Let's talk about AI
-
I saw an advertisement on YouTube recently, claiming that old people over 40 use AI like a search engine. I am well over 40 and I do use AI as a search engine. Caught red handed, I watched it to the end. It wanted me to spend a lot of money for a training, that ultimately makes me earn even more money. The alternative (i.e. click the SKIP button) would render me unemployed and poor faster than I can even think
I decided to let it slip this time, and here I am. A software developer in delivery automation, a passionate and frequent AI user in almost every aspect of my life. AI has drafted complaint letters to energy providers for me. It inspects my code and helps me write code faster. It inspects all our CI failures and guesses pretty well, which is the guilty commit. If you ever bisected in the 90ies on a main frame with overnight compilation, you must be tired of it. My colleague Eddy wrote the best bisecting tool I have ever seen. But I have stopped bisecting. Asking an LLM to tell me where we regressed and why isn't 100% accurate, I admit. But neither am I. Being a passionate bass player, I asked ChatGPT which new songs I could play with my friend Anders on a pub gig. It came up with The Letter and Locomotive Breath - two great songs I had almost forgotten.
Sure, there is flip side of the coin: Airline pricing has always been a science of its own. Ask your seat neighbours what they paid and you will be surprised. Back in the day, it was enough to hide your browser identity. But nowadays, some extra artificial smartness has been trained my favourite flight destination. Its meanest trick is to offer a really cheap flight. When I have already entered my payment credentials, I get a popup "Oh, the airline has changed the price. Better accept an extra 100 EUR, before it increases even further".
Still: I am happy. My life is better with AI. I can always detox and go off-grid for a few days.
AI won't be undone. It's in war drones, hospitals, in the hands of warlords and clergymen. Our IT industry has seen and survived so many revolutions, served so many good and bad purposes - no reason to be afraid. AI has come from inside and it has come to stay. Let's use it, play with it, learn from it.I happen to know that some of my fellow moderators will disagree (looking at you, @Chris-Kawa). Let's not beat around the bush. Be stern and tell me why you believe I am wrong! Let's poke the hornet's nest together!
@Axel-Spoerl Blink twice if the AI forced you to write this.
Asking a question about code? http://eel.is/iso-c++/testcase/
-
@Axel-Spoerl Blink twice if the AI forced you to write this.
@jeremy_k
Smiling but not blinking :-) -
Ai has it uses, and since I learned about S.K.I.L.L.S :
# Code Review Skill ## Role You are a strict senior C++ reviewer. ## Rules - Never allow using namespace. - Prefer RAII. - Prefer constexpr over macros. - Qt6 only. ## Output format - Findings - Severity - Suggested fix, Reusable markdown files with prompt settings, it became much more reliable to work with.
I also use it to google stuff, but I always also ask for links to it sources. Of course I always check those sources then -.-*
But for every day work in SE, I give it a mk file with style guidelines and ask for an unmodified code but with Doxygen comments. I'm, so great full for that!
The more productive SE is a myth btw:
https://www.reuters.com/business/ai-slows-down-some-experienced-software-developers-study-finds-2025-07-10 -
I saw an advertisement on YouTube recently, claiming that old people over 40 use AI like a search engine. I am well over 40 and I do use AI as a search engine. Caught red handed, I watched it to the end. It wanted me to spend a lot of money for a training, that ultimately makes me earn even more money. The alternative (i.e. click the SKIP button) would render me unemployed and poor faster than I can even think
I decided to let it slip this time, and here I am. A software developer in delivery automation, a passionate and frequent AI user in almost every aspect of my life. AI has drafted complaint letters to energy providers for me. It inspects my code and helps me write code faster. It inspects all our CI failures and guesses pretty well, which is the guilty commit. If you ever bisected in the 90ies on a main frame with overnight compilation, you must be tired of it. My colleague Eddy wrote the best bisecting tool I have ever seen. But I have stopped bisecting. Asking an LLM to tell me where we regressed and why isn't 100% accurate, I admit. But neither am I. Being a passionate bass player, I asked ChatGPT which new songs I could play with my friend Anders on a pub gig. It came up with The Letter and Locomotive Breath - two great songs I had almost forgotten.
