Newcomers Who Bet $9 Billion on AI for MCAD, Part 2

Opinion by Ralph Grabowski

In the past year, several companies announced plans to create an AI system that models reality with ‘World Model AI’ using physics, 3D spaces, and so on.

World Model AI In MCAD

Old MCAD systems boast of AI to stay relevant, while brand-new ones use it to get noticed and be funded. Vendors like Dassault Systemes and Autodesk state they’ve been working on AI for over a decade already, but their reference point is generative design, which a decade ago they weren’t calling ‘AI’.

A common AI scenario in today’s CAD systems is asking a built-in chatbot to render a scene with specific parameters, or to select all lines on layer “Wall-External”; this kind of AI is consists of shortcuts to tracking down the commands needed, and is little different than a search engine. Sometimes, there is no solution, because the chatbot hasn’t been programmed with enough answers yet [source], a sign that we might not be dealing with artificial intelligence after all.

Autodesk says it has lots of AI in its Fusion design package, but further reading reveals it tends to be, in my opinion, script-like automations to replace many steps with one, such as automating tool paths, automating drawings, and automating fastener replacements [source].

Marble generating a room image from World Labs

To get onto the World Model AI bandwagon, Autodesk last month invested $200 million in World Labs, whose Marble software creates editable 3D environments. One approach is this: Start “With a world-model-based sketch in World Labs (say, of an office layout) and then drill down on certain design aspects (like the design of the desk), which is where Autodesk’s tech might come in,” said Autodesk chief scientist Daron Green [source].

A new firm launching with World Model AI is AMI Labs, short for Advanced Machine Intelligence. It wants to provide industrial process control, automation, wearable devices, robotics, and healthcare software. The very sparse Web site carries the tagline, ‘Real World. Real Intelligence, because “real-world data is continuous, high-dimensional, and noisy” [source]. As it is based in Paris, might Dassault Systemes be a possible future investor?

“Wait a minute!” we might imagine Archetype saying. “We were first.” It is coding the entire physical world through its Newton foundation model that reads data from motion, power, and electrical sensors, and then generating human-like interpretations. The company landed $48 million in funding since 2024. www.archetypeai.io.

The one firm that to me appears to be the furthest along the AI route is an architectural one, Snaptrude. When you enter a one-line prompt, such as “Design a 3-story office building with open workspace, meeting rooms, and cafeteria,” it spends ten minutes or so generating and optimizing an initial room layout that meets specs of a building code — cheerfully chattering to itself as it carries out the determinations.

Initial rooms layout generated by an AI prompt in Snaptrude

What Ralph Grabowski Thinks

The closest MCAD has come to World Model AI so far is in digital twins, where the CAD model is a replica of what will be, and has been, built. It lets designers run simulations and optimizations on products like expresso machines and airplanes before building them, and then monitor their operations once they are built. Digital twins are still manmade.

Adding AI could help identify problems designers might not have noticed [source]. The problem for us is that LLM-based AI has a 30% failure rate by returning hallucinations — making up information due to its stochastic nature [source]. Worse, a single agent could cost plus-or-minus $100,000 a year to run [source], as estimated by organizations like Riseup Labs.

Running more than one agent also suffers. “Because individual agents cannot independently verify truth, one agent’s confident hallucination becomes the factual foundation for the next agent’s output,” says Gerard Sans [source]. “Errors do not cancel out. They compound.”

As for World Model AI, it is too new to know if it will work at what we call LoD 5 (the highest ‘level of detail’). Given the complexity of Revit modeling a single building, I suspect the complexity of our world is not particularly modellable. Gerard Sans speculates World Model AI could work in areas “Where a domain has high repetition, structured representation, consistent terminology, and deep public documentation.” Most do not.

Just like it is concerning that unexperienced engineers could access simulation software early in the design process, so too we ought to apply caution with blackbox AI assisting designers [source].

---

[This material first appeared in Design Engineering magazine, and is reproduced with permission. It has been updated.]