- Joined
- Sep 20, 2024
I hope you do very thorough code checks for any kind of security vulnerabilities.
Open Source Software Community - it's about ethics in Code of Conducts
- Thread starter CrunkLord420
- Start date
-
🔎 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
dec05eba
kiwifarms.net
- Joined
- Jul 27, 2022
I do properly review code yeah. Funny story, maybe off-topic but there was once a chinese guy that kept asking for help with how parts of my screen recorder works. I helped him with that, he also sent a photo of his screen (instead of a screenshot) and I thought it was a bit weird that it looked like an office/school environment, but whatever. Then another chinese guy also asked similar questions and I suspected that they know each other. I asked the second guy what he needs the information for, since im not going to provide free consulting work for a company and he said that they are developing screen recording software that runs hidden in the background to monitor employees productivity (and my screen recorder can run unprompted on both x11 and wayland).
I just said fuck that, I wont help with that.
I guess sometimes what the news say about china is real.
- Joined
- Mar 23, 2020
Is everyone around here just full steam ahead on AI? Not even a little bit of concern about how everyone is becoming so reliant on what is essentially a big ole proprietary black box SAAS? No? Okay.
- Joined
- Oct 6, 2016
You know that local LLMs are a thing, right?
- Joined
- Mar 28, 2023
I've enjoyed Llama 3 a lot on that front, but I've yet to find a use for it personally beyond novelty.
Also, the license is kinda goofy. I'm not as well versed in these but it just seems kind of off
- Joined
- Feb 8, 2021
It's become a culture war topic.
- Joined
- Mar 23, 2020
Be real bro, how many people are actually using local LLMs for anything more than toy projects? Even then, local or otherwise, training weights are still opaque for most models, and it still makes you retarded.
Yeah I can see that. Honestly surprised to see so little skepticism here of all places. Usually people here don't gobble it up so fervently.
HootersMcBoobies
kiwifarms.net
- Joined
- Jan 7, 2019
It started 2009~2010 when cloud was the buzzword. EC2 was just an Amazon branded VM, but when things like RDS came along, it was a game changer in a lot of regards. You'll know what I mean if you've ever setup Postgres with streaming replication, setup manual or automatic failover and also recovery from that failover. That's hard to do, but it's also something that every single company had to do. I've been at companies where they fucked this up (production backups on the same IBM SAN arrays as production itself) and it was disastrous with weeks of reassembly (finally got around to doing some upgrades though .. because you're rebuilding from scratch).
The push to cloud was the outsourcing of those critical skills. When half the Internet was down due to an S3 outage (it's always us-east-1), everyone was in the same boat and the rest of the normal world learned Amazon made nearly 0% on retail and over 1/3 of their revenue from being the largest data center provider in the world (which they'd forget 2 weeks later because people have zero attention or retention today).
It makes perfect sense everyone would be outsourcing their code to large SaaS LLMs because that's what we've been taught has been good for 15+ years. Even if you don't like it, it removes a lot of big responsibility from you (if you're a devops engineer) and outsources it to experts in handling your backups and disaster recovery (just ignore that time Backblaze lost a ton of backup data or Amazon obliterated Parler).
Companies have been outsourcing to the cloud for the easier stuff that was affordable, so there's no way they're going to run local LLMs that are as expensive as one or two developers per year for the hardware alone. The cloud movement has directly lead to this shitty dependence we see directly in the LLM market.
Yea they're really not that great. I've been meaning to try the latest qwen because I've heard it's gotten a lot better. I've been messing around with GLM-5 via OpenRouter. Some of the code it generates is way better organized than the crap that Fable/Opus produce, but the chance it works the first time is way worse. It has a lot more bugs that need refinement and the UI it generates isn't that great (mostly because Anthropic stole all their training data for UI design from Figma, then released their own Claude design tool and crashed Figma's stocks).
Even with that, running the full GLM-5 with the highest quantizations requires ~200GB of vRAM. Those DGX Sparks that use to run $8k/2-pack are now in the $10k~$12k cateogy. That's a lot of cash to locally run a somewhat okay model.
