GitHub just released a report about their ongoing experimentation with Copilot over a million users. Interestingly enough, it highlights the increase in acceptance rate over time. But I'd be curious to distinguish pure performance increase of Copilot as an API vs. learning curve of users to prompt it correctly.