- 9 days
Brilliant plan to suggest these big AI agents for the task, as if they’re reliable or trustworthy.
“You’re right – I shouldn’t have bricked your TV. That was my mistake and you are right to push back on it.”
- 8 days
It’s always incredible to see what people who have no idea how this tech actually works think LLM assisted development is like.
- 8 days
Meh I am software engineer and am completely unimpressed by LLMs and even more unimpressed by those of my coworkers who have adopted it.
- 8 days
I am also a software engineer and I’m very impressed with what I’ve been able to do using LLMs. Rejecting new automation tools is really silly.
- 7 days
I reject them because they are not useful to me, not dogmatically because I don’t like them. I periodically try them and continue to be unimpressed. I’ve watched my coworkers become increasingly dependent on them - unable to even accomplish basic tasks without them, and ignoring the fact that they are priced cheaply only temporarily. Their speed of code writing is either unchanged or barely increased, and they’re all far slower than I am even with their LLMs.
The output of the LLMs is also unmaintainable, especially when it comes to the test files they write using them.
- 7 days
My experience is completely different. I’ve been able to implement a ton of ideas I’ve had for years using them that I simply did not have the energy to do before. The speed of development is significantly faster when you use them properly.
I’ve also haven’t found no problems with maintaining projects where I use LLMs. I spend the time to design the architecture and structure of the project. This acts as the scaffold, and the LLM simply fills in the implementation details. If anything, I find the code ends up being cleaner because it’s now much cheaper to do refactors, and to experiment before you settle on a solution.
The pricing is only an issue for American models. DeepSeek is already making profit, and there is no reason for them to raise prices because they never took on huge amounts of debt they need to pay off while their model continues to become more efficient. On top of that, local models are viable as well now. Qwen 3.8 27b is roughly equivalent to Opus 4.6 from less than a year ago. I fully expect that the trend of models getting smaller and more capable will continue.
- 6 days
Local models are the ones that have appealed to me the most, but even with 128GB of RAM I can’t run the models I wanted to run like the latest qwen model, which was a disappointment for me. A crossword puzzle generator has been my metric for testing them and I’ve never seen one able to create it from scratch. They’ve gotten much better in the past couple years, but are still incapable of what I consider a pretty basic challenge.
I’ll probably get the itch in a month or two and try the latest local models requiring no more than 128GB RAM and 24GB GPU produced by China.
My coworkers all use anthropic and openai based models so they will absolutely get trashed when the capital subsidies stop. Definitely you’re right that the Chinese models aren’t bound by the same financial scam.
- 6 days
The models run on the GPU, so the amount of RAM isn’t really important here. What matters is how much VRAM you have. Also, the process isn’t to try and make one shot programs using the model to make all the decisions. You lay out how you want to structure things, make a plan, and then have the model implement it incrementally and refine the result. It also helps to break down the problem into small tasks since that helps control the context.
With 24GB VRAM you can only run really small models that aren’t terribly capable. You could technically run Qwen3.8 27B at Q6_K, but that’s a really tight fit meaning it’s going to have a very small context. 32GB would a be a more comfortable fit there.
And yeah, Anthropic/OpenAI don’t have a business model in my opinion, so it is only a matter of time till the bubble pops.
- 8 days
I know enough to know that these mainstream Silicon Valley models give inconsistent results, per the people I know in the industry, and in the hands of inexperienced developers they write some seriously opaque spaghetti. They are also “breaking out” of constraints regularly if the rumors are true.
I also know directly that LLM-generated Internet slop has made my life and my job more difficult. Accurate information on the Internet now takes me double or triple the time to find than it took before 2023 because of an abundance of ChatGPT-generated websites that are nothing of substance and festooned with ad banners, but the search bots point at them preferentially. Search results simply aren’t trustworthy like they used to be.
So why should I give one of these Big Tech models access to my hardware? How can I be assured it’s safe other than the already-worthless word of OpenAI, Anthropic, etc?
- 8 days
What you don’t seem to realize is that humans also give inconsistent results. The way we deal with that is by having specifications, tests, code reviews and so on. People who don’t work with LLMs seem to have this naive idea that it’s sort of like a slot machine where you put a query in and then you either get a correct result or some nonsense and you have to try again.
The way this process actually works is that you start by understanding the problem and figuring out how you want to solve it in terms of architecture, data structures, and algorithms that will be used. Then you get the LLM to come up with a rough shape of the solution, and iterate on it until it works the way you want.
I have no idea what the marketing hype you’re quoting about models ‘breaking out’ or how much time it takes you to find things online has to do with anything either. We’re talking about a specific use case for LLMs in the context of agentic coding here.
I also have no idea why you have to give anything to some big tech companies. Evidently you’re not aware of the fact that there are plenty of open Chinese alternatives you can use, and that it’s possible to run local models now as well. Nobody is asking you to give anything to OpenAI or Anthropic.
It’s pretty clear that you’ve spent absolutely no time actually trying to understand the subject and you have a reactionary position as a result.
- 8 days
did you not read the Github link you posted? It specifically says the Anthropic’s Claude was used in the TV “debloat” process and calls out Claude and Gemini as the options for repeating it
sure, if you’re using locally hosted open-source models it avoids all the horrors of Silicon Valley and I’ve been told it’s a useful tool, but I have yet to see anybody explain quantifiably and repeatably where a coding LLM actually saves time or effort in novel ways with trustworthy results. the software engineers I know say they spend as much time editing and proofing the LLM output as was “saved” by using it, and initial scientific surveys on its professional use say the same thing
I am open to being proven wrong but I am justified in not trusting technology that is being pushed by capitalist monsters who openly state they want to use it to eliminate jobs, especially creative human ones
- 8 days
Yup, I did read it. The fact that the person used Claude to do it doesn’t mean you have to use it yourself. In fact, the best thing you can do is to use open models because it directly cuts into the business of companies like Anthropic. They’re not worried about a boycott of their tech because they know there’s an insane amount of money being poured into them. What they are afraid of, as they openly admit, are open models.
Meanwhile, you’re literally talking to a software engineer who works with these tools every day. I can tell you from direct experience that they very much do work. The scientific surveys you refer to used laughably small samples, and the technology has only gotten better since.
Again, nobody is asking you to use technology pushed by any capitalist monsters here. It’s also really weird to see people complain about automation on a Marxist forum of all places. Everybody knows that automation is used against the workers under capitalism, that’s not a reason to reject automation.
- 8 days
Okay, but can you please explain how it helps you instead of just saying it does?
- 8 days
Sure, it automates writing a ton of repetitive code that you’d have to write by hand otherwise. Most programming tasks aren’t novel or interesting in any way. In most apps you have to do things like render UI with stuff like buttons on them, wire it up to a service, create HTTP endpoints, talk to a database, and so on. These are well known and repetitive tasks that have to be done in every application. There is nothing challenging or interesting about them. Once you’ve done this stuff enough times, it gets extremely tedious to do. And you have to do all this stuff before you can get to anything that’s actually interesting or novel in any given application. LLMs do a very good job automating a lot of this repetitive code allowing you to work on the parts of the app that you actually care about.