CJ Handmer details a workflow using Claude to parse and upgrade decade-old academic simulations for modern GPUs

@CJHandmer this is how we figured out zero radar cross section with the skunkworks programs. denys overholser realized mainframes could do the compute at the time

A decent fraction of the parts of academia I'm from would basically make their career by upgrading simulations from ten years before using the latest supercomputers.

Claude and I have been having fun doing this at scale for the last few months.

Incredibly alpha in scraping old papers for "computationally prohibitive" and throwing GPUs at it.

@JaraySzabolcs I feel like the magnetohydrodynamics of the Earth's magnetic field are due for another crack.

@CJHandmer searching arxiv for "computationally prohibitive" sounds like a goldmine honestly

@emm0sh You can do it by hand for the faceted airplanes.

@CJHandmer Interesting point. There must be lots of stuff from the 70s and 80s like that - interesting ideas that lacked the compute to pursue in enough detail/volume.

@CalebJoelFuller Check out soviet journals.

@CJHandmer That's exactly the kind of interesting problem I've been trying to ponder up as a sort of utility benchmark... lemme see if GPT can pull a solid list heh.

@CJHandmer That's what we have been doing with deformation sim, moving from CPU to GPU. Lots of room for speed up.

@CJHandmer Claude scraping old papers for 'computationally prohibitive' is hilarious. I have a stack of propulsion papers from 2015 that would fold instantly under modern GPU load.

@CJHandmer Throwing GPUs at old sims is fun, but shaky assumptions from back then just scale right alongside the compute.

@CJHandmer I think it's cool how both you and @esrtweet are using AI to bring old things up to date.

@CJHandmer Can you share an example?

@CJHandmer Did this to Gunnar Farneback's master's thesis for a term project a month ago

@CJHandmer the fun part for me is changing the numerics or math itself to be optimal for the new architecture

@CJHandmer Mining the past.

This is why the intro/discussion section has a long legacy

@CJHandmer Good idea!

@CJHandmer the real alpha's revisiting old ones with tools that didn't exist yet.

This is one of the most underrated uses of AI nobody's talking about.

Decades of "computationally prohibitive" results just sitting in old papers — effectively frozen in time because no one had the GPUs to finish the job.

Now you can thaw them all in months.

AI isn't just accelerating new research. It's resurrecting everything we couldn't afford to finish.

@CJHandmer What level of trouble would that cause?