Its also possible to run these things locally… its slower than a cloud LLM service, but its still very usefully fast, and isn’t funding some corpo datacenter.

But it wasn’t done locally, which begs the question: can local models reverse engineer on any kind of reasonable time scale?