When people ask what I do for a living, my response is “software engineer”. Not “programmer” or “coder”, because the actual programming is only part of the job.
The big concerns for software engineers are reliability, maintainability, and scalability. Using LLM-generated code without reviewing and understanding it is antithetical to all three of those big concerns. Personally, I find that the amount of review and revising it takes doesn’t end up saving me any time at all compared with me writing it myself.
Almost anybody can build a bridge given some materials. But there’s a reason we have licensed engineers who review and sign plans for them. If your methed out neighbour built a bridge in an hour by himself, would you drive over it?
Albertatech has a video about a manager who fired his last software engineer because “he could code faster with Claude than an SE” and she was all “most programmers can code faster than an SE even without Claude, the point of an SE isn’t speed, it’s making sure your codebase is maintainable, scalable, and reliable”
These managers decide who stays on their teams without any understanding of what their team does. It’s like a 5 year old firing Davinci because he can draw a lady smiling with long brown hair faster with Claude crayons.
If you know in your heart of hearts that if you don’t show up for work, everything’s fine, but if all the workers don’t show up for work, everything stops, you have to, like, confront the possibility that maybe you’re not in the driver’s seat, maybe you’re in the back seat with a toy Fisher-Price steering wheel. And AI is like the possibility of wiring that steering wheel directly into the drive train.
I’m very much like you. I do find that for some specific circumstances it does save probably 40% of my time, especially for greenfield projects. For existing code it’s very useful at straight refactoring from old code standards to new ones.
I’m not even a software engineer ( not full stack or anything at least), and I would never use AI code if I didn’t understand and read it first. It’s fast and taught me some tricks I didn’t know, but do you all have management that just lets you submit shit that you have not read or understand?
I get ugly looks for suggesting that maybe we review code before we merge it because not doing that is only a half step away from just pushing straight to main.
There’s talk about full AI development loops with no human intervention, just people designing implementation plans and throwing them at Claude to maintain autonomously forever.
In other news, this week I’m reducing thousands of lines of AI generated code that repeats itself and has no concept of patterns into an instantiable module, the way it should have been written from the get-go. This code does things like say it grants access to a load balancer from a particular single application source when in fact it grants that access to the entire cluster CIDR and every single thing running there.
Most management doesn’t even have a working knowledge of this stuff, so they have no idea whether what you submit is good or bad, only whether it “works”.
When people ask what I do for a living, my response is “software engineer”. Not “programmer” or “coder”, because the actual programming is only part of the job.
The big concerns for software engineers are reliability, maintainability, and scalability. Using LLM-generated code without reviewing and understanding it is antithetical to all three of those big concerns. Personally, I find that the amount of review and revising it takes doesn’t end up saving me any time at all compared with me writing it myself.
Almost anybody can build a bridge given some materials. But there’s a reason we have licensed engineers who review and sign plans for them. If your methed out neighbour built a bridge in an hour by himself, would you drive over it?
Albertatech has a video about a manager who fired his last software engineer because “he could code faster with Claude than an SE” and she was all “most programmers can code faster than an SE even without Claude, the point of an SE isn’t speed, it’s making sure your codebase is maintainable, scalable, and reliable”
These managers decide who stays on their teams without any understanding of what their team does. It’s like a 5 year old firing Davinci because he can draw a lady smiling with long brown hair faster with
Claudecrayons.Cory Doctorow has a really good quote about why AI is so appealing to managers:
I’m very much like you. I do find that for some specific circumstances it does save probably 40% of my time, especially for greenfield projects. For existing code it’s very useful at straight refactoring from old code standards to new ones.
I’m not even a software engineer ( not full stack or anything at least), and I would never use AI code if I didn’t understand and read it first. It’s fast and taught me some tricks I didn’t know, but do you all have management that just lets you submit shit that you have not read or understand?
I’m just amazed.
I get ugly looks for suggesting that maybe we review code before we merge it because not doing that is only a half step away from just pushing straight to main.
There’s talk about full AI development loops with no human intervention, just people designing implementation plans and throwing them at Claude to maintain autonomously forever.
In other news, this week I’m reducing thousands of lines of AI generated code that repeats itself and has no concept of patterns into an instantiable module, the way it should have been written from the get-go. This code does things like say it grants access to a load balancer from a particular single application source when in fact it grants that access to the entire cluster CIDR and every single thing running there.
Most management doesn’t even have a working knowledge of this stuff, so they have no idea whether what you submit is good or bad, only whether it “works”.
I usually just tell people I’m really good at pressing buttons.