I had gpt-5.5 translate microgpt.py into a C++ version recently. I had to steer/convince it to use data oriented design to avoid excessive pointer chasing, but the end result was as expected: Now 500 LOC instead of 199, but speedup was 100x. That speedup is definitely worth doubling the line count. And frankly, modern C++ can read very nicely, even compared to Python.