The story of frontier AI has been a story of tokens. Words, then code, then images decomposed into sequences a transformer can attend over. Each modality unlocks a new domain of reasoning, and each one moves the field measurably closer to general intelligence. But there is a domain the dominant paradigm has barely touched: space.
Engineering is the discipline of reasoning about space under constraints. A wing, a chip, a heat exchanger, a bracket — every real-world artifact is a negotiation between geometry, physics, and manufacturing reality. Today’s models can describe these things in language. They cannot reason through them.
The current wave of “generative design” tools illustrates the gap. They produce plausible-looking shapes that fail under load, ignore manufacturability, or violate constraints their training data never encoded. We call this Gaussian noise generation — outputs that look correct without being correct. Surface-level fluency, no underlying model of how the physical world actually behaves.
CAD is the right wedge for solving this. It is the most structured spatial dataset humans have ever produced — parametric histories, constraints, materials, simulation results, every iteration captured as a graph of decisions. When an engineer edits a model, they are not generating geometry. They are traversing a decision space under physical and economic constraints. That trace is exactly what a spatial foundation model needs to learn from.
This is why we think the path runs through engineering software, not around it. The orchestration layer we’re building today — the agents, the physics-aware latent space, the iterative validation loop — is also the data-generation infrastructure for a Large Spatial Engineering Model. Every workflow we accelerate produces another reasoning trace. Every constraint adjustment teaches the system something language alone cannot.
AGI will not be built purely out of text. It will be built when machines can reason about the same physical, geometric, constraint-laden world that humans operate in. CAD is the shortest path to that capability — and the engineers who use it every day are the teachers we have been overlooking.
If you build CAD, simulation, or engineering tools and this resonates — we’d like to hear from you.