The domain of ML-driven design optimization isn't exactly new, is quite specific, and I would need convincing that Claude has anything to contribute to it.
It’s not a problem of the ML model. It’s a problem of the human setting up the description of the problem in such a way that the model can operate. OpenSCAD gets a long way towards that target. Use it with a model thats been trained for the purpose - just the same way that ML has been used to produce optimal rocket engine nozzles and fuel transfer systems.
Not that difficult, really.
> Claude has been trained on high school physics textbooks, yo
The very fact that physics textbooks have little bearing in the real world is the whole damn reason why Mechanical Engineering exists as a separate discipline, yo
> It’s a problem of the human setting up the description of the problem in such a way that the model can operate.
It's a problem of defining the initial state (which is the trivial part that OpenSCAD may be a contributing element of), defining constraints and variables (what is allowed to be changed and not, for what can, in which ways, to what extent, i.e. what is the library of allowed material, fastening, machining, assembling techniques available to your very specific situation), defining evaluation and fitting criteria.
I see very little adequacy of general purpose LLMs in that
> the same way that ML has been used to produce optimal rocket engine nozzles and fuel transfer systems.
Which has nothing to do with "just use Claude, yo"
- text description to 3D geometry gap: you can't easily describe complex geometry in a language that's common and convenient to both humans and machines
- 3D geometry to physical model gap: material , geometrical constraints induced by manufacturing (machines, tools, costs, …)
- physical model to model fit for simulation gap: meshing, constraints, stress modelling, …
- simulation outcome to fitness assessment gap: now you have a high-dimensional and numerically heavy simulation to weigh against a non-rigorously defined acceptance criteria
- simulation outcome to geometry profiling gap: how do you even start to guide the LLM into the vast space of possible changes to apply to the 3D geometry and reboot that loop?
LLMs don't strike me as a particularly relevant technique to apply here, to be honest.