Has the decline of knowledge work begun?

> A recent academic paper found that software developers who used an A.I. coding assistant improved a key measure of productivity by more than 25 percent, and that the productivity gains appeared to be largest among the least experienced developers.

I dunno about this citation. I just read the paper and it considers "productivity" in this context to be the number of builds, commits, and pull requests in a developer week. Interestingly, there was no statistically significant difference in build success rate between those who used the AI tool (Copilot in this case) and those who didn't.

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4945566

Personally, I've seen junior developers most negatively impacted by AI. It seems to stymie learning and reasoning skills.

There is definitely an effect of ”you forget what you don’t need to remember” happening even to me.

In my experience, the best developers out there have all worked through hard problems by themselves. "Been through the trenches", so to speak. That means getting right to the bottom of difficult bugs etc., no matter how long it takes.

On the other hand, the worst developers reach out for help the moment things start to look unfamiliar to them. Over the years I've tried to encourage such developers to think for themselves, but I've come to realise some just can't do it if they know someone else on the team probably has the answers.

Availability of LLMs seems to be the worst thing for this kind of developer. Now they don't even have any kind of social barrier to reaching out for help. I just don't see how they're ever going to learn to do anything.

Code monkeys have always existed, and can be useful, but why wouldn't the engineers just go straight to the LLM instead of going through code monkeys?

There is a recent paper by Microsoft that describes exactly that. (Although they didn't restrict the negative impact to junior developers)

https://www.microsoft.com/en-us/research/wp-content/uploads/...