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.