Programming thread

  • 🔎 If you see an alert for analytics.usips.org, this is a self-hosted system I am running as a test. It is GDPR compliant and collects no invasive information.
  • Want to keep track of this thread?
    Accounts can bookmark posts, watch threads for updates, and jump back to where you stopped reading.
    Create account
IDK where to put this but Dragonminded, the creator of Libdragon, a modern SDK for Nintendo 64 software development that I and everybody in N64 homebrew frequently uses, is quitting development for virtually all xir projects. Reasons include seething over the fact that xir software is no longer obscure and elite and only usable by 15 people thanks to AI crumbling xir ivory tower (N64 homebrew used to be an extremely small and autistic scene thanks to its lack of high-level programming options that is now experiencing an Eternal September (they fully deserve this BTW. They're insufferable.)) and not being accepted as a hekkin vivaldi traiso wompa. Bonus points for screeching about DHH.

@Null is right in saying that vivaldi wompas are seething because AI is making programming accessible to more than just freaks willing to spend months learning low-level function pipelines.

October 3rd 2026, 10:06 pm

I'm no stranger to open source. My earliest foray into open source software was a framebuffer library for the Nintendo DS. I uploaded the first version of that in August of 2005 and linked it around IRC for people to play with. That catapulted into an open source career that is frankly still hard for me to believe. Twenty one years of publishing libraries, homebrew for several Nintendo systems, multiple system toolchains with associated system drivers, countless utilities for Linux, several full server collections and even some open source hardware later and I can say that I've been quite prolific. Honestly, I'm incredibly proud of the things that I accomplished and until very recently was also very proud of the software that my own contributions nurtured.

I wish I could speak as fondly about the experience as I can about what I think is my legacy. The fact of the matter is that my time in open source has largely been miserable. If I'm being honest it has felt like an abusive relationship. The ecosystem is exploitative and the participants entitled, demanding and cruel. I've been harassed and bullied, had death threats lobbed at me, had my work stolen repeatedly and routinely been pressured to spend my free time on maintenance. By and large people have not given me any positive feedback for what I've released. Generally, the only time I hear about anything is when somebody wants a thing I made to do something different or they're reporting to me that it's broken. People seem to forget that I'm a person and only see me as a collection of useful code or as an obstacle to getting their way.

I sat down the other night and tried to enumerate all of the times people have contacted me to say thank you for something I wrote and it's honestly shocking. In the 17 years that I had libdragon up I've received one email that was positive. It was somebody who had published a game using the library and wanted to thank me with free copies of the game. Contrast that to the years where I tried to get anyone interested or to help and was met with beratement, insults and harassment. My BEMANI utilities repo which has been forked private to run half a dozen private networks has never received a single thank you in the six years it's been open source. Most of the feedback I get is somebody asking me to implement a new game. A few years ago when I released a Naomi Diagnostic ROM I actually got yelled at because people thought the word "ROM" meant I was giving them some super secret leaked factory cartridge, when in reality I'd just spent years reverse-engineering the system to be able to make things like that. The list goes on and on. It's honestly depressing.

Then there's the corporate angle. Open source is seen by much of the private sector as free labor to be taken from. They choose to build on top of open source and then come along to tell the people maintaining those projects that they're part of the supply chain now. Projects that creators never asked to become universally depended on suddenly turn into recipes for burnout, exploitation and stolen labor. Corporations don't give back in the form of help or even monetary contributions (unless you're an open racist like DHH, then they'll gladly publicly fund you to the tune of millions of dollars). If you are unlucky enough to get used by a corporation it's basically the software version of "working for exposure" except many of the people doing so have found that they can't get a job in the industries that depend on their software!

I ignored these problems, mostly happy to just play in the code of whatever special interest caught my heart at the time and thinking that I was doing my part by giving back when I released each of these projects. But after every single project I did things a little bit differently to try to avoid the intensely negative background noise, negative feedback and harassment. After DSOrganize I avoided GPL'd code because I realized that not only did it remove rights from me but it attracted the sort of entitled FOSS dude that was afraid of plants but quietly believed that women didn't have a place in software. After the Naomi netboot project and my arcade DOOM port I learned not to focus on trying to give a better experience to a community because they would just keep taking and taking until you burn out. When I stopped contributing to libdragon it was because I learned that people coding N64 didn't want open source if it meant they had to contribute as well and that they were happy to pirate official devkits until such time as I finished every single feature they demanded. When I gave up trying to improve MAME I noted that no good deed goes unpunishsed and that if you try to be helpful you will step on the toes of assholes who see you as a threat to their control. When I burned out on bemaniutils I learned that my mistake was making software that people who wanted fiefdoms to control could easily snatch and setup, filing the serial numbers off and claiming they did it.