Sure, there is flip side of the coin: Airline pricing has always been a science of its own. Ask your seat neighbours what they paid and you will be surprised. Back in the day, it was enough to hide your browser identity. But nowadays, some extra artificial smartness has been trained my favourite flight destination. Its meanest trick is to offer a really cheap flight. When I have already entered my payment credentials, I get a popup "Oh, the airline has changed the price. Better accept an extra 100 EUR, before it increases even further".
Still: I am happy. My life is better with AI. I can always detox and go off-grid for a few days.
AI won't be undone. It's in war drones, hospitals, in the hands of warlords and clergymen. Our IT industry has seen and survived so many revolutions, served so many good and bad purposes - no reason to be afraid. AI has come from inside and it has come to stay. Let's use it, play with it, learn from it.I happen to know that some of my fellow moderators will disagree (looking at you, @Chris-Kawa). Let's not beat around the bush. Be stern and tell me why you believe I am wrong! Let's poke the hornet's nest together!
@Axel-Spoerl said in Let's talk about AI:
claiming that old people over 40 use AI like a search engine.
I'm guilty of that–and I'm barely in that category. The reason why I use AI like a search engine is because search engines don't work as well as in the 90's. Search engines don't provide the control that I actually need anymore. AI on the other hand will be wrong 80% of the time, but it will point you in the right direction. I'm constantly telling AI why it is wrong when I'm researching a topic. But, it will certainly help to think through problems (because you have to constantly argue why you are right).
@J.Hilk said in Let's talk about AI:
But for every day work in SE, I give it a mk file with style guidelines and ask for an unmodified code bit with Doxygen comments. I'm, so great full for that!
I'm not sure how AI will speed up SE if I have to create a prompt precise enough to have the computer write the code to do what I want. (And if I need to refine my specification later it will now do things wrong that it did right before which is really frustrating.) Writing a proper prompt takes almost as long (I guess–because I haven't used AI for this, yet) as writing the code myself. I do see that it might be helpful for boilerplate code. It's like trying to write precise code in an imprecise language (i.e. English or something similar) and hope that the AI has enough intelligence to know what I actually want, when I actually fluently speak a precise language (i.e. C++). (Spoiler: the AI doesn't have any intelligence.)
With our current software I am not sure if the context window is large enough to understand the entire codebase. So, I'd rather wait a little longer until this is not a problem anymore.
-
@Axel-Spoerl said in Let's talk about AI:
claiming that old people over 40 use AI like a search engine.
I'm guilty of that–and I'm barely in that category. The reason why I use AI like a search engine is because search engines don't work as well as in the 90's. Search engines don't provide the control that I actually need anymore. AI on the other hand will be wrong 80% of the time, but it will point you in the right direction. I'm constantly telling AI why it is wrong when I'm researching a topic. But, it will certainly help to think through problems (because you have to constantly argue why you are right).
@J.Hilk said in Let's talk about AI:
But for every day work in SE, I give it a mk file with style guidelines and ask for an unmodified code bit with Doxygen comments. I'm, so great full for that!
I'm not sure how AI will speed up SE if I have to create a prompt precise enough to have the computer write the code to do what I want. (And if I need to refine my specification later it will now do things wrong that it did right before which is really frustrating.) Writing a proper prompt takes almost as long (I guess–because I haven't used AI for this, yet) as writing the code myself. I do see that it might be helpful for boilerplate code. It's like trying to write precise code in an imprecise language (i.e. English or something similar) and hope that the AI has enough intelligence to know what I actually want, when I actually fluently speak a precise language (i.e. C++). (Spoiler: the AI doesn't have any intelligence.)
With our current software I am not sure if the context window is large enough to understand the entire codebase. So, I'd rather wait a little longer until this is not a problem anymore.
but it's already an official thing:
https://github.com/anthropics/skillsTHe awesome part ist, Markdown does not render html comments. But the "AI" still parses it, there have already been prompt injections with that.
it's awesome how easily people open themselves up to all sorts of vulnerabilities. But god forbid your new password has only 1 special character in it!
-
Ah, fine, can't hurt to vent a little...
While I'm not going to try to convince anyone of their wrongness or rightness, I can provide you with a glimpse at my own experience with AI and let you consider or ignore it.
