Posts for 2025

As the progress of Artificial Intelligence marches on, I find myself needing to reiterate a point I may have made earlier, simply because it cannot be overstated.

While doing some reading on the internet, I came across this thread and quickly became immersed in the arguments and counter-arguments presented. Something stood out to me: the assertion that LLMs are incapable of genuinely writing software and will always churn out code that is inferior to what is produced by human developers who actually understand what they're doing.


It was a fascinating discussion, not least because of the developers who came out to say how useful they found A.I in their work. But in all this back-and-forth, there's one thing which none of them seem to have mentioned. Namely, the importance of quality.

LLM-generated software can work. However, here's the thing software developers know that laypeople don't - working software is the lowest possible bar to meet. This is like calling myself a "good person" just because I've never molested a child. Of course software has to work. But properly-crafted software does not only work; it is also maintainable, clean, extensible and testable. It is robust and won't break after a few changes.

That's quality software.

Is it true? Do LLMs really produce rubbish?

I think people who aren't tech workers, even some tech workers themselves, really underestimate just how defective all software is. Software these days does not exist in a vacuum. It's built on layers of existing software platforms, and connected to other pieces of software, each with its own set of flaws and vulnerabilities.

The end result is that a staggering amount of software in the world today is held together by duct tape and prayers. And that's only a slight exaggeration.


What has that got to do with LLMs, you ask? Well, what do you think LLMs are trained on? That's right - existing software. The crap code we've all committed at some point or other, if we're all being honest. And LLMs are gobbling all of it up, warts and all. If rubbish is the input, what are LLMs likely to produce as output?

That's right - more rubbish.

So what if it's true?

The world at large is made out of non-technical people. They are perfectly happy for software to just work, and give zero shits about maintainability, code cleanliness and all those high-minded ideals. They aren't the ones who have to deal with software on that level.

So what if the LLM-generated software is generic bug-ridden rubbish? It works half the time. Maybe even three-quarters of the time. It's good enough until the next LLM-generated software comes along. The important thing is that we meet those business needs now.


Let's use food as an analogy. Why do we eat? If we disregard luxuries such as pleasure and nutrition, at the core of it, we eat because the human body is designed to starve (and die!) if it goes without food for too long. So, what's the difference between a double-tier cheeseburger with fries slapped together by a fast food chef, or a gourmet meal lovingly prepared by culinary genius Gordon Ramsay? Forget nuances in taste, empty calories, cholesterol, salt and all that jazz. Both do the job of staving off starvation. One is cheap and produced in minutes. The other needs significantly more time and money, plus you have to put up with that smug bastard Gordon Ramsay. What do the majority of people generally choose most of the time? It's a no-brainer; Option A wins.

Let's examine another example. What about art? Some collectors or connoisseurs will pay a hefty sum for certain pieces or works from certain artists. But what if they weren't interested in art? What if they could only afford cheap imitations? What if their purpose was simply to cover the walls with something, anything? Would cheap A.I-generated art, a simulation of the genuine article, suffice? I think we all know the answer to that.

Finally...

The reality is that quality doesn't matter. At least, not as much as most of us think it should. Especially in software. Questions like "will it accommodate future changes?" and "have all security holes been closed?" and "what happens if we give it unexpected input?" have been replaced by "does it generally work?", "is it too costly?" and "is it ready right now?". Artificial Intelligence didn't cause this; years of business culture did. But that's also a big reason Artificial Intelligence isn't going away - because the world is addicted to quick and dirty solutions.