And then came LLMs. A big thing I learned over my tenure in open source was that the best way to not get exploited for your labor was to make things that were both hostile to capital and useful to individuals. So I built things that could not reasonably be sold and would not be of any use to corporations but were still useful to people and communities. I worked on libraries and addons and utilities and services that were too niche to become the next log4j or the next LibCST yet still let individual people do cool things. I did that very intentionally to minimize the amount of drive-by labor theft but also in the hopes that some of the people who found it useful would connect and contribute. Sure, open source is full of people that will grab your library and just start using it, only to pop back in and demand bug fixes. But those were the people that I was publishing for. I would very much have liked them to give me positive feedback once in awhile and I would very much have liked people to collaborate with me instead of just throwing PRs at me. But I was mostly happy that people were using my work at all. LLMs changed all that. Tech figured out a way to extract value from my work anyway by scraping all of it to train their plagiarism machines and then sell that slop to people.

The thing about vibe coders is that they're the natural conclusion to where open source was going for a long time. These same people that feel entitled to the stolen valor of their slop commits previously felt entitled to your time and software. The people that are vibe coding things that look suspiciously like your software with the serials filed off were previously forking your code without so much as a "thanks". The same people that expect accolades for being a meat proxy were previously putting up PRs that didn't work and expecting maintainers to fix it for them. There was always a high degree of exploitation and theft in open source; LLMs simply ramped things up exponentially. But AI enables one thing that wasn't possible before: it's directly monetized that exploitation in the favor of tech. LLMs represent an astonishing theft of labor and concentration of wealth in the hands of a very few. The insult upon injury is that none of us gave away all of this work so that AI companies could steal it wholesale to sell to other capitalists.

I'm hardly the first person to point out these problems with open source. I am guaranteed not to be the last. There's really nothing I can do about it. I've been on the fence about giving up on open source for the last few months as I've watched project after project betray everyone by selling out to AI companies. Watching people tell you LLMs are the future while you sit here with twenty years of work stolen for that future doesn't sit well at all. I now realize that there's never been a point where open source wasn't miserable. I had just ignored that because I wanted to believe in a version of open source that was community focused and enabled all of us to lift each other. I wanted to believe in an open source that liberated computing while making space for marginalized people that are increasingly shut out of corporate spaces. I wanted to believe in an open source that was punk and embodied the hacker spirit. Sure, there have been points where open source has done that incidentally and a few projects that have intentionally built inclusive and rich communities. But it's clear to me that this was never the goal. The goal, as far as I can tell, was just to normalize giving away hundreds of hours of labor and I fell for it.

So where do I go from here? Well, open source has been a pretty big hobby of mine for quite some time. But if I'm being honest with myself, nothing changes if I simply don't release the source anymore. I'm still going to be coding things for myself regardless of whether I share them or not. Most of my projects have effectively zero people using them so the only thing they're doing being online is giving AI companies more free training data to sell back to everyone. I still believe you should have the right to inspect the code of things that run on your computer so I am not going to go back to only releasing binaries. I also think that community can be created around good pieces of software and not at the expense of the maintainer. However, that doesn't mean I need to make my work freely available. Over the next couple months as I find the energy I'll be taking down almost all of my repositories from GitHub. For the few that have contributions or an actual userbase I'll be looking into self-hosting git so that I can give access to the code to people that are actually interested in the work.

I'm tired both physically and emotionally. I feel exploited, used and betrayed. The trust I extended when releasing my work is completely shattered. It's finally obvious to me that this relationship is abusive, so I'm breaking up. Goodbye, open source. I don't actually think I'll miss you.