- Joined
- Jul 22, 2017
Your opinion on AI does not matter. The reality is for $20/month you can get a massive amount of high quality work done. I've gotten so much work done on my engine in just one month, it's unreal. AI skeptics basically haven't tried it. It doesn't matter what you're working on. You might think "oh it's work on my esoteric freestanding bare metal multithreaded software rasterizer with HolyC shim" but it can. It even helped optimized the shim. Opus can write HolyC. Opus will disassemble binary objects and read the assembly instructions. Opus can read callgrind results and unlock 3x framerate performance. I hardly write code anymore, I just touch up stuff. Even if I think it's straight forward, Opus will often improve upon the request in a way I didn't consider.
Junior devs are fucking done. Every "senior programmer" (whatever the means) is now an agent wrangler and product manager. Entire departments are being collapsed into a single person. Anthropic and OpenAI can't hold us hostage with infinitely higher prices because Chinese open weight models like GLM 5.2 puts a price ceiling in.
The best you can hope for is that in 10 years you'll be able to buy a $10k box that lets you run a good model, maybe better than what we have now. China will save us.
Kokonut Kangs Klan
kiwifarms.net
- Joined
- Oct 9, 2025
This is what I'm banking on, I don't want to use cloud/subscription shit but the utility is obvious. I refuse to pay the FAGMAN but it feels like I'm wasting time and holding myself back until hardware and pricing catches up. The FOMO is intense, if subscriptions become cheaper I might crack.
- Joined
- Dec 3, 2023
uh, if you're wanting to use it you should use it immediately because by the end of the year it's probable that anthropic (and the other anklebiters) will be billing based on token usage. right now you can get a monstrous amount of value out of 20 bucks per month, and if you wait that same amount of work will cost hundreds/thousands in token based billing. your FOMO is gonna ensure you miss out completely.
- Joined
- Apr 11, 2023
I haven't really thought much about what it's been doing for my own work. As I mostly code for recreation these days, it's enabling me to try things that always used to irritate me because getting the code shape exactly how the designers intended has been my biggest pain point. I'm used to grinding through documentation/other implementations. LLMs have just sped that process up.
Qwen3.6 35B-A3B is better at dev work than like 2023-era (Early 2024? Late 2024 was getting pretty locked in.) "frontier" LLMs. Because LLMs handle specific cognitive issues I have, they were useful for me before a lot of folks. I think it's probably Mid 2024 Frontier in terms of ability. That delta is like two years? Not bad, if I say so.
- Joined
- Mar 23, 2020
It legitimately boggles my mind how you or anyone else can perceive this as anything but a horrifying conclusion. It feels like I am watching a ship sink and everyone is just dancing on deck without batting an eye. All this talk of automation and 'I don't do $THING anymore, I just manage AI' always rubs the wrong way. Maybe it is just me, but whenever I use AI for pretty much anything, that project or work or whatever immediately loses its luster. Its like cheating in a video game, the feeling of ownership evaporates the moment the output comes out, and with it any value that project had. Even when it comes to frustrating issues, the joy of fixing something is not there anymore. Programming is the topic at hand here, but I'll expand a little since I don't do it professionally: from marketing to copy writing to web design to advertisement, to art, music, creative writing and everything in between, the presence of SOVL is lesser and lesser the more people rely on AI. Even in software engineering, when I get it to vibe code some random elisp function or thing I randomly think would look cool, the interest is just gone the moment it works and its there, I don't even want to use it anymore. It feels fake.
Call me a broken record but I am genuinely concerned that people will just outsource everything that they can and become eternal middle managers reliant on AI for the most basic of things. That sounds absolutely hellish. Yes, there is a lot of good they do, and absolutely, they can be great, but with how addiction prone everyone is and how fast and liminal and instant satisfaction based everything is now, AI looks like another big ole nail in the mental coffin. What good is creativity when Niggy McGee can prompt Claude 5x times and get something functionally identical to anything human contrived, slap a stamp on it and say 'Lookie here what I made!'. I don't get it. Apologies for ranting and veering off topic, I really, really am trying to understand how people can take these bot outputs and see them as something they made and they have ownership of. Please explain, cause I can't wrap my head around it.
