I think your points on abstraction and deployment are both good. Abstraction in particular seems to be the skill that will be valuable for copiloted programmers. To generalize from the EOQ example, it seems like anything with a Wikipedia page or known reference implementations could be implemented in a variety of programming languages by a LLM, but doing something novel will be more of a stretch. So, the way to do something novel (or, more likely, domain-specific) will be to break it down into well-known pieces that compose into something new.
That’s not too different from the way good programming is done now, but as anyone who’s maintained research code knows, we often don’t live up to that goal. I’m cautiously optimistic that GPT-era models can improve overall code quality while also raising the total amount of code in the world.
Enjoyed the article. Interesting world we are evolving into. I wonder how different a solution would have been produced if you said "write an R program"?
I think your points on abstraction and deployment are both good. Abstraction in particular seems to be the skill that will be valuable for copiloted programmers. To generalize from the EOQ example, it seems like anything with a Wikipedia page or known reference implementations could be implemented in a variety of programming languages by a LLM, but doing something novel will be more of a stretch. So, the way to do something novel (or, more likely, domain-specific) will be to break it down into well-known pieces that compose into something new.
That’s not too different from the way good programming is done now, but as anyone who’s maintained research code knows, we often don’t live up to that goal. I’m cautiously optimistic that GPT-era models can improve overall code quality while also raising the total amount of code in the world.
Enjoyed the article. Interesting world we are evolving into. I wonder how different a solution would have been produced if you said "write an R program"?