4 ms·
I suspect that with a little iterative modelling of the whole system with derivatives, efficiency could be bought from ~40% to ~80%, giving you an extra few hun
by londons_explore 2mo ago
I suspect that with a little iterative modelling of the whole system with derivatives, efficiency could be bought from ~40% to ~80%, giving you an extra few hundred mph...
- repeekad 2mo agoI wonder if Claude has guardrails against hyper optimizing a trebuchet projectile into a proper bullet
- ezst 2mo agoHow would you even get Claude involved in that? This isn't your git repo that a language model can scrape for textual content.
- bonoboTP 2mo agoIt could make simulations and read the CNC templates and the files for the 3D printed parts.
- ezst 2mo agoAnd then do what? Do you have any evidence that Claud would do anything relevant with those?
- bonoboTP 2mo agoIt could come up with design adjustments and measure their impact in the simulation. But the sim2real gap may be too large.
- ezst 2mo agothere might be several gaps before that, even. From the top of my head: - 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.
- bonoboTP 2mo agoNext year some LLM will do something like that, and y'all gonna claim it's still unimpressive because reasons.
- ezst 2mo agoYou can probably keep this kind of unconstructive comments to yourself. Even if this was a place to discuss one's beliefs, nobody was attacking yours to begin with.
- fragmede 2mo agoI mean I can link a conversation with Claude that says it knows about materials and various details on their various strength but I don't think that'll convince you of anything.
- ezst 2mo agoIndeed, it won't: I can link a conversation with XYZ-bot that tells me that I'm handsome and my ideas are brilliant and I'm en route to tremendous success. The point is that LLMs are, by design and current models norms, sycophantic and incapable of describing their own knowledge and workings/processes. This is not a matter of "my opinion vs yours", that's what few minutes spent legitimately trying to understand how this works will tell you.
- londons_explore 2mo agoI have had Claude run fluid simulations for me to check machines work. Seems to manage fine.
- bonoboTP 2mo agoMany people are stuck in 2023.
- MomsAVoxell 2mo agoConvert the model to OpenSCAD, put it in a repo, and off you go. OpenSCAD is very LLM-friendly.
- ezst 2mo agoA LLM may produce OpenSCAD compliant syntax, I don't doubt that, but how do you make it produce designs that fit engineering constraints? How do you even express those to a LLM? How do you validate those? How do you iterate upon them? 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.
- MomsAVoxell 2mo agoI’m pretty sure Claude has been trained on high school physics textbooks, yo. And if it hasn’t there are plenty of other models out there that have. 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.
- ezst 2mo agoAre you familiar with the problem space at all? > 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"
- voidUpdate 2mo agoI'm pretty sure the only reason the video trebuchet wasn't a proper bullet is because it's made out of plastic instead of a ball bearing. Aren't most handguns subsonic?
- malfist 2mo agoThis is nonsense. You're going to double the efficiency by using "derivatives". As if the expert in this video didn't do any modeling themselves (shown in the video) and wouldn't have considered acceleration at all. I say, put your money where your mouth is and prove you can double the efficiency.
- londons_explore 2mo agoAll the energy is lost to air resistance and left in the machine in the form of parts still moving when the projectile detaches. Both of those can be modelled - rather precisely if you have the compute power. By modelling the derivative of all of those things with respect to every parameter of a parametric design, losses can be minimised with gradient descent and a huge Jacobian matrix. A few parameters are obvious (string length, arm length, etc), but there is no need to keep the set small - you could model the exact shape of the projectile latch as a spline with 100 parameters for example, and let gradient descent find the best shape to minimize losses. AI can set up and run all those simulations, and all you have to do is pay the AWS bill for all the open foam instances you'll be running.
- chakintosh 2mo agohttps://www.thebrighterside.news/post/spinlaunch-looks-to-slingshot-satellites-into-space/ https://www.thebrighterside.news/post/spinlaunch-looks-to-sl...