when ai becomes an exercise in existisitential boredom
Social media's AI zealots are cheering an exercise in complexity and reinventing the wheel for the sake of complexity and reinventing the wheel.
One of the more frustrating social trends I see every day on tech-oriented social media is the constant high praise lavished on AI celebrities like Andrej Karpathy, formerly of OpenAI. Every week, he seems to discover a dozen new frameworks, and patterns, and abstractions, and proceed to have an existential meltdown on how he's never been so behind as a programmer and how AI is just so damn good at coding now you guys, and everything is just too complicated for the feeble human brain.
As someone who focuses on architecture and system design, this irks me. It's not that I'm against learning new technology or am vehemently anti-AI in code. LLMs do have a place in modern tech stacks. They're built for natural language processing, so throwing them at oceans of unstructured raw data and sentiment analysis is ideal. They're also a huge help with testing, documentation, and setting up common boilerplates, stuff that eats up a lot of hours, and is both demanded and very under-appreciated by businesses since they're not directly tied to revenue.
But when it comes to actual programming, they can be more of a hindrance than help, especially in code that requires a high degree of precision. Likewise, even though some devs think they're almost a quarter faster in writing day to day code with AI assistants, studies show that they're actually almost a fifth slower. Worse yet, the biggest reason why so many companies can brag that so much of their code is now being written by AI is thanks to mandates more than results, which have been less than stellar.
In other words, if you unleash it on busywork you can outsource, on tasks you can clearly define and do not have an immediate impact on core functionality and do not need to be actively babysat, then yes, LLMs boost your productivity and speed to market. If you have to keep going back and forth, or review tens of thousands of lines of output with automated tests that look like they pass and everything works, working on a tighter and tighter deadline? Not so much.
Basically, bragging about the exponential adoption of AI today, when bosses demand that employees use the chatbots or else, is kind of like bragging that 99% of people give you your wallet if you just ask them, and all you had to do was pull out your gun. And now that you've forced people to use the tool, you bring in someone like Karpathy to warn them that they better plunge into every AI trend as if their livelihood depends on it, because, well, it does. Then the same people ask why there's a massive backlash to LLMs, as if selling it through mandates and fear is somehow a sign of a truly amazing and beloved tool.
Still, even that's not what really makes me groan every time I see another edition of Karpathy's Complaint being regurgitated online. No, what bugs me is that our software is supposed to do two things. First, it should solve an actual problem. Second, it's needs to be as simple and efficient as the task allows for fewer bugs and easier maintenance. Building 87 layers of abstractions across 43 brand new frameworks, with 18 agents, and all of this is constantly in flux, being reinvented every day to maintain frenetic FOMO, as LLM makers keep on advocating, is just wasteful, self-indulgent absurdity.
I remember that when I was starting out as a professional, I'd look at code which was created by people who sold themselves as humble geniuses who could use the latest tools to solve problems in new and unique ways, and was amazed at just how complex the code was and how many things it completely reinvented, and thought that I just need to work harder and learn more to appreciate their genius insights.
Then, after spending years cleaning up their "amazing, innovative new code" so it could actually work for real users, at a quarter of their hourly rate, I'd realize that no, these guys weren't visionary genius coders whose work would take me lifetimes to start to understand. They were just codestrubating, creating elaborate frameworks, utilities, and toolkits no one asked for just for the sake of creating them, then dumping the real problem of making all this work in the real world on suckers like me.
Maybe it's just the early onset curmudgeonness talking here, but it feels like the tech industry has been so wildly and uncritically successful that it's eating its own tail. This is, after all, one of the biggest weaknesses of any coder. If left alone to work on a tool with no deadline, we're going to recursively add and polish its internals, features, tests, and frameworks forever, looking to experiment with the newest toy or try the wildest idea. It's a problem common enough to have a name: gold plating.
At some point, the technology ceases to be for the customer and exists more as an exercise for the coders. The actual problem is long solved, but the work still marches on by sheer inertia, and with modern product management philosophies like Agile, the project is never done. And so, projects are created for the sake for there being a project to work on, to pad resumes, and to look busy, maxing out metrics. Since the utility for the user is no longer a concern, what they think doesn't matter, the project just needs to be justified as "the future" and continue.
Honestly, this is where I think we are as an industry. We've solved so many big, thorny, and complex problems that we're, well, bored. So, we keep reinventing the wheel and making things harder for ourselves to feel productive because the truly, really out there ideas depend not on VCs with cash to burn in hope of 100x returns, but scientists and government agencies which are stingy in the best of times.
What people like Karpathy – if his public musings aren't an act – and his fans are doing with their pro-LLM lamentations are finding their own meaning in a world that is now increasingly in need of tweaks, nudges, and long, quiet deliberation, instead of The Next Big thing, which keeps turning out to be weird monkey jpegs, or magic internet money, or chatbots who drive you up a wall after enough time with them, on a two year hype cycle. But in that search, we're no longer being helpful, and waste tremendous mental and real world energy on solving problems of our own making.
