Neural geometry thread explores AI-CAD

István Csanády, the founder and chief executive of CAD software company Shapr3D, used a thread on X on May 24 to argue that the next wave of AI design tools will need to learn geometry and physics, not just images or text. His posts focused on a longstanding problem in computer-aided design: the industry still relies on boundary representation, or B-rep, to define editable solids, even though those models can break under complex edits and simulation-heavy workflows. Csanády tied that argument to recent neural-CAD research and to a separate OpenAI mathematics result on discrete geometry that he linked as evidence that machine reasoning is moving deeper into formal spatial problems. ### Why did a CAD executive make this case now? May 24 was the date of Csanády’s thread, which circulated around a familiar engineering complaint: generative systems can sketch shapes, but production CAD requires geometry that can be edited, constrained and manufactured. Csanády, whose company builds CAD software for iPad, Mac and Windows, has written repeatedly about geometric kernels and 3D modeling algorithms. (shapr3d.com) Shapr3D’s own materials describe B-rep as the mathematical basis for describing geometry in CAD, where a shape is defined by its boundaries, including faces, edges and vertices. That representation became the industry standard through commercial geometry kernels such as Parasolid, ACIS, CGM, Granite and ShapeManager, according to Shapr3D’s technical overview. ### What exactly is the problem with B-rep? Boundary representation is widely used because it is precise and editable, but Shapr3D says it also carries “inherent problems” that show up in CAD models. (shapr3d.com) The company’s technical explainer says B-rep can contain both geometric and topological inconsistencies, which is one reason CAD kernels are hard to build and maintain. Csanády’s argument in the thread matched that critique. (shapr3d.com) He pointed readers toward approaches that are more robust under constraints and better suited to coupling design with simulation, especially when a model needs to survive optimization loops rather than a single forward modeling pass. That position lines up with current research on differentiable CAD and differentiable simulators, where gradients from geometry or physics engines can be used inside optimization systems. ### What does “train on geometry and physics” mean in practice? Recent papers show one route. A March 2026 paper titled “DreamCAD” said existing CAD generation methods are limited by small annotated datasets and B-rep labels, and proposed a framework that produces editable B-reps from point-level supervision using differentiable parametric surfaces and differentiable tessellation. A separate review of differentiable simulators said their value is that they compute gradients of physical processes, allowing them to plug into gradient-based optimization schemes. (arxiv.org) Another architecture paper published in May described “physics-in-the-loop” CAD generation as a way to improve reliability and physical validity in generated designs. Those papers do not endorse Csanády’s thread, but they support the technical direction he was describing. (arxiv.org) ### Why did he link an OpenAI geometry result? OpenAI said on May 20 that one of its internal reasoning models had disproved a longstanding conjecture in the planar unit distance problem, first posed by Paul Erdős in 1946. The company said the proof was checked by external mathematicians and published companion remarks with background on the argument. Tim Gowers, writing in the companion material cited by OpenAI, called the result “a milestone in AI mathematics,” and number theorist Arul Shankar said the paper showed current AI models could have “original ingenious ideas.” Csanády’s link appeared to place CAD in that broader discussion: if models can reason over formal geometry, engineers are asking whether similar methods can be pushed into design systems that must obey geometry and physics at once. (arxiv.org) That connection is an inference from the timing and content of the thread. (openai.com) ### What comes next for AI-CAD work? May 2026 research papers and public CAD writing point to the same near-term test: whether generative systems can output models that remain editable, simulation-ready and valid inside existing engineering workflows. Shapr3D continues to publish technical material on B-rep and direct modeling, while neural-CAD researchers are releasing systems aimed at editable B-reps, longer CAD sequences and physics-coupled generation. (shapr3d.com) (openai.com)