Can AI Tell You If You’re Reinventing the Wheel?
Is our AI code assistant quietly telling us more than just what the next line of code should be?
My most popular post on LinkedIn this year, by some margin, was about vibe coding, This, from someone who largely talks about HPC and supercomputers.
Whilst I haven’t vibed with Cursor in quite the same way since then, I have recently had the opportunity (yay!) to turn my hand back to crafting software. I’d love to do it the old fashioned artisanal, organic, home grown, slow roasted, handmade way but the reality is I no longer have that luxury. Whilst I’m not doing it based purely on vibes, I am often accelerating what I do with a bit of electron addled assistance.
I don’t think it’s any secret at this point that LLMs work particularly well when they are operating in domains where they have good training data. Even Andrej Karpathy (who coined the phrase Vibe Coding) says so. And who am I to argue with him. He switched to writing code old school style for a recent project because the AI just didn’t have the beans for it.
I’ve certainly found the same, and often in anything that’s HPC related my AI powered chainsaw of a coworker tends to go off the rails rather quickly. No surprise there either. What I wonder though, is that if we shouldn’t be using this fact in itself as a signal?
Can we use how well the AI assistance is working as a gauge to indicate how unique what we’re doing is? If so, can that serve as a signal as to if we can reuse something that exists already (assuming that’s a good idea) or if what we’re doing even needs doing again?
Since I’m sharing random AI coding thoughts, I do still think we haven’t cracked the tooling (or maybe the process) for AI software development. A few weeks ago I put down a new subfloor in my kids’ bedroom. This meant driving over 1000 screws. Clearly, I used a torque driver for that. That’s my AI copilot. Critically though, I also used Torx head screws, not Philips or Pozidrive and certainly not slotted heads. That seemingly tiny difference saves hours and more importantly, reduces frustration massively. I don’t know what the equivalent in AI tooling for software development is, but I know we haven’t seen it yet.
Just a thought. Probably nonsense but its Friday and close to Christmas 😛
