Programming thread

[DavidS877]

is this shit even english

i like llms but what the fuck am i reading

Presumably the output from an LLM.

 

[Nice Stick]

Good? Bad? Satisfying or not? Does it matter? Did it ever matter? Can you imagine a future where compiled binaries are diffused from noise just like how images are generated right now, skipping programming languages entirely?

That's the craziest thing I've ever read that might actually come true. Truly mind bending.

 

[Sneed's Feed n Seed]

I like telling the AI to type faster than me so I can shitpost.
I like telling the AI to first-pass shit it can actually reliably check and then I do finer grained things.

I also like making the AI say nigger 50 times.

 

[Private Tag Reporter]

Over the course of two days, I have created an app to the point I'm confident enough to hand it to my friends and ask them to test it. Previously, something like this would have taken two months, and I would have lost interest half-way through because of various configuration retardation and just finding a shiny new project to work on.

LLMs are just a force multiplier for competency. It just so happens that the majority of people using them have a negative coefficient.

 

[macrodegenerate]

I had an unemployed software Dev. friend tell me today that he refuses to use AI and never will as a programmer. He's been out of a job for 2 years now, and while I respect the sentiment, I'm worried he'll never be able to find a job in the industry.

 

[analrapist]

I'm worried he'll never be able to find a job in the industry.

At 2 years, he's probably not going back, especially if he refuses the LLM-pill, at this point. He'd have to be employed and practiced to compete with the LLM force multiplier.

 

[macrodegenerate]

At 2 years, he's probably not going back, especially if he refuses the LLM-pill, at this point. He'd have to be employed and practiced to compete with the LLM force multiplier.

Blaming the career gap is a quiters mentality. It's learned helplessness. There's generally a way back into your industry if you've taken a break. It's just going to suck as you eat shit for a year or two at a lower wage or having to make some sacrifices. After that you're fine though.

He really is screwing himself though, and needs to lose everything sooner rather than later instead of the slow decline.

As much as they want AI now, the programmers who came out of school before it are becoming more of a sought after commodity. There's a difference between knowing how to do something, and having your AI do it for you, and having your AI do something for you because you don't know how to do it. The later will always produce architecture inferior to the former.

 

[SchizoDaemon]

I had an unemployed software Dev. friend tell me today that he refuses to use AI and never will as a programmer. He's been out of a job for 2 years now, and while I respect the sentiment, I'm worried he'll never be able to find a job in the industry.

What is there to respect about refusing to learn how to use something? It's not going to be un-invented.

 

[deleuzeon]

I had an unemployed software Dev. friend tell me today that he refuses to use AI and never will as a programmer. He's been out of a job for 2 years now, and while I respect the sentiment, I'm worried he'll never be able to find a job in the industry.

Then he should probably write some software that people want to buy. But if he wants to collect a big salary just for existing at a company regardless of actually accomplishing anything, the way tech people are used to, then you have to do the monkey dance for people with manic AI fixations.

 

[mnt_ms]

I had an unemployed software Dev. friend tell me today that he refuses to use AI and never will as a programmer. He's been out of a job for 2 years now, and while I respect the sentiment, I'm worried he'll never be able to find a job in the industry.

I doubt that his conviction not to use LLMs is the main reason for him being unemployed.
It's not like LLMs do some crazy shit that you can't do with your head alone.
I myself do use them occasionally, but only for guidance - all the code I submit is written manually. For example, I know jack-shit about modern C++, and it's a bit faster to ask a question than to search for an answer in the literature. It's easy to read about some language feature and understand it, but it's a completely different story when it comes to some tricky stuff that needs to be done via those features.

Most of the mundane stuff I need to do is easily done with a little bit of typing and some vim macros - no need for LLMs. Furthermore, if you use some normal fucking languages in your day-to-day job, you also do not need LLMs. They are made exactly for those complex monstrosities like C++.

Maybe there will be a day when I am versed enough in Bjarne's brainchild, but I doubt that this day is in the near future.

 

[Meth Reindeer]

I had an unemployed software Dev. friend tell me today that he refuses to use AI and never will as a programmer. He's been out of a job for 2 years now, and while I respect the sentiment, I'm worried he'll never be able to find a job in the industry.

Others have said it, but he might need to do the song a dance about AI, just a little bit to get himself hired somewhere modern that isn't explicitly anti-AI. I had to do a coding test on an interview recently where they asked me to use AI. It was weird, but much easier than I expected. A lot easier than doing stuff by hand or some kind of take home test.

He should probably build some kind of portfolio and apply for a more junior position somewhere with that gap. I've never been out of a job that long since I was old enough to work though, so I'm not sure how it goes. Presumably, he's not retarded?

As much as they want AI now, the programmers who came out of school before it are becoming more of a sought after commodity. There's a difference between knowing how to do something, and having your AI do it for you, and having your AI do something for you because you don't know how to do it. The later will always produce architecture inferior to the former.

I'm trying to get this kind of idea into junior developers heads. Learn what the fuck you're doing so you can tell your AI what to do more clearly, correct its mistakes, and not need it 100% of the time to work.

 

[ADHD Mate]

I am honestly baffled how many of you love LLMs for writting code. My experience is that, sure then can generate working code, but they will also generate seemingly good code that's just enough subtly wrong that understanding what went wrong sets you back so much, that writing it from scratch would probably be faster.

Not to mention that having to review it anyway makes gains negligible, even in best case scenarios, at expense of having to do more unpleasant work (reviewing) over one that's more fun (writing).
Not to mention that writing code after feature specification is agreed on is usually trivial.

I personally only use LLMs as semantic search, kind of like a more context aware grep, it makes code exploration a lil bit faster.