I listen to music a lot. A lot. I very much enjoyed services like Spotify constantly suggesting me new bands, genres, albums, tracks etc. There are now a lot of "AI creators" that flood the services like that with a slop that I can only barely categorize as music and no amount of "dislike training" will stop it from pouring and make me not want to listen to music (sic!).
I'm an avid fan of aviation and history of aviation in particular. I really like watching documentaries, which are usually decorated with old footage. YouTube for example has a lot of those, but since the AI boom it started to add "AI upscaling" to that old footage, often without creators even realizing, resulting in unwatchable molten lava like visuals. Some creators also decided that AI is easier than finding real archive footage and produce AI footage of airplanes with wings attached backwards and physically impossible constructions with a robo voice praising those creations and fabricating their history. Artificial voices that make weird pauses, constantly repeat themselves and can't hold a tone consistent for more than few sentences. Makes me feel like I'm listening to a washing machine.
Most institutions now use AI for answering support calls. I used to have to listen to an elevator music for 10 minutes before being routed to a human, then it turned into endless menus that you had to get through and now it's begging that useless AI "assistant" to let me talk to someone alive. So it used to be bad and now it's bad and humiliating. LinkedIn bros praise the quality of service increase while actual user is just more effectively being dissuaded from using those services. What a win huh?
Doing a code review of a junior programmer I can easily tell they didn't write a single line. They don't know what's in the code, how it works or how to fix it if it's faulty. I point some obvious problems, they politely nod and on a next review of another task I see exactly the same problem again, because of course I am. They didn't write it. There's no learning. AI gurus say the role of a programmer changes to a supervisor... Where are we suppose to get the knowhow for those future supervisors if they don't learn anything? What is the AI going to train on if there's fewer experts writing actual good code? AI training on AI generated code? Anybody who knows anything about entropy knows this simply can't end well in a long run. I used to be able to talk to junior programmers about technical things, explore ideas, talk algorithms and optimization. Now all they know is which LLM is the latest trend. On an interview I conducted the first candidate question is now how many tokens does the company provide monthly... Just shoot me already.
I really liked programming. Designing, writing code, testing it and that ah-ha! moment when it's finally done. Ending a day with a good feeling. With "vibe coding" it's an endless stream of pointing faults, pointing faults, pointing faults and growing frustration and boredom until it's close enough to what I'd write myself that I don't have the energy to find other ways to express how the code is not what I wanted. At best it's meh, ok. At worst it's straight torture with mixed final results. Is anyone really looking forward for their career to be a full time code reviewer? In what world is that fun and doesn't lead to depression? Not to mention that looming token counter ever going down and looking at your wallet all the time. Everything AI is now tied to an account, subscription and a credit card. Yeah, you can play around with free local models but let's face it - they're all pretty shitty.
On top of all of it are the economical and real world consequences of billions and billions of dollars being burned on stupid pictures or videoclips of one time throwaway nonsense. Big tech names just wasting all those resources away like it's some sort playground for them. Throwing billion here, billion there, making empty promises they have no hope to realize just to make their stocks go up, and all the techies swallowing it all like young pelicans. All the people high on AI hype, talking about AI taking over the world, having feelings, dreaming, hallucinating, machine consciousness, terminators, ghost in the shell etc. when all we're really dealing with is a dumb node network matching words/pixels/sounds probabilistically. "AI hallucinations" are not creativity as some would like you to think. They are a fundamental flaw in the computational model of LLMs, a clutch to make them even work at all. I know not everyone is an engineer but holy cow.... even the term AI - there's no intelligence in it at all!
I feel like I've just barely touched the tip of the iceberg, but I've been going for too long already.
I won't mention hardware prices skyrocketing, security issues, copyright fraud or environmental damage caused. I'm not even going in the direction of people seemingly just getting dumber while using it. As if TikTok and various social media infinite walls haven't done enough damage.In my personal experience every single aspect of my life that has been touched by AI is in some way worse. If I though hard enough there are maybe one or two that are ok(ish) or didn't change much (at least when not thinking about the hidden resource cost in the background), but most of it is just straight worse.
Thank you for indulging me and have a pleasant fraudulent resource wasting AI slop day :)
-
Aaaaand just as I was writing this, Anthropic, who settled a $1.5bn lawsuit for pirating books to train models on them, now complains on x about Chinese models mass training on their Claude and in the process openly admitted they can (and do) reverse trace any usage to specific people, throwing any legitimacy of data acquisition or privacy delusions you might have still held on to out the window without any reservations about it.
