I bet they’re counting code written while someone had an AI plugin installed as “written by AI” and I bet that accounts for almost all of that 30%. On top of that, I’m betting that they made it mandatory to have such a plug in, and the other 70% is just code written before they mandated this.
I wish this shot from The Terminator had the camera showing Sarah Conner’s face instead of Reese’s, because it’d be such an appropriate meme image on multiple levels for when someone makes a misleading claim about some current AI system.
Also, having 1/3 lines with obvious code that can be auto suggested correctly would make sense, but that is hardly code “written by ai” in the way they suggest.
I’d guess a lot of the people writing the code don’t even have it turned on, it’s just installed because management said it had to be, because management wants to be able to tell investors they’re “innovating work flows”.
I am a small sample to confirm that’s exactly the reason in my brother’s company.
And in my company we’re pressured to make X prompts every week to the company’s own ChatGPT wrapper to show we’re being productive. Even our profit shares have a KPO attached to that now. So many people just type “Hello there” every morning to count as another interaction with the AI.
It is exactly the opposite, with simple, predictable auto-complete you didn’t have to worry about that anymore, with LLMs you always have to look at it in detail because every little thing could be just plain completely different and wrong.
I can read way faster than I can type though. You still check it, but it’s pretty good as that kind of stuff once you have an example for it to follow.
Reading code is usually orders of magnitude slower than writing code. Sure, typing might be slower than reading but to check if it is what you intended you have to understand it too.
Well, I’m generally very anti-LLM but as a library author in Java it has been very helpful to create lots of similar overloads/methods for different types and filling in the corresponding documentation comments. I’ve already done all the thinking and I just need to check that the overload makes the right call or does the same thing that the other ones do – in that particular case, it’s faster. But if I myself don’t know yet how I’m going to do something, I would never trust an AI to tell me.
Well, okay, I can see how it would be useful in languages like Java that are extremely verbose and have a low expressiveness. Writing Java pretty much was already IDEs with code generation 20 years or so ago because nobody wants to write so much boilerplate by hand.
I bet they’re counting code written while someone had an AI plugin installed as “written by AI” and I bet that accounts for almost all of that 30%. On top of that, I’m betting that they made it mandatory to have such a plug in, and the other 70% is just code written before they mandated this.
I would be very surprised if 30% of their code lines had even been touched at all by anyone since AI coding assistants became a thing.
I wish this shot from The Terminator had the camera showing Sarah Conner’s face instead of Reese’s, because it’d be such an appropriate meme image on multiple levels for when someone makes a misleading claim about some current AI system.
I could see stuff getting small changes and them claiming that the entirety of the new version is “written by AI”.
Also, having 1/3 lines with obvious code that can be auto suggested correctly would make sense, but that is hardly code “written by ai” in the way they suggest.
I’d guess a lot of the people writing the code don’t even have it turned on, it’s just installed because management said it had to be, because management wants to be able to tell investors they’re “innovating work flows”.
I am a small sample to confirm that’s exactly the reason in my brother’s company.
And in my company we’re pressured to make X prompts every week to the company’s own ChatGPT wrapper to show we’re being productive. Even our profit shares have a KPO attached to that now. So many people just type “Hello there” every morning to count as another interaction with the AI.
Every few months I turn it on for a few days just to see if it is better.
Then I go back to the old AST based autocomplete that actually knows something useful about my code.
Those are the easy time savings though, the safe easy stuff the developer doesn’t have to worry about anymore. (Giving them time do the gnarly stuff)
It is exactly the opposite, with simple, predictable auto-complete you didn’t have to worry about that anymore, with LLMs you always have to look at it in detail because every little thing could be just plain completely different and wrong.
I can read way faster than I can type though. You still check it, but it’s pretty good as that kind of stuff once you have an example for it to follow.
Reading code is usually orders of magnitude slower than writing code. Sure, typing might be slower than reading but to check if it is what you intended you have to understand it too.
Well, I’m generally very anti-LLM but as a library author in Java it has been very helpful to create lots of similar overloads/methods for different types and filling in the corresponding documentation comments. I’ve already done all the thinking and I just need to check that the overload makes the right call or does the same thing that the other ones do – in that particular case, it’s faster. But if I myself don’t know yet how I’m going to do something, I would never trust an AI to tell me.
Well, okay, I can see how it would be useful in languages like Java that are extremely verbose and have a low expressiveness. Writing Java pretty much was already IDEs with code generation 20 years or so ago because nobody wants to write so much boilerplate by hand.