Ben Okonkwo: Marcus, hey — I spent part of this week trying to explain to someone why 3D modeling takes as long as it does, and I genuinely struggled. Because the honest answer is: a lot of the time is skill acquisition, not thinking. And then yesterday Formas just... removes that.
Marcus Vale: Cartesian. September 15 launch.
Ben Okonkwo: Cartesian, yeah — 'Anything to 3D… structured, editable 3D,' that's the actual framing from the @formas_ai announcement. And Carlos Banon, co-founder, SUTD professor, had posted a demo the day before that pulled 296 likes. The tagline was 'Powered by ASTRA. Built for the real world.' Now here's what I want to sit with — a 2025 CHI study found that workers who grew confident in generative AI reported measurably less critical-thinking effort. And my instinct is that the thinking that used to happen inside the hard parts of modeling was... load-bearing somehow.
Marcus Vale: Load-bearing for what, though? Because — look, the Architizer and Chaos survey: 1,227 architecture professionals, 46% already using AI in their projects. If the thinking that lived inside manual modeling was essential, you'd expect that number to come with some professional crisis signal. I'm not seeing that reported anywhere.
Ben Okonkwo: That's — okay, that's a fair challenge. Though I'd note that 46% adoption doesn't tell us what they're using it for. Visualization is very different from production geometry.
Marcus Vale: Which is exactly the Cartesian claim — that it closes that gap. That's the bet.
Ben Okonkwo: Right — but that gap-closing is the part I want to actually understand mechanically. Like, what is Cartesian literally doing when it takes your sketch?
Marcus Vale: Okay, picture this. Structural engineer, three in the afternoon, sketches a curved wall section on her tablet — rough, not precise — types 'glass curtain wall, twelve meters, slight inward lean' — and Cartesian hands back a file she opens in Rhino and starts editing. Not a render of a building. An actual geometric object she can pull vertices on. That's the promise.
Ben Okonkwo: And the thing that makes that different from, say, a nice-looking image of a building is — what exactly? I mean, I think I know but I want to hear you say it.
Marcus Vale: Topology and surface type. A generative render is basically a photograph — looks right, structurally meaningless. What Formas calls 'clean topology and native surfaces' means the geometry underneath has actual mathematical structure — not approximated polygon meshes, real surfaces you can modify without the whole thing collapsing. That's what editable CAD means in a professional workflow. Formas and Yiping Goh are making that claim explicitly.
Ben Okonkwo: Hm. And ASTRA — OpenAI's model — is doing that conversion. From sketch to that precise output.
Marcus Vale: ASTRA is doing it, yeah — Formas owns the interface layer, not the substrate. Which is worth flagging, but the deeper issue right now is actually simpler: no independent practitioner has published a workflow test. Nobody's said 'I imported Cartesian output and here's the vertex count, here's where it broke.' The precision claim is — I mean, it's marketing. Asserted, not demonstrated. And the thing that makes this consequential later is that even if the geometry holds up, there are two other blockers — export compatibility and who actually owns the IP on AI-generated output — that we haven't touched yet.
Ben Okonkwo: So the honest answer today is: plausible mechanism, zero independent verification.
Marcus Vale: And those two blockers aren't hypothetical — they're already on the record. The replies to Banon's September 14 demo, on X, people asked directly: what export formats does Cartesian support, and who owns commercial rights to the geometry it generates? Formas hasn't publicly answered either question. That's not a PR miss, that's a product reality.
Ben Okonkwo: Okay — that's the specific gap I want to make concrete. Say a mid-size firm, Friday afternoon, deadline slipping. Someone uses Cartesian to generate a structural massing model. Geometry looks right. But before she hands it to the structural engineer, she needs two things: can it actually export into the Revit pipeline her firm runs, and does her firm — not Formas, not OpenAI — own what just got generated? Neither question has a public answer.
Marcus Vale: And the selective preview gate doesn't help. September 18 start, prioritizing Founding Circle and Studio members — the gap between announcement and anyone actually using it is three days minimum, probably longer for most architects.
Ben Okonkwo: Right — but is that gate quality control or is it... I mean, is it that the precision claims haven't actually been validated at scale yet?
Marcus Vale: Probably both? The scarcity play is real — Founding Circle members first, that's a deliberate signal of demand. But the honest read is: nobody outside that circle has stress-tested the export pipeline. The AEC software market is supposedly hitting thirteen to fifteen billion by 2030. That's the prize. You don't walk away from that over mesh artifacts — but you also don't win it if your output breaks the moment it hits a real Revit workflow.
Ben Okonkwo: And here's what actually worries me about the IP question — Deloitte found 31% of workers are already using GenAI without their employer knowing. 17% paying out of pocket. So the bottom-up pressure is already there in design firms, people are already doing it quietly, and now the licensing question lands on them individually, not on the company.
Marcus Vale: Which means the liability travels with the individual practitioner until someone — Formas, a firm's legal team, someone — actually publishes terms. That's the unresolved thing. The market is ready. The tool is announced. The answers that would make it billable aren't public.
Ben Okonkwo: And I think that's actually — I mean, that's the honest place the whole thing lands, right? Exciting tool, market clearly ready, Architizer numbers confirm that, but 'ready' and 'proven' aren't the same word. Formas hasn't shown the receipts on the precision claims, hasn't answered the export questions, hasn't touched licensing publicly.
Marcus Vale: The real question — if Cartesian actually works at production scale — isn't speed. It's whether the architect can tell you why the building is shaped the way it is. Or whether that judgment got quietly outsourced to ASTRA.
Ben Okonkwo: Yeah. That one I actually buy.
Marcus Vale: Good conversation. Genuinely.
