Colm Murphy

sorry in advance for the tl;dr... part ranting/part venting here :) i really really really don't see how anyone thinks llms are going to take over all software engineering. i say that as someone who does a lot of software engineering, a ton of vibe coding, and employ a lot of developers (read: humans). llms def aren't going to do that. it's a symbiosis still evolving, but a symbiosis nonetheless. that's the only way it ACTUALLY works. every time i try to one-shot vibe code a project, i end up with 300+ commits fixing things the agents got wrong. not just bugs (i mean there are a ton of those too) but also deliberate decisions. two examples: front end. models can create beautiful designs, but they don't understand what it's like to actually use them. spacing, visual hierarchy, scrolling, four font sizes fighting for attention in the same field of view...it looks great until you start using it and then you realize it sucks to use...agents don't actually use software, so they don't understand things that people who do intuitively grasp back end. i built something using my personal inference API key to prototype... when it looked good, i gave the project its own key. the model left some calls on my old key without telling me, because it calculated that was cheaper (which btw, it was not). a human engineer would understand why i wanted everything going through one provider. the llm did not because it never works on a budget, never has to reconcile the books, never has to trace a credit card charge back to a specific invoice, etc. humans make mistakes all the time. but a human would understand why every endpoint needs to move and not just some of them... if they miss one, it's an oversight not a deliberate decision. models will just decide to leave some on the old key, on their own. its wild. this is why claims that llms will replace all programmers, work entirely autonomously, or take over the world seem so very dubious to me. llms are astonishingly capable and still struggle with incredibly basic shit. i think love AI so much precisely because i've spent nearly 25 years delegating work to humans. the process hasn't changed: i define specs and hit send... a human or model builds. i review. give feedback. revise. repeat. the difference is speed. and that part is truly fucking incredible. but you do sacrifice for that speed. it's not free. but at least with my experience, i know what to look for -- i've spent decades building software. it's hard to bullshit me. so my worry isn't that models will take over the industry, or that llms will kill saas by giving everyone their own bespoke applications, or that they will lead to mass unemployment. my real worry is that models are already *just good enough* for people who don't know what they're doing to ship software that's subtly broken with chaotic ux and weird business logic. its a different type of slop. we'll get less xss risk and more cost risk, less race conditions and more weird user experiences... these are real hidden costs but they're usually not broken enough to fail the quick smell check.... just enough to create problems that nobody notices until later, when cleaning them up will be a lot harder. that's why i'm encouraging my son to learn computers, code, networking, business, and delegation skills etc. in a world of people who understand nothing, having your head wrapped around the nuances of how things work will give him a big leg up, as it does me. that advantage isn't going away anytime soon. the best software of the last 20 years was built by human engineers. the best software of the next 20 will be built by human engineers working with llms. there is no world where llms alone build full applications without human involvement. the roles will ebb and flow, but at the end of the day, the best, most profitable outcomes will be a team effort. so beef up on technical knowledge and get good at delegation and adopt as much ai in your work as you can... just don't let it make you dumb. if you've got this, the sky is the limit but if not and not willing to learn....good luck! </end rant>