 

[mnt_ms]

I personally only use LLMs as semantic search, so kind of a more context aware grep, as it's make code exploration a lil bit faster.

One more thing I use LLMs for is to ask what kind of tense to use in a sentence, when I want to express a particular idea, or to reassure myself that the tense I picked is correct.

No offense, but that grammar you guys have is kinda retarded.
Still a beautiful language, though. Vocabulary in particular is a blast.

 

[ADHD Mate]

One more thing I use LLMs for is to ask what kind of tense to use in a sentence, when I want to express a particular idea, or to reassure myself that the tense I picked is correct.

I would gladly install grammarly if there was FOSS version.
For now you have to excuse my ESL.

 

[DavidS877]

one that's more fun (writing).

Writing a ton of boilerplate to make a GUI is not fun, it's cut and paste. Asking the LLM "make 3 columns, 2 buttons per column, here's what each one needs to do" and then wandering off for a while is good.

Now, ripping a C++ program apart to make it able to use an I2C GPIO instead of hard-wired GPIO pins, that's fun. Now I need to chat with the LLM to do the emulation of the I2C GPIO on the RP2350 on the other end and decide who is going to write that code. Might use it as a first test of Qwen 3.8 27b.

 

[analrapist]

I am honestly baffled how many of you love LLMs for writting code. My experience is that, sure then can generate working code, but they will also generate seemingly good code that's just enough subtly wrong that understanding what went wrong sets you back so much, that writing it from scratch would probably be faster.

I'm much faster at reading code than I am at writing it. LLM output often does require aggressive refactoring, it's true, but refactoring working code is easier for me than getting to working in a lot of cases. I'm good at reading docs, but often they don't answer my questions.

But I use local and free models exclusively, and I hear a lot of talk that people get better results from the expensive American models. I'm quite fond of the Chinese models.

For another example, Con Kolivas has recently revived his old Linux kernel scheduler because LLMs take the drudgery out of keeping the code up-to-date with the kernel tree. https://lore.kernel.org/lkml/CABqErrH=oQ3povVuSPhRON97v63=mB85jQmZjf443ofdYAuxxw@mail.gmail.com/

I think of LLMs as accessibility patches for specific cognitive deficiencies. That's how it functionally works for me.

 

[SIMON THE PIEMAN]

I am honestly baffled how many of you love LLMs for writting code. My experience is that, sure then can generate working code, but they will also generate seemingly good code that's just enough subtly wrong that understanding what went wrong sets you back so much, that writing it from scratch would probably be faster.

If you'd ask me a year ago about coding with LLMs I'd probably give the same answer. Nowadays my opinion is that the paid models can do truly amazing things and that the $20 subscription is well worth the price. My opinion will probably shift again once the investor money dries up and they'll be forced to jack up the prices through the roof. For now though I will continue doing incredibly dumb and stupidly fun shit while I imagine Sam Altman begging me to stop burning tokens on the coding equivalent of a shitpost.

 

[Least Concern]

I am honestly baffled how many of you love LLMs for writting code. My experience is that, sure then can generate working code, but they will also generate seemingly good code that's just enough subtly wrong that understanding what went wrong sets you back so much, that writing it from scratch would probably be faster.

Thank you. It's really starting to feel isolating to be someone who still enjoys coding by hand, not just here but on HN and LinkedIn and pretty much everywhere else coding-adjacent nowadays.

I guess this LLM revolution is creating a clear divide among programmers that couldn't really be seen before; those who truly enjoyed the puzzle and problem-solving aspects of coding, and those who thought the coding was just a means to an end. Unfortunately those of us in the former category seem to be in the vast minority if online discourse is anything to go by.

Frankly it's gotten to the point that I'm starting to consider new career paths despite me being way too old to do that. The robots aren't going to replace cattle ranching yet, are they?

 

[SchizoDaemon]

I am honestly baffled how many of you love LLMs for writting code. My experience is that, sure then can generate working code, but they will also generate seemingly good code that's just enough subtly wrong that understanding what went wrong sets you back so much, that writing it from scratch would probably be faster.

No it is often slower. I can iterate faster with an LLM.
I can often have the same LLM work on many features at the same time, while I solve other problems.
If I could afford another AMD AI 9700PRO (which won't be for a while). I could have do more parallelism.

Most of this requires research on how to do and implementation is often the easy bit.

Not to mention that having to review it anyway makes gains negligible, even in best case scenarios, at expense of having to do more unpleasant work (reviewing) over one that's more fun (writing). Not to mention that writing code after feature specification is agreed on is usually trivial.

Writing isn't that fun anymore since I am normally writing very boring code. I've been writing code since the early 90s (started when I was 9/10 years old).

I guess this LLM revolution is creating a clear divide among programmers that couldn't really be seen before; those who truly enjoyed the puzzle and problem-solving aspects of coding, and those who thought the coding was just a means to an end.

This is quite frankly bullshit. I enjoy the problem solving aspect. But professionally I never asked to solve anything more complex than "when I click this button, thing is bringing up this error". It is very boring work e.g. find error in AWS console, work out what line, write a test, PR, review and then deploy. It is fucking tedious.

Computers are very good at doing tedious work. It is not good at doing work that requires creativity and problem solving. Tedious work gets shifted to the AI. Most of my work is tedious so most of my work get done by the AI.

 

[Sneed's Feed n Seed]

I am honestly baffled how many of you love LLMs for writting code. My experience is that, sure then can generate working code, but they will also generate seemingly good code that's just enough subtly wrong that understanding what went wrong sets you back so much, that writing it from scratch would probably be faster.

TDD. Read it. Use a good one.