Key Takeaways

  • AI is at the "very, very beginning" of impacting CAD, lacking the foundational understanding needed for complex engineering design.
  • Current large language models (LLMs) and video models cannot grasp physical properties like friction, weight, or surface texture, making them unsuitable for intricate hardware tasks.
  • The most significant obstacle to AI-driven hardware design is securing access to proprietary CAD data, which companies guard as their most valuable intellectual property.
  • Early advancements in AI for CAD are more probable through hobbyists or on-premise AI training systems than from large, established incumbents.
  • Caitlin Kalinowski envisions a future where AI can generate complex 3D CAD from simple inputs, leading to a world where "robots building robots" becomes a reality.

AI's Blind Spot: No Sense of Gravity, Friction, or Form

Many founders are buzzing about AI designing products, but Caitlin Kalinowski, a hardware leader from Apple, Meta, and OpenAI, brings a dose of reality. She argues that AI's impact on Computer-Aided Design (CAD) is still in its infancy. Why? Because the core AI models we have today—the LLMs and video models everyone talks about—are missing something fundamental: a sense of the physical world.

“These LLMs and even video models, they don’t know how to do that,” Kalinowski explains, referring to the nuances of physical engineering. “They don’t have the ability to understand friction or weight or contact, uh, pressure, uh, friction, surface texture, like they’re just not able to do these things.” This isn't a minor detail; it’s the entire ballgame for hardware. Designing a functional component requires a deep, intuitive grasp of how materials interact under stress, heat, and motion. Without that, AI is essentially guessing words, not engineering parts.

Your CAD Data is Gold – And That's The Problem

Even if AI developed a sense of physics, there’s a giant wall in its way: data. High-quality CAD data is critical for training any AI that hopes to design. But this data is highly proprietary, locked away behind corporate firewalls. Kalinowski points out the tension directly: “The biggest challenge here, Lenny, is actually the data. This CAD data is some of the most valuable IP that anybody has… Where is this data going to come from is a big question I have.”

This makes sense. If you’ve spent millions developing a unique engine component or a specific circuit board layout, you’re not likely to hand over the raw CAD files to an AI company to train its general model. This proprietary wall means that the companies most capable of benefiting from AI design are also the ones least likely to provide the fuel for it. Kalinowski suggests that early breakthroughs might come from less restricted environments. She sees hobbyists, unburdened by corporate IP concerns, or small, focused on-premise AI training systems as potential initial adopters for creating complex 3D CAD from simple inputs.

The "Robots Building Robots" Dream is Real, Just Not Yet

Despite the current limitations, Kalinowski paints a compelling picture of a future where AI fundamentally changes hardware design. Imagine going from a simple 2D sketch to a fully functional 3D CAD model, complete with assembly instructions, vendor communication, and iterative feedback—all driven by AI. “The idea that you could even as a hobbyist go from a 2D picture to complex 3D CAD to assemblies to communication with vendors of how to make those parts and getting their feedback to iterating on that and doing a couple builds like that is possible I think in the future,” she says.

This vision of “robots building robots” isn’t science fiction, but it hinges on solving the two major hurdles Kalinowski outlined: giving AI a profound understanding of physical laws and finding a workable solution for accessing or generating the vast amounts of proprietary CAD data needed for training. Until then, the promise of AI for truly generative hardware design remains a significant opportunity for those willing to tackle these hard, foundational problems.

What to Do With This

If you're a founder in hardware, understand that generative AI isn't replacing your core design team next year. Instead, invest heavily in protecting your CAD intellectual property; it remains your core competitive advantage. If you're building AI, look past LLM applications and focus on the fundamental physics-modeling gap. There's a greenfield opportunity in building AI that understands friction, weight, and surface texture, rather than just guessing words.