at home now.
As a next step, putting that knowledge in a custom interface is also not that hard. Famous last words - the first 80% are probably easy, and that might be good enough for personal use. I can easily imagine a wizard that guides me or a child through the process. Ask some questions about “describe what you want to do”. “take some photos”, the wizard then uses a model to identify strange angles and tapering if the model doesn’t do it yet.
This is likely to impact many jobs. Some work will be less necessary, and having confidence in the output for professional work might be tricky. Specialised commercial software will become even more for niche applications. And to stay relevant will need to have good APIs so they integrate with AI tools.
I guess like in software this can free up space to improve verification and validation. And also this opens up the possibility for countries to re-industrialise. Some things you can only imagine if you have a factory. But if you have no factory, where do you start? 3D printers are now fairly cheap, and if the barrier lowers, more people can start doing this as a hobby, and learn making by doing.
As for myself, I am very much a software person, although I often work with companies that make machines. . Our kids want a 3D printer, but I held off so far, as I have seen with my friends how time consuming it can be to do anything useful with it.
Seeing this video, 3D printing and creating our own objects is now the Adjacent Possible (opens in a new tab). As Michael Lydick mentions, this is exciting and terrifying in equal measure. What a time to be alive!
Further Reading
It looks like, with software two years ago, small local models are not far behind. Training and research on these also has sped up, so the gap is probably closer. This morning my LinkedIn feed deliverd this new research paper by Leon Chlon et al on training small models. These are around 3 to 6 Billion parameters, which is something that can run on a laptop fairly easily, espicially with a GPU, but even without it is manageable as inference improvements keep coming.
Small physical ai models, the github repository is called Areals (opens in a new tab) , as it’s tagline says: “Learn surface geometry. Transfer the representation to pressure prediction.” It was announced in this message on LinkedIn by Leon Chlon (opens in a new tab):
Proud to announce our Physical AI breakthrough! Millions gets spent on excessive simulations to fine-tune physical AI from scratch because modern foundation models learn a poor representation of geometry. Our breakthrough: 86.9% fewer simulations needed when starting with a pretrained geometry model at no loss in accuracy, building on centuries old work by Euler and Lagrange.
On "Making machines" . I try to avoid jargon, but if you want to search further, engineering managers at clients sometimes talk about "capital intensive goods" and "cyber-physical systems".