Do you have a dumb question that you're kind of embarrassed to ask in the main thread? Is there something you're just not sure about?
This is your opportunity to ask questions. No question too simple or too silly.
Culture war topics are accepted, and proposals for a better intro post are appreciated.
Jump in the discussion.
No email address required.
Notes -
How did AI annoy you this week?
I'm pretty pro-AI generally, but I've been frustrated by a huge uptick in vibecoded applications that have some use, but are pretty unimpressive and not well thought out, for what they are.
E.g. people in my socials:
And I have a mess of feelings about it. On the one hand, we live in an age of technical wonders, and I'm glad people are discovering them. On the other, because we live in an age of technical wonders, the bar for quality has gone up so much in the last year or two, and these people seem to lack any self-awareness. The default vibe coded design tropes are immediately apparent in these apps, like how you can sense AI writing with em-dashes or "it's not X, it's Y!", or just its general tone. And like, it's fine. It's okay. The apps work, but they should be so much better.
It's not like I haven't vibecoded some turds. I've made websites and android apps and tools too. It's just that these are for me alone, or to be shared in person, if someone requested it. To release one to the public, the actual utility of the thing would have to be unimpeachable. Tracker website for <thing that is already tracked> does not meet that bar. Map of <thing that is already mapped> does not meet that bar. Yet another app doing the same thing as a hundred others does not meet that bar.
Does it count if it's AI and also the entire software help ecosystem?
I can see the frustration that went into Hello World — Enterprise Edition, because any time I try to look up a one-line action, it spits out a hundred lines of code with a few dependencies. My average script is about 50 lines total, and the biggest program I've ever made for work is a few thousand lines spread between a few systems. I don't need a
MessageFlyweightFactoryfor that.More options
Context Copy link
LLMs have been VERY helpful to me around the house when I hit on normal homeowner problems that (usually) have surprisingly simple fixes or workarounds. Its probably saved me 1000's of dollars in theoretical professional repair bills (blunted by the time I actually have to spend to implement the solutions).
It once diagnosed a simple plumbing issue from a single photo and basic description of the problem. It helped me fix my water purification system by finding specific parts that I needed based on a couple photos.
However, certain real world issues with complex physical interactions seem to elude the thing.
I'm trying to diagnose a water leak in my washing machine, and I give it photos of the washing machine, along with model info, and photos of the water leak and general description. It has me go through some diagnostic steps but is very shaky on what the most likely failure modes are.
One of the issues, I think, is that I can't convey to it the SOUND that the machine is making very easily, or feed it video of the thing while its operating. I'm describing things to it, then it tells me to take a particular step and describe the outcome, and I think too much is lost in translation, and a washing machine is more complex than most appliances.
So its telling me workable steps to diagnose and possibly fix it, but its prognosis is all over the place depending on how exactly I describe the problem, and some of the fixes are involved and probably uneconomical.
So once I pop the washer open and figure out how bad things are, I'm almost certainly just going to buy a new washer.
Its not that I expected 'better' from the AI, but I think if I just had a handyman or repair guy come out they could figure things out within 5 minutes just by looking, listening, and poking around a little. THEN I could query ChatGPT as to whether their proposed price was fair or if they were likely yanking my chain.
More options
Context Copy link
I'm probably one of the more AI-sceptical people on this board. I don't think the God Machine is going to techno-rapture us to cyber-heaven anytime soon, but I do try to keep an open mind around the idea that it might have domain specific value, particularly in coding tasks.
I've noted here in the past that I haven't seen much value, even when using frontier models. The responses that I get are:
I had a chance at work to try using Codex with GPT-5.4. This is allegedly a top tier stack, and so far as I can tell, as close to the frontier as you can get.
I targeted a fairly straightforward performance issue in our codebase, where some JPA code was generating an inefficient query when two tables each grew two orders of magnitude larger than we usually see them grow. This is the kind of thing that would normally take me 30 - 40 minutes to write a few automated end to end tests, then ten minutes to fix.
Since I clearly have a problem with Prompting It Wrong, I spent almost two hours working with the planner describing the problem, and the root cause, and where the failing method was used. I described what might be at risk of breaking, and what tests we would need to write to prove out the fix. I described the architecture of the automated test system, and what the tests would need to verify.
After doing all this, I let Codex churn.
It generated tests that verified the wrong thing.
Then it did the fix wrong, in the name of "efficiency".
Then, rather than fixing the issue correctly, it tried to rewrite all the sites that called the method.
After losing most of a day to this, I fixed it my own damned self. I'm starting to think Djikstra was on to something.
My job is mostly just routine ports of legacy code into a new framework, with a few minor architectural decisions thrown in here and there. It took me some time to set up sufficiently exact instructions, but by now GPT-5.4 can do it pretty well. The end result isn't any worse or better than if I had done it myself (which makes sense, since I'm telling the copilot to utilize my own methodology), and it's just moderately faster, but notably it can do it while I'm off doing something else, with a handful of corrections on the way.
I don't like it, but it looks like it can do my job, at least.
More options
Context Copy link
Wow love that Djikstra essay!
More options
Context Copy link
More options
Context Copy link
Not AI itself, but I did see someone on HN claim they feel like they are finally good at programming because of AI... then go on to say that they haven't written any code in months.
I recently remarked to a friend that I probably wouldn't care about AI if it wasn't causing people to act retarded. Whether through stunning displays or ignorance like the one I mentioned, corpos shoving it down my throat when I never asked for it, or the insane tech industry push that you simply must use it for work, the hype train annoys the absolute shit out of me. But it's really not the tech's fault - if not for people acting stupid I would just go "meh, not very useful to me" and move on with my life. As usual, humans ruin things.
I'm not exposed to this personally, but a lot of my acquaintances in the industry are. I think your use of "insane" really nails it. When was the last time a tool in software genuinely produced a huge efficiency gain but didn't see widespread, enthusiastic adoption by the rank and file? Hell, most of the time, leadership would try to keep it from us at first in cases like that. CI servers and hosted version control come to mind.
I'm old. I've experienced a similar tool that was shoved down my throat by management at an F-500 company because it was going to eliminate the need for programmers. It was Rational Rose, and it turns out its main value proposition was that executives got really nice dinners from the sales team.
Maybe it's different this time, but it sure seems familiar.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link