I'm less interested in turning programs into transformers and more interested in turning programs into subnetworks within large language models.

Which the blog post brings up as a research direction, but never actually elaborates upon. And the interface between the two is a hard problem.

I'll check out the link though, thanks.

This seems like it has some potential, but is pretty much useless as it is.

Shame there are no weights released - let alone the "compiler" tool they used to actually synthesize computational primitives into model weights. It seems like a "small model" system that's amenable to low budget experiments, and I would love to see what this approach can be pushed towards.

I disagree with the core premise, it's basically the old neurosymbolic garbage restated, but embedding predefined computational primitives into LLMs could have some uses nonetheless.

Have you seen the average person trying to use technology?

I mean, a real average person, in a natural environment. Not in a movie or in stock footage. The real deal.

I have, and, holy shit. I cannot find the words to express just how unsettling it was of an experience. I still haven't fully recovered from it.

The biggest flaw with your logic is the utter lack of it.

If I could rip K-Pop Demon Hunters with a screen capture app to obtain a file I could share with a friend, I still wouldn't do it. Because finding a torrent is simpler and faster. I would get a very similar file, but so much faster, because I didn't have to keep the screen running at x1 for the full duration.

And finding a shady website that has it available is simpler and faster still.

No. The bottleneck isn't "getting the files", it's sharing them.

If you can ask a friend with basic tech know-how to "rip a CD", you can also ask a friend with basic tech know-how and a VPN to "rip a movie".