This is cool, but still is affected by the same local vs global scope problem that all bottom-up algorithms suffer. The NN will piece together small snippets (l

This is cool, but still is affected by the same local vs global scope problem that all bottom-up algorithms suffer. The NN will piece together small snippets (local scope) to produce large snippets and then larger snippets and so on. The problem occurs when it can't understand what it has created (global scope), and therefore can't streamline the code in obvious ways: 1. combining loops. 2. removing repeated logic. 3. understanding memory problems (i.e. when to use a list vs a lazy list). 4. finding performance hotspots and understanding how to fix them. 5. non-maintainable output likely, and seeing as humans need to edit the final code, this becomes a big problem. Those last three points are why this stuff is fine for contests, but no one is going to want to use it in production. Not even getting started on how do you define a complex algorithm without getting into a ton of ambiguous cases.