I've seen dozens of studies arguing that
#AI makes them more or less
#productive.
My
#theory is that a
#team where one person uses AI, while the person may (seem to) be more productive than without AI, the team as a whole is less productive. In fact, I think productivity goes down as the number of AI users goes up. Does anybody have concrete (non-anecdotal and unbiased) numbers for or against this?
To me, AI is that person browsing the H̵u̵b̵ Tube/news all day at work, but has a spreadsheet ready in case somebody walks by. They give the semblance of producing work and might even turn in some barely thought-thru
#slop that needs heavy editing, but in reality force team-members to pick up the slack and fix their non-work.
My rationale is code reviews. I hate writing code reviews, but do them to improve our shared code and help the other (often junior) person improve. If I get code to review from a person, I do not check it in detail: I assume there is rationale behind it – even ask people to write down the rationale – and check the idea rather than every character of each line. I skim the tests and usually trust they can write documentation and frontends that match.
When I get a sloptacular commit straight from the s̵e̵w̵e̵r̵
#LLM, I need to check it much more thoroughly. I cannot trust that the documentation is true, because LLMs do not describe the idea or code (there is no idea and the code is mish-mash of security holes from StackOverflow), it writes what looks like documentation. Sometimes it's correct, sometimes it's not and bad documentation is worse than no documentation. But if I don't read it in excruciating detail, I won't know when it's either. After that, I need to do the same with the code. It'll look fine, but there's no underlying idea to check, and even if there were, the code is unlikely to reflect that. The same for the tests.
So, the slop-cowboy may have produced 1000 lines of code in half the time, but if it takes me 3 times as long to review, we're not
#succeeding as a team. It's basically
#Brandolini's law: it takes longer to review ultra-productive slop than it does to "write" it. Not only does that turn a junior job into a senior job, it makes that senior job much less enjoyable by amplifying the bad parts while – according to my theory – reducing productivity.