A polarised world is usually a disaster but the entertainment-quotient is enormous. And so, as much as I dislike the kind of polarisation that's happening in my small little world of software developers about AI and LLM-driven software engineering, I tell you, it is incessantly entertaining.
On the one hand, the clickbait headlines and the trying-too-hard-not-to-sound-like-a-promotion-for-my-business opening salvos on LinkedIn posts have made all pronouncements there are to be made about the game-changing nature of AI-driven engineering and product-building. On the other, small little indie blogs are silently declaring war on the same thing, citing dead-end goose-chases and wasted productivity from the very same AIs.
What is actually happening? This is left to academics who have enough time to study. The engineers — both the ones vibing their way into a feature or out of a bug, and the rest who just canceled their Cursor subscription and went back to neovim — simply don't have the bandwidth.
The verdict from the academics (to the simple-sounding question "Does AI make the software dev more productive?") seems laced with the most common phrase you hear in software circles: it depends. (Well, they reported a 19% decrease in a narrow group).
The landscape is quite divided and there are no clear answers yet. Vibe coders are producing software artefacts in the time it would've taken a senior developer to setup the frontend. But elsewhere, the state-of-the-art model has already misdirected someone in a 2-hour hair-pulling bug hunt. The joke is that the bug was written by Claude in the first place during that vibe coding session.
A single vibe-coded artefact brings surprise and shock at the same time — founders and top-level collars are surprised at the result because they go, "imagine the possibilities now!", and software engineers are shocked at the codebase because, "who the f— writes code like this?" (plus, "oh god, now I have to clean this mess up within unreasonable timelines).
Amidst the haze, I think some clarity has been achieved collectively. We now understand that this AI-driven productivity gains is quite possible in some scenarios; notably, building prototypes from scratch and grunt-work that's easy to verify. Elsewhere, disaster rules.
Founders and top-level managers who seldom code but have hundreds of ideas they'd like to build and profit from have the ultimate tool — building prototypes of what they have been dreaming about is easier than ever before and involves very little money, very little time and, often, very little coaxing of the LLMs.
Grunt-work that you would otherwise debate between writing an automation for or just doing manually — and end up spending way too much time trying to write that automation script, immortalised by Randall — is now a couple of prompts away (usually).
Between these poles of birthing a prototype through vibe-coding and letting an LLM loose on mindless grunt-work, there is a large swath of land. A land of complex, large codebases needing a feature add. Of imperfect code with duplications and edge-case bugs and occasional reflection of human fallacies requiring large-scale refactors. Of codebases in obscure languages or frameworks asking for bug fixes. The folks at the poles have tasted immense success and productivity through AI and what they expect is that the folks dealing with this large, unruly land tame the wilderness with AI. It is a reasonable expectation in a perfect world where patterns are unblemished and consistent. Alas, LLMs are anything but.
But beyond all this, there's something that bothers me.
What to do with the productivity gains? Are you going to empower people to spend less time on work and more time on their personal lives or are you going to make them march to death to the tunes of profit-maximisation and shareholder happiness?
If we go by the historic behaviour of the industrialised world, there's only sad news. Technological progress has brought more work-hours and less work-life balance — exactly the opposite of things technological progress was supposed to offer — to the lives of the workers. This new kid on the tech-progress block is no different.