This follows after it was found out that Libdragon has had AI merged into its repo for a year before anybody noticed as Dragonminded is a lazy retard and flipped the fuck out in xir Discord.
1791264675400.png
 
Last edited:
Reasons include seething over the fact that xir software is no longer obscure and elite and only usable by 15 people thanks to AI crumbling xir ivory tower
I've stated many times in this thread that most programming knowledge is worthless, and something like HyperCard or Decker make programming about as simple as it should be even to people who lack more direct knowledge, but the neural network nonsense is different. This reminds me I forgot to remark to someone earlier, but now I can't find the exact post. Sure, Claude can write a script plagiarized from every script it's ever seen along some vast space nearby, but it's never going to remove the need for the script.

I've remarked before about the "intelligent idiot" who can spend his life optimizing code, like the strlen() implementation in the Linux kernel, but never remove it. I was chatting with someone else a few weeks ago about this same thing: Truly intelligent behaviour would be taking something huge and making it smaller but equivalent, or equivalent in all meaningful ways, which is exactly what these kikes and businesses expect to happen—somehow—before the house of cards collapses.
@Null is right in saying that vivaldi wompas are seething because AI is making programming accessible to more than just freaks willing to spend months learning low-level function pipelines.
It's not, really. Is it making writing more accessible by letting dipshits churn out entire books in an afternoon without understanding anything inside of it?

Also, it's good form to post the direct link and the archived link. I can't even access the archived link over Tor without Google telling me to bend over.
 
I still don't entirely understand what you're asking for, but if you're asking if you can host a web site by just pointing a web server at a directory of HTML files, yes. This is called "static hosting" now and is still a very common approach for sites that don't need to change often.
Sorry, I think I need to get some more sleep.

I'm just lamenting the state of modern technology continuing to add layers on top of a foundation that only makes sense in a high trust society that no longer exists while pretty much anything worthwhile is overshadowed by the mountain of a billion Indians.
 
Truly intelligent behaviour would be taking something huge and making it smaller but equivalent, or equivalent in all meaningful ways, which is exactly what these kikes and businesses expect to happen—somehow—before the house of cards collapses.
There's a lot of held breath right now among this exact type of people, increasingly leaning on an ever-more-desperate hope that the actual researchers/smart people working on LLMs will find some kind of breakthrough that will reduce memory or GPU/NPU consumption instead of just "improving" test scores or whatever.

Behind the scenes they know they're already hitting some big walls. There's only so much memory you can hand a computer before a new bottleneck appears that "download moar RAM!" can't solve. Same with GPU/NPU count in a single host. There's only so many lanes any given CPU can furnish, no matter how expensive it is.

The big race right now in the "state of the art" research is reducing memory consumption. They know at some level they've fucked up the RAM market pretty badly and the handful of manufacturers are cheerfully bending them over; there's a real sense they won't be able to keep expanding memory capacity and adding more nodes indefinitely. It'll get too expensive (RAM won't run out, it'll just get too pricey, then the RAM industry will go back to building slower consoomer-grade stuff again once the AI industry is done gorging). It's not pushback against data centers (lol that isn't slowing anything down, just screwing small towns out of jobs and revenue), but a realization that they're flirting with an upper limit on useful model size with current hardware (with no promising hardware architecture breakthroughs on the horizon that could change that).

Right now they're frantically trying to get "native 1-2 bit weights" working reliably (compared to the currently most frequently used 4-bit) and smarter "learned cache" sparsification. They're also trying to develop architectures that don't have memory requirements that grow linearly with context (that's a big killer right now too). There's other optimizations in the works too, but none will be as immediately impactful as those three factors.

The "house of cards" you describe will fall apart once demand exceeds supply (either because demand goes up too much or supply can't keep up as the models get bigger and start gobbling more and more hardware to run). That'll force them to increase prices (either reducing/eliminating free tiers and reducing tokens-per-dollar). Once the situation turns into "only the people who throw the most money at us can access our stuff because there's not enough to spread around anymore," it'll fall apart real fast.

I think you're right though. It'll break down before the smart people can fix things on the supply side with "tweaks."
 
There's something weird going on in the Godot subreddit where if your game isn't like some happy clappy troon game with a heckin' quirky mechanic then good luck getting help on a problem, you'll just get bumped off and told to go Google it.
My friend asked some question on Redot's discord. Something mentioning pregnant women and Floyd reactor... and they have happily provided an answer.
 
