Three frontier releases landed in the same week: GPT-6 Astra, Claude Fable 5.1 and the World Labs Atlas. AI has found its way into the world of modelling geometries.
The line that stuck with me is that BenchCAD checks whether the shape occupies the right space, not whether the model understood why the part is shaped that way — and your Fable demos show the verifier earning its keep (the garage roof pitch, the dangling schedule activities). From the project-review side, the error class I worry about is the one BIM checks are structurally blind to: omission. Clash detection, IFC schema validation and rule sets all test what is in the model; a missing sleeve, an undetailed connection or a drainage run nobody modeled passes every one of them, and that is exactly where a "looks complete" AI model will be most confident. Do you see a practical way to spec "what must exist" per system so the verifier can flag absent elements, or does that part stay with the engineer's responsibility for now?
The line that stuck with me is that BenchCAD checks whether the shape occupies the right space, not whether the model understood why the part is shaped that way — and your Fable demos show the verifier earning its keep (the garage roof pitch, the dangling schedule activities). From the project-review side, the error class I worry about is the one BIM checks are structurally blind to: omission. Clash detection, IFC schema validation and rule sets all test what is in the model; a missing sleeve, an undetailed connection or a drainage run nobody modeled passes every one of them, and that is exactly where a "looks complete" AI model will be most confident. Do you see a practical way to spec "what must exist" per system so the verifier can flag absent elements, or does that part stay with the engineer's responsibility for now?
Thanks G, I think more and more our job will be, only, to make verifiers.