New Research On CoPilot And Code Quality

So if you don't want to spend 40mn, the paper suggests that AI makes developers more productive—46% more code written 55% faster. But it also increases code duplication, churn, and lowers overall quality, leading to more tech debt and production bugs.

"Productivity" can only be truly measured if "negative productivity" is also taken into account. That means stuff like "writing bugs that take a long time to hunt down and fix" or "bugs that end up never getting fixed." It can also mean "poor design decisions that screw your whole project over in the long run." My hypothesis is that AI use massively increases "negative productivity," making it a big win for good developers' job security in the medium term.

Article should be linked in the description. Make it easier to find his work if you are going to cover it.

There was a study released in the past week by Microsoft that showed the use of AI is causing people to become worse at critical thinking, as if it sounds "reasonable" then people will just accept the AI output, without actually thinking about it. Goes directly to this discussion about duplicate code accelerating at a huge rate.

This research basically confirms my own experience coding with AI. It makes me SUPER fast when starting something new or something I might not be super knowledgeable about. When there are inevitably problems with this code because it wasn't super well thought out, I have to personally go in and refactor things. If you attempt to AI code your AI code you will literally make whatever problem you have worse, with some exceptions of course. Usually I think it's best to use AI to help you get started, once things are hacked together understand what's happening and dig in and really make it your OWN code. Without using AI. You'll be MUCH happier with the end product and won't feel as much like a prompt engineer!

Bringing AI slop articles to the code of major open source projects !! Amazing !!

Some years back, a certain retail biz decided to measure productivity as "item count at checkout per hour". So, quite obviously, the cashiers stopped using the multiplier keys and started scanning all items one by one. You can guess how well it went, how long the lines became, etc... So now it's "commits per day" or whatever. With same results.

I already have problems with two developers (sometimes even the same developer), solving the same problem differently within a project. And of course copy/paste is common. I've even seen copy paste of entire solutions where only a small portion of the code was actually needed. And all of this is without an AI assistant!! I didn't realize it could get this much worse!

So I have a theory. A) the current AI coding workflows are not properly incorporating a test-driven development approach or test-gated commit acceptance process. B) AI coding workflows require an explicit cycle of identifying duplication and refactoring that away - and the on-commit workshops aren't making this a blocking linting error.

I love the hard data confirming what the wisdom of experience engineers already knew and have been barking about for 4 years.