Anybody who has worked through the life-cycle of large projects knows that as time passes and the software gets adjusted and expanded, code just accumulates problem and brittleness - especially because multiple different people change it and they tend to each do it their way, often without full understanding of the code base - not just at the code level but also at the software design level - and eventually that code gets so hard to change or fix that a whole new system has to be built from the ground up.
In my experience this happens maybe at around 5 - 8 years of age of a codebase.
So I expect we’re headed for a spectacular industry-wide explosion because using AI code vastly accelerates this because for just about anything but small projects that can entirelly be generated in one go, AI isn’t consistent in coding style, much less software design.
Throwing software engineers at it right now only works if they end up spending even more time reviewing and ajusting the AI code than they would if they did the work themselves, since having AI coding is pretty much the equivalent of outsourcing to a pool of random junior developers.
Hell, even with smaller projects, you’re going to have debug cycles and if AI is driving those cycles, it will be acting as a new coder for each invocation (which happens multiple times per prompt for systems like claude code).
So you’ll get shit like duplicate helper functions, other code not using those helper functions anyways, debug code added and then not removed, errors and warnings using a variety of styles, overly verbose and redundant arguments, support for enhancements that don’t even make sense in that context, confidently incorrect assertions about what is and isn’t happening or possible, etc.
My manager wants me to make a presentation that sells some AI debug solution but the hand holding I have to do for it to actually understand and not give useless conclusions means I don’t even believe in it. Or the case where it did help, turns out it didn’t even use the tools provided by the solution and was just CC.
I’ve mentioned the cycle of being impressed with what these LLM-based systems can do and feeling like I might have been unfairly critical, and then running in to a major issue that justfies the earlier critical view. Last times I mentioned it, I said I was in the impressed (but skeptical) part of the the cycle. Well, I’m back to the “this might just be a complete waste of resources” part of the cycle.
I remember back in the day when they had to pull Cobol programmers out of retirement to update mainframe software because of Year 2000 and they got paid a bundle for it.
Similar thing for customization of older SAP systems after SAP changed the language used to Java but those systems were still done in the old language.
So I expect that freelance senior designer-developers are going to get paid A LOT of money to come fix things in a few years’ time, especially since in places with high AI adoption this is going to be way bigger in terms of size, complexity and seniority of expertise needed that either mainframe code updating for Y2K and updating customizations in old SAP systems.
Anybody who has worked through the life-cycle of large projects knows that as time passes and the software gets adjusted and expanded, code just accumulates problem and brittleness - especially because multiple different people change it and they tend to each do it their way, often without full understanding of the code base - not just at the code level but also at the software design level - and eventually that code gets so hard to change or fix that a whole new system has to be built from the ground up.
In my experience this happens maybe at around 5 - 8 years of age of a codebase.
So I expect we’re headed for a spectacular industry-wide explosion because using AI code vastly accelerates this because for just about anything but small projects that can entirelly be generated in one go, AI isn’t consistent in coding style, much less software design.
Throwing software engineers at it right now only works if they end up spending even more time reviewing and ajusting the AI code than they would if they did the work themselves, since having AI coding is pretty much the equivalent of outsourcing to a pool of random junior developers.
Hell, even with smaller projects, you’re going to have debug cycles and if AI is driving those cycles, it will be acting as a new coder for each invocation (which happens multiple times per prompt for systems like claude code).
So you’ll get shit like duplicate helper functions, other code not using those helper functions anyways, debug code added and then not removed, errors and warnings using a variety of styles, overly verbose and redundant arguments, support for enhancements that don’t even make sense in that context, confidently incorrect assertions about what is and isn’t happening or possible, etc.
My manager wants me to make a presentation that sells some AI debug solution but the hand holding I have to do for it to actually understand and not give useless conclusions means I don’t even believe in it. Or the case where it did help, turns out it didn’t even use the tools provided by the solution and was just CC.
I’ve mentioned the cycle of being impressed with what these LLM-based systems can do and feeling like I might have been unfairly critical, and then running in to a major issue that justfies the earlier critical view. Last times I mentioned it, I said I was in the impressed (but skeptical) part of the the cycle. Well, I’m back to the “this might just be a complete waste of resources” part of the cycle.
I wonder how many engineers they’ll have to pull out of retirement in a few years to fix a lot flawed logic.
I remember back in the day when they had to pull Cobol programmers out of retirement to update mainframe software because of Year 2000 and they got paid a bundle for it.
Similar thing for customization of older SAP systems after SAP changed the language used to Java but those systems were still done in the old language.
So I expect that freelance senior designer-developers are going to get paid A LOT of money to come fix things in a few years’ time, especially since in places with high AI adoption this is going to be way bigger in terms of size, complexity and seniority of expertise needed that either mainframe code updating for Y2K and updating customizations in old SAP systems.