GitHub Copilot's telemetry does collect data on whether generated code snippets end up staying in the code, so presumably models are tuned on this feedback. But you haven't solved any of the problems set out by Karpathy here—this is just bankshot RLHF.