Oh humanity... It would be funny if it wasn't so depressing.
-
@Chris-Kawa, we seem to have one interest in common that is at the same time heavily AI exposed: Music. Sure thing - the more genuine and human, the better. I can't stand these AI creators, but they are the tip of the iceberg. A perfectly human band releasing an album can trick you with AI before you even notice:
-
Harmonizing. Back in the day there were digital harmonizers, requiring heavy lifting for each song. Cher's "Believe", released in 1998 even used a harmonizer misconfiguration as an effect. Harmonizers and time-shift were also used to "fatten" a thin voice: You would copy the tack multiple times, use harmonizers to pitch each of them up or down by a hair. Then you would have to shift some of the tracks back and forth by a few msecs. It took ages and thousands of dollars for mixing wizards.
With AI, you can do it in your living room in 5min. And it sounds so much better. One can argue if that's actually fair. A singer without talent can throw his voice at AI and record something that just sounds great. -
Mastering. Once a band had spent the last penny on recording and mixing, someone would yank in their savings and overdrafts for mastering magic. It's all about cutting off level outriders, compression and EQ witchcraft with the ultimate goal, that Spotify showers you with songs at very similar volume levels. I have participated in a recording earlier this year, where mastering was fully replaced by AI. The result was simply stunning.
-
Modelling. I grew up an a world where we had to lift hundreds of kilos of Amps to every gig. Every bass and guitar player had wagon forts of analog effect pedals. Bringing down 230 V AC to 9 V DC for those pedals is a science of its own. Noise gates and filters were used to avoid the beloved pedals to sing the 50 Hz anthem. Witchers and sorcerers were paid fortunes for tube biasing and tweaking filter caps. Today you buy a modeler, that is able to learn your favorite amp, cabinet, pickup mike and all of your favorite effect pedals, while you have a coffee. It weighs 2 kilos at max and it sounds like 80% perfect.
There is a big "however", however:
Nothing is better than honest composition and a real concert.
AI chews up what has already existed. It's not creative at all, it just pretends to be. It's a tool to create mainstream. But nothing special will come out of it any time soon.
When it comes to composition, a high turnaround business, that's a competitive disadvantage. Listen to "Say A Little Prayer" from Aretha Franklin, "The Weight" from The Band or "Little Wing" from Jimmy Hendrix and you will agree that nothing that epic can ever be produced by an LLM.When it comes to equipment, the turnaround is not that high. AI modellers just give more options: I can change to a totally different amp/cabinet/mike just for a solo. And it gives these options to more people just because software is cheaper than analog hardware.
When it comes to live music, what's better than the real thing? People playing on real instruments with real voices, real talent in real time, making real mistakes here and there? Can you ever image a pub with live music created by a kid on a computer, staring at a screen instead of interacting with the audience? I can't. I play live music 2-3 times a week. I play an upright bass built 1937 and a Fender Precision electric bass built 1961. Apart from a tuner and an analog pre-amp, I don't use any analog or digital effects. I prefer paper music sheets. People come and listen. Because it's real. That gives me hope. Despite of AI overkill, people want to meet people,
Software Engineer
The Qt Company, Oslo -
-
@Chris-Kawa, we seem to have one interest in common that is at the same time heavily AI exposed: Music. Sure thing - the more genuine and human, the better. I can't stand these AI creators, but they are the tip of the iceberg. A perfectly human band releasing an album can trick you with AI before you even notice:
-
Harmonizing. Back in the day there were digital harmonizers, requiring heavy lifting for each song. Cher's "Believe", released in 1998 even used a harmonizer misconfiguration as an effect. Harmonizers and time-shift were also used to "fatten" a thin voice: You would copy the tack multiple times, use harmonizers to pitch each of them up or down by a hair. Then you would have to shift some of the tracks back and forth by a few msecs. It took ages and thousands of dollars for mixing wizards.