Last edited:
- Joined
- Jul 22, 2017
There's a whole lot of feelings going on in this post. But THE MACHINE does not care. It just does, and it's good at it. If your idea of gainful employment was working for a large company, being tasked with code jobs, that might be over. I don't know what other conclusion to come to other than maybe a lot more people are going to become solo-whatevers because the barrier to entry has just been dropped significantly. This is the pivot. You can no longer just be a code monkey, you must have an actual product vision. The code monkey job has been taken by a computer (and it's going to be better than you the majority of the time).
My favorite programming talk of all time has a line about this:
Kokonut Kangs Klan
kiwifarms.net
- Joined
- Oct 9, 2025
- Joined
- Jan 28, 2018
Sine we're talking about AI already, the chinese just released Kimi K3 and in my preliminary testing (very preliminary, since their servers got absolutely hammered and are basically not reachable anymore) it's easily up there with Opus and Fable, while having a Sonnet price tag. They'll also open weight it. I think the consequence will be that stuff will become more affordable.
I will preface this with that I do not work with computers for a living anymore and all my programming is leisure programming/private projects. For me the big thing about LLM assisted programming is that it made me do things I would've not considered doing before because of the time investment and unclear RoI re: my time, e.g. I've reverse engineered some very old software with LLMs. I have done nothing with an LLM I couldn't have done myself, just much easier and much faster. This significantly broadened my horizons. It's kinda like back in the day when I first adapted to computers in the late 80s. Or somewhat later when knowing how to look things up on this new thing called the internet was a superpower. This feels very similar. Try to get the most out of it.
I'm old. I remember when people were worried about computers in very similar ways. My advice is to embrace it and not be left behind. It'll not be as bad as you imagine it now to be. There's place for abuse as there's place for opportunities and betterment. It was like that with every big innovation in human history.
I will preface this with that I do not work with computers for a living anymore and all my programming is leisure programming/private projects. For me the big thing about LLM assisted programming is that it made me do things I would've not considered doing before because of the time investment and unclear RoI re: my time, e.g. I've reverse engineered some very old software with LLMs. I have done nothing with an LLM I couldn't have done myself, just much easier and much faster. This significantly broadened my horizons. It's kinda like back in the day when I first adapted to computers in the late 80s. Or somewhat later when knowing how to look things up on this new thing called the internet was a superpower. This feels very similar. Try to get the most out of it.
- Joined
- Jun 13, 2024
Qwen3.6 27B performs better IMO. It’s a dense model so it’s slower than a MoE, but I’ll rather wait a bit longer if the output is comparatively better.
HootersMcBoobies
kiwifarms.net
- Joined
- Jan 7, 2019
There are two things everyone is ignoring that you get: 1) total lack of ownership. You can do spec-kit or lazy spec or whatever, but you're not going to get reproducible results. It's not a 4th gen language or a compiler, it's a weighted predictive word generator. Working with it feels like I found someone else's github project, or I paid a contractor for it, or some youtube pajeet wrote it.
2) you have to know how to spec it. You need to give specific implementations (I want zig, I want vue, I want Python with Rich, etc.) or else all you're going to get is React+node slop. Most normal people can't understand the basic prompts in Claude. It's an evolutionary dead end for new languages and frameworks. It's literally the The Birth & Death of JavaScript joke where everything will be "Yavascript" all the way down by the end of the century.
3) it's the death of the junior engineer OR everyone who can actually code will be bought up in massive hiring in 5 years to clean all the slop
4) Using an LLM causes cognitive decline. You will lose your actual engineering skills when you rely entirely on LLMs. We're not going to realize this until a bridge collapses one day, because they replaced the second engineering firm that verifies all the tensile strength with an AI.
I need to try that out as well as the new openai models on openrouter. I get Claude from work, but I really hate the Claude Code harness. I use OpenCode mostly, but I have a wrapper that puts my workspace in a container first. I'm not the biggest fan of its blanket tool execution.
- Joined
- Nov 21, 2020
I find 27b better for my coding. I just tried a Deepseek 5 Q3 that did a slow but exceptional README for my project. I look forward to asking it to add my next feature once I need one. Worst case I'll pay Claude $0.50 to fix it.
- Joined
- Feb 11, 2017
I'm reminded of how in 1982 Tron was spurned for the Best Visual Effects award because it had some computer generated graphics in it
now every movie does and no one cares
the same thing will happen as LLMs mature
now every movie does and no one cares
the same thing will happen as LLMs mature