I'm not as optimistic about the future of AI. Yes, products will get pumped out faster. Pretty sure the quality will be lower, but of course I guess AI can fix it faster too. Now I'm a bit biased because I work in the intelligence community. I produce software that can affect whether people live or die, or national security. And it gets fielded in places where I just can't say a week later, here's a patch to install to fix the first 7 bugs found. It has to be designed, coded, and thoroughly tested (using test software that also has to be written, guess AI can do that too?) At best AI will produce software that produces the correct outputs all the time or damn near all of the time. Which could be a good thing. But if it is completely created by prompting, it will never have a coherent, logical design that someone who has to maintain it can understand. AI doesn't work that way. I guess a good software engineer could do at least the high level design and then have AI code the various components to specs. That's better. But there is a problem with that too. A software engineer can execute a good design because they have years of experience writing code and understands how design translates to implementation. They understand all the stages in between the design and code that actually works on the machine. Also, a good software engineer can look at the code that AI has produced and evaluate it, make sure it meets standards, test it, and so forth. So that's good. But this hypothetical software engineer can only do all the stuff described in the previous paragraph because they have written tons of code, made all the mistakes and learned from them, and understand why doing it one way is better than doing it another. And as the years pass and software engineering becomes more and more about prompting AI at a high level to produce something that works -- there won't be any of us left who have made all those mistakes and have the background to intelligently evaluate the code that AI is producing. We'll all be retired (377 days from now, in my case). In the end you will have AI writing code, AI writing the test code that tests the code, AI checking the code to make sure it meets the design criteria, and you will have code that no human fully understands, and you won't even have humans who have the background to fully understand and evaluate any large and complex codebase. And the more code you produce that was not written by actual developers, and which has either never been really looked at and evaluated by competent people, or certainly hasn't been fully understood by competent people and created specifically to fit in an overall coherent design -- it will be FAR easier for security problems to occur. Either security flaws that could be exploited will be in the code because no one knew how to look for them, or else malicious actors actually exploiting the AI to put malicious things in the code. It's a disaster waiting to happen. And then you look a few years down the road -- we all know about how AI can produce something that isn't "quite right" because it is just mimicking patterns of the trillions of bytes of data it has access to -- whether that is pictures, or large volumes of human writing, or code in open source repositories. But it's 98% there. Which is often good enough. But a few years down the road, most of the code in the open source repositories will have been written by AI, not by people. So instead of mimicking the best code out there that people have been written, and approximating that using the pattern-matching algorithms that make AI work, and getting "98%" there -- it will be trying to mimic other AI code that was probably at best 98% there to start with. So that generation's AI will produce code that is "96%" there. And so on. Quality will degrade over time. I guess the AI creators will have to work harder and harder to improve their AI to prevent that degrading from occurring. Who will be coding these AI models? Other AI, I guess. I don't know, maybe I'm just an old timer, like those folks who think that someone isn't really driving if they are using automatic transmission. Obviously there have been other advancements in software that have created abstract levels above the basic machine -- first assembly code, then compilers, then higher and higher level languages, abstractions of network layers, etc. And they've all been for the better. And there's always been old timers who felt they understood what was going on under the hood better because in their past they had written in assembly code and lower level languages that were much closer to the "bare machine". But each of those layers of abstraction was at least fully understood and definable. You don't always need to know every bit of implementation of how a compiler for a high level programming language turns source code into machine instructions -- but it is a concrete, repeatable, definable process. A compiler can be tested to make sure that it meets all the requirements of that translation or it won't be considered a true compiler for that language. It could have bugs in it like anything else, and one compiler could be faster or more efficient or create faster running machine code than another or handle errors differently. But at some point a compiler just takes a fully described programming language standard and creates machine code that does what it is supposed to. There is no such definable standard for AI. You could give it a bunch of prompts today and it could produce, say, 10000 lines of code to do something you want. You could give it those exact prompts a week from today and it could produce 11000 different lines of code. Because its model has ingested more data since then. Both pieces of code might meet your requirements, in a different way. But you are no longer using something like a compiler which has a definable and testable standard to judge what it is producing. Or you could change something unnoticeably small in your prompt (maybe you say "a" instead of "the" in the middle of a 2 paragraph description of what you want done). And the output could be different. As a software engineer, that scares the heck out of me, because I have lost the ability to say that I, or my team, actually created the code. Yet I am still responsible for it. Glad I'm retiring in 377 days. ** Edited by SteveInBaltimore at 2/8/2026, 3:12:06 PM