Pretty good post from The Verge talking about where coding models are actually sourcing improvements from... They've hired a lot of humans to feed more code in for the model to regurgitate.
I've been watching this for a bit - and I think people underestimate how much these models are still faking intelligence by just brute force pushing of code in. It also explains why models fall off a cliff - they haven't hired anyone for that use case/API yet.