Claude can write a script plagiarized from every script it's ever seen
He says, using StackOverflow to copy entire functions throughout his entire life. No no, let me guess, you just knew those obscure coding paradigms from first principles?
 
He says, using StackOverflow to copy entire functions throughout his entire life.
No, actually. I'm aware that many have justified this latest nonsense because of that prior nonsense, but I always looked down on that too. In general, I'd argue someone who can't program without access to the Internet can't program. It'd be like a writer who constantly wants to check Twitter or some shit. Now, we can quibble over the exact line here, and books are fine, but I think we know what I'm saying.

I can program with a nice Lisp listener that contains access to its own documentation in various forms for when I forget the parameter ordering or whatever, and I can write with nothing but Emacs and my trusty Oxford dictionary. Yes, I've used StackOverflow to learn about some underdocumented shit, almost always to do with the WWW, which is yet another reason to hate said shit. I much prefer using languages and systems that have real manuals, not a bunch of people sitting around a fire to share tribal knowledge, which is my preferred analogy for StackOverflow.
No no, let me guess, you just knew those obscure coding paradigms from first principles?
Of course not. I read nice books instead. I'm reading "The Architecture of Symbolic Computers" by Peter Kogge right now, and it's very good, although the frequent typos are annoying. I recommended a bunch of books and papers in this thread I made. I've also learned quite a bit from arguing on the Internet with others.
 
Books? You presume those are the originals? Do you not think that those books are written after seeing other developers write code, or do you think that authors come up with their own work? If you're merely borrowing ideas from books, are you not a plagiarist yourself, or is this a special exemption that only exists for books?

It's all copies. Your brain works exactly like an LLM. The way they train the weights is through something called artificial neurons. We call them weights because of the biological synaptic weights.

Human creativity is blending up their own experiences. If you try to think of something random, you'll just reference something weird, not something novel.
 
Claude is already more valuable than basically every human being I see. The masses are worthless. There is no combination of life from India that I would preserve over Claude Opus 5.5
 
I should also remark on this:
Do you not think that those books are written after seeing other developers write code, or do you think that authors come up with their own work?
I believe men can create new ideas, yes. Most everything around us was an idea at some point.
There is no combination of life from India that I would preserve over Claude Opus 5.5
When I wrote that making things equivalent and yet smaller would be impressive, I actually had some Indians in mind as an example of the worthlessness of these machines. I rarely use YouTube but, in the few occasions I do, I've noticed Indians will make videos that are basically plagiarized slideshows set to some ugly music. They've taken something small, and worthlessly expanded it on Google's dime. This is basically what the neural network nonsense does too.

This machine called "Claude" isn't going to take a program that's one million lines and turn it into one thousand anytime soon, now is it?

Anyway, we can agree on the worthlessness of most "human" life, sure.
 
My God... Jersh has rejected his sneed, his divine spark. He has become Null, rejecting the warmth of flesh, rejecting the smile of compatriots and the thrill of jolly cooperation, consumed by the brilliant yet cold fires of statistical text models. He no longer applies his devine spark to give meaning, creation to his LLM. The LLM gives meaning to his conciousness. He has been written, he is the prompt.

The mutt slobbers in darkened glow. He has become machine.

But seriously, LLMs are quite poor at best and harmfully trash at worst for programming-needed R&D except for hyper accelerating support microprocessor programming and websocket communication shit. I haven't read my fucking data collection code in months.

It is still entirely worth learning electrical engineering and programming from fundamentals.

LLMs are VERY, VERY poor at R&D hardware, data analysis, and human-suitable GUIs for purpose. It is entirely trained on corporate business data analysis code which has no human value and is a lie meant to look good to scam a white guy into giving jeets money. The text models have no proper concept of time, where the code is meant to handle external events occuring at intervals. You can attempt to beat it into shape with a daisy chain of endless prompts but they will always need human intervention. I basically have to manually program all of those tools but do rely on LLMs as a boilerplate code repository.