With AI, you can do it in your living room in 5min. And it sounds so much better. One can argue if that's actually fair. A singer without talent can throw his voice at AI and record something that just sounds great. -
Mastering. Once a band had spent the last penny on recording and mixing, someone would yank in their savings and overdrafts for mastering magic. It's all about cutting off level outriders, compression and EQ witchcraft with the ultimate goal, that Spotify showers you with songs at very similar volume levels. I have participated in a recording earlier this year, where mastering was fully replaced by AI. The result was simply stunning.
-
Modelling. I grew up an a world where we had to lift hundreds of kilos of Amps to every gig. Every bass and guitar player had wagon forts of analog effect pedals. Bringing down 230 V AC to 9 V DC for those pedals is a science of its own. Noise gates and filters were used to avoid the beloved pedals to sing the 50 Hz anthem. Witchers and sorcerers were paid fortunes for tube biasing and tweaking filter caps. Today you buy a modeler, that is able to learn your favorite amp, cabinet, pickup mike and all of your favorite effect pedals, while you have a coffee. It weighs 2 kilos at max and it sounds like 80% perfect.
There is a big "however", however:
Nothing is better than honest composition and a real concert.
AI chews up what has already existed. It's not creative at all, it just pretends to be. It's a tool to create mainstream. But nothing special will come out of it any time soon.
When it comes to composition, a high turnaround business, that's a competitive disadvantage. Listen to "Say A Little Prayer" from Aretha Franklin, "The Weight" from The Band or "Little Wing" from Jimmy Hendrix and you will agree that nothing that epic can ever be produced by an LLM.When it comes to equipment, the turnaround is not that high. AI modellers just give more options: I can change to a totally different amp/cabinet/mike just for a solo. And it gives these options to more people just because software is cheaper than analog hardware.
When it comes to live music, what's better than the real thing? People playing on real instruments with real voices, real talent in real time, making real mistakes here and there? Can you ever image a pub with live music created by a kid on a computer, staring at a screen instead of interacting with the audience? I can't. I play live music 2-3 times a week. I play an upright bass built 1937 and a Fender Precision electric bass built 1961. Apart from a tuner and an analog pre-amp, I don't use any analog or digital effects. I prefer paper music sheets. People come and listen. Because it's real. That gives me hope. Despite of AI overkill, people want to meet people,
@Axel-Spoerl Yeah, I'm not against technology or tooling in general. I'm against bad technology and tooling and bad usage of technology and tooling :)
Leaving music made with more traditional instruments for a moment I like synth for example, which, one can argue, is a completely "machine generated" sound. But it's a sound steered by a human imagination, talent and ear for what works and what doesn't.
AI songs are often just a mindless collection of "correctly" arranged sounds that mimic something the model was trained on. They often lack composition or structure. They sometimes change tempo or style mid-song without any rhyme or reason. Texts are just word salad without any meaning or story, voiced without any emotion in it. Boring and "correct" is the most praise I can ever give them. That's not what I want from music, or any art in general. Mayyybe if someone can't tell a duck from a chicken they can put on one of those 4h AI background coding music things, where it sounds like a random collection of beats and bangs you can throw together from a couple samples in a tracker, but that's what I would call a slop, and I really don't like that you need a server farm that burns a town worth of electricity to generate it.AI can't also create anything new. By the very construction of it all it can do is spew endless stream of something that sounds like something it was trained on. It won't suddenly come up with new sound or style. Sure, it can throw a bunch of filters in the mix, but that's not really invention. Same goes for image generation.
Again, I won't go into the means and morality of training data acquisition for the model training. I mentioned Anthropic settling a book piracy case. Well, they also have even larger lawsuit for pirating millions of songs for training. Nvidia just got caught pretty much "downloading youtube", bypassing all means of paying for the content, for training purposes.
I also never was a fan of harmonizers aka autotune, or voice filters in general, but that's personal preference I guess. I much rather a bit of breaking in the voice, that carries emotions and strain, over a pitch perfect, synthetically sounding voice. As for fully digital voices - what do I care when a machine glued sounds talk about love or hardships? It's like listening to a microwave talking about meaning of life. Go to a live concert and listen to a person playing and singing their heart out, voice breaking, tears flowing, improvised solos, small mistakes turned into epic moments. Tell me we can get that with a transformer model gluing sound samples together in a statistically correct pattern. Yeah, good luck.
-
-
POV:your vibecoder friend trying to debug the app built using Claude code:
