Here’s the lifecycle of the AI experiment in a software development company:
Honeymoon phase: “AI greatly increases productivity, we can reduce the number of programmers, get the product finished more quickly and gain a competitive edge; let’s use as much of it as possible and evaluate coders by token use, because it’s a productivity multiplier so coders who use more AI are more useful to the company”.
Greed phase: “AI greatly reduces or completely negates the differences between programmer skill levels; it just requires precise instructions. This means we can eliminate expensive programmers and replace them with writers of precise instructions for the AI. This is a generic job that requires far less training and qualifications and makes the employees easily replaceable and cheap. Oh goody.”
Mature phase: “We no longer hold any competitive advantage since everybody uses the AI. Also, our code base became a black box since no human in the company can actually inspect it and evaluate AI’s work. We are experiencing unfixable bugs, random outages, and AI intelligence level drops that make it impossible to understand its previous work, let alone fix it. The customers are complaining, but we can’t do anything about it now since our entire code base is AI-produced and AI-maintained and our human employees are all merely human interface for the AI. All our human coders either quit after we offered them pay cuts and/or limited promotion potential, or we fired them. We might have to start looking for programmers to supervise the AI, but since the code complexity is enormous, the cost would be order of magnitude greater than what we started with. Also, since we started experiencing outages and service degradation, the customers started leaving, and we lost 30% of revenue because they refuse to pay for our “AI slop”. We don’t actually have the finances that can handle the cost of employing high-end coders.”
Nightmare phase: “The AI providers understood that they have us hooked and are steadily increasing the token prices for the last year. They are also introducing tiered pricing scheme, where only the top tier is useful for our needs, but it’s astronomically priced and we can no longer afford it. Our token consumption went hyperbolic since the AI constantly generates problems that require enormous numbers of tokens to resolve. Also, there are all kinds of sanctions so American and Chinese providers are limiting access to their most useful models. We lost more than half of our customer base because they say our product became AI slop; the web interface is crashing, the data contains all kinds of nonsense, every now and then things get deleted or show nonsensical results, and they are done with us. We need to start downscaling just to keep the lights on, but if we do so, the product will instantly stop working because nobody will be able to control the AI’s blunders and we might as well close the doors. Why did we even think AI was a good idea in the first place?”
Actually, AI writes better code than vast majority of programmers because it seems later iterations of models were trained on enterprise level code, so you won't see AI slop like in generated text.
Plus, coding agents have debugger and tests which include not only automated unit tests but also UI tests that verify actual application. This eliminates bugs and bad code because tests will detect important regressions. And finally, human testing is final step. So it either works or it doesn't.
Sure, it still requires control and prompts require constant optimizations as models change and evolve and sometime devolve.
Price … well, that's an interesting one. For example, I wasted 450€ trying to do something, it was not successful and I had to throw everything away – huge waste you might say (and it did feel like it).
However, in the same time, in last couple of months I ported big chunk of old application which is 2 year of work for 2-3 people. That's close to 100.000€ in salary alone, let's forget missed opportunities because application was not ready for 2 years or more.
So, there is HUGE space between now and the point where I would say "AI is too expensive".
Sanctions and stuff … this is a problem, Fable and Fugu showed us that. However, we are at the brink of running frontier level models locally.
Currently it takes 4 x Mac Studio 5 Ultra to run Kimi 3 without quantization – which is Opus 5 level model – and we do not even need that level for most of daily work.
Sure, thats around 100k € … but that is reachable, unlike millions of euros for a data center.
Motif 3 can run on single Mac Studio 5 Ultra with enough room to work on it as main computer and is more than enough for every day work (it falls short on complex reasoning).
New models are popping daily, many of them open and more and more of them capable enough for daily work. There is no going back.
Now, I am sure there are many companies that will do exactly like in your article, but those will be mostly startups eager to grab investor money ASAP or AI oriented companies that sell AI based stuff.
Everyone else … I don't think so, mostly because you actually need a fairly good programmer to write good prompts. Only at first glance it seems like everyone can do it, but there are so many things that are implicit in developers prompt that regular person can never do it and get something meaningful since it won't even understand when AI gets it wrong let alone approach optimizations.
Let's not forget, coding AI models are trained on "developer brains", not "average person brain". Sure, average person might do some stupid simple tool or simple mobile app, but anything complex – forget it – you still need a good developer for that, not a story writer.
Could I continue without AI? Nope. But I loose nothing, since I would not be able to continue without AI anyway. Hiring 3-4 developers simply wasn't an option and I would try to last as long as possible without further development and then close the company (probably in a year or two).
Now we are back on the market and we can keep up – even though everyone has AI. However, there is still massive organisational difference – by the time big(ish) companies finishes their meetings and approval chains for each and every feature, I have finished application long before.