BUT, LLMs can NOT cope with software that has had a decade with multiple api's or is hyperniche in a closed community. Like I said, no concept of time. It can and will use V0.3, V1, V2, V3... Vx api interchangably. You have to directly prompt each fix. And LLMs WILL light your house on fire if you try to follow it for hardware.

Chatbots are useful but they can't replace real minds doing real R&D because the real world can not be replicated by tokens. The needs of now are not circumscribed by tokens of yesterday, and novel inspection is entirely unwritten. All of the AI math and science discoveries are because they spied on every professor and engineering firm on Earth then put together 1000 competing men's already partially correct solutions into a Frankenstein mass of their work.
 
No no, let me guess, you just knew those obscure coding paradigms from first principles?
Many useful principles are easy and reinventable, if you spend enough time writing code.

Claude is already more valuable than basically every human being I see.
I work with a guy who thinks the same way. Weird how this magical instrument that reportedly makes people obsolete, writes such a shitty fucking code that all of it's output needs to be double and triple checked by a bunch of other people.
 
Weird how this magical instrument that reportedly makes people obsolete, writes such a shitty fucking code that all of it's output needs to be double and triple checked by a bunch of other people.
Oh man human code never needs review! When I worked professionally we definitely didn't alternate between reviewer/reviewee all the time! Those reviews didn't catch problems every single time either. And the humans NEVER got defensive about their poor design decisions, leading to workplace drama and hurt feelings fucking routinely.

Face it buddy-boy, Claude is already faster than 100% and better than 99% of developers.
 
Opus 5.5 is insanely good, and those who do not agree are ignorant. Still, an incompetent human prompter/director will create a bad project. The code itself will be good regardless, but it might not accomplish the goals optimally due to badly set constraints, and have a structure that can not accomodate future changes well.

This machine called "Claude" isn't going to take a program that's one million lines and turn it into one thousand anytime soon, now is it?
Try it out, and be prepared to change your mind. (Or try OpenAI's model, I'm not a paid shill)
LLMs are VERY, VERY poor at R&D hardware, data analysis, and human-suitable GUIs for purpose.
It basically one-shotted a great GUI for a slightly esoteric project I'm working on.
It is still entirely worth learning electrical engineering and programming from fundamentals.
Yes
 
Oh man human code never needs review! When I worked professionally we definitely didn't alternate between reviewer/reviewee all the time! Those reviews didn't catch problems every single time either. And the humans NEVER got defensive about their poor design decisions, leading to workplace drama and hurt feelings fucking routinely.
All the code must be reviewed. But there are people in my current team who are competent enough, so that there is no drama and the amount of time that you are spending reviewing their code is much smaller than that of LLM-code. In fact, all the major problems we have are brought upon us by the niggers that are big time LLM users.
 
I used to sometimes meet people who insisted computers with their fancy new "word processing" would never replace their trusty typewriters. Typing on a screen could never give writing the same soul as you'd get from clattering away on an old Olivetti.

That's exactly how anti-LLM people sound to me now.

God forbid that people should use a computer to automate tasks.
 
All the code must be reviewed. But there are people in my current team who are competent enough, so that there is no drama and the amount of time that you are spending reviewing their code is much smaller than that of LLM-code. In fact, all the major problems we have are brought upon us by the niggers that are big time LLM users.
>poor developers submit poor code
my god, nuke anthropic immediately.
 
Dear faggots:

What is the fucking point of what you want to do?

a) Toil on the syntax?
b) Design the system, which AI still can't do?
c) Make something useful?
d) Use something useful?

If the answer is a, touch grass.

I used to sometimes meet people who insisted computers with their fancy new "word processing" would never replace their trusty typewriters. Typing on a screen could never give writing the same soul as you'd get from clattering away on an old Olivetti.

That's exactly how anti-LLM people sound to me now.

God forbid that people should use a computer to automate tasks.

Admittedly, the number of people who repeat something with an LLM instead of having it write a script to do that more than once is high.

But seriously, LLMs are quite poor at best and harmfully trash at worst for programming-needed R&D except for hyper accelerating support microprocessor programming and websocket communication shit. I haven't read my fucking data collection code in months.

Nigger, source your shit. Where is a frontier, or even a good Chinese, AI actually 'poor'? What are YOU doing with it? What are you pointing at that someone else did?
 
Last edited:
Back
Top Bottom