Eh, well, if your guns are trained on the "copyright" portion of the ship and you can sink it from there, no need to waste ammo or time trying to figure out if code bits are as explosive as the copyright bits are. Probably the code is just as sinkable, e.g. here's a recent response to some other AI slop:
I didn't look closely at most of the code but one thing that caught my eye, pid is not safe for tempfile name generation, another user of the system can easily generate files that conflict with this. Functions like mktemp and mkstemp are there for a reason. Some of the other "safety" checks make no sense. If the LLM code generator is coming up with things which any competent unix sysadmin (let alone programmer) can tell are obviously wrong, it doesn't bode well for the rest.
To me, SOTA is just bad at DRY, KISS, succint, well architected, top down, easy to test code and has to be constantly steered to come close. Even the article suggests that. YMMV.
TDD makes the code test-passable, but it is still rng. As for goals, you can't foresee every stupid thing it will generate. It will look at a state machine, and rather than using the existing event structure, write its own loops and conditions. This is very different compared to human devs. No goal will help. You just keep yanking its chain until it generates as described. It can't even put imports at the top as you described. It can't help making circular refs in c++ despite being specifically told to use a hierarchical structure. Left alone you will get truly unstructured random mess.
People keep making trivial apps with open source examples thinking they found god. Another dismissive comment and I swear.
The next AI winter can't come soon enough…