24
GitHub - Gathin23/cleanyourtv: Debloat any Android or Google TV over Wi-Fi and get an Apple TV-style home. (github.com)
submitted 4 days ago by to c/technology@lemmygrad.ml
26 comments fedilink hide all child comments
you are viewing a single comment's thread
view the rest of the comments
[–] [S] 3 points 4 days ago (20 children)

It's always incredible to see what people who have no idea how this tech actually works think LLM assisted development is like.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 20 child comments
    •
  • [–] 1 point 4 days ago (9 children)

    Meh I am software engineer and am completely unimpressed by LLMs and even more unimpressed by those of my coworkers who have adopted it.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 9 child comments
    •
  • [–] [S] 3 points 4 days ago (8 children)

    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.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 8 child comments
    •
  • [–] 1 point 2 days ago (7 children)

    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.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 7 child comments
    •
  • [–] [S] 5 points 2 days ago (6 children)

    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.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 6 child comments
    •
  • [–] 3 points 2 days ago (5 children)

    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.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 5 child comments
    •
  • [–] [S] 5 points 2 days ago (4 children)

    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.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 4 child comments
    •
  • [–] 1 point 2 days ago (3 children)

    I don't think a crossword puzzle generator requires all that much context. Maybe they will get there at some point. They are still consistently confidently wrong, which makes them feel super untrustworthy to me. If a coworker said to me "oh, you're right, this fixes it" 4 times in a row without fixing anything, I would not trust that coworker. LLMs do the same thing.

    You can't even run them if you don't have enough system ram, but yeah, graphics vram is for sure very important.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 3 child comments
    •
  • [–] [S] 2 points 2 days ago* (2 children)

    LLM assisted development has nothing to do with trust at all. Think of it as a genetic algorithm. The model makes a rough approximation as a first pass, then it runs tests or specs, gets feedback, and iterates on it. And each iteration gets closer to what you specified. The better the specification that you provide is, the better the end result will be.

    Meanwhile, context isn't some abstract thing. It's the amount of tokens the inference engine can hold in memory. With the memory constraints you have, it's going to be tiny, which means the model isn't able to hold the whole solution in the context even for a small thing like a puzzle generator which will be thousands of lines of code.

    Good news is that pretty much any problem can be broken down into a series of steps. And then if you structure things well, each step can be treated as its own isolated problem which doesn't requite much context. With the example of a crossword puzzle, what you'd want to do would be to create an outline of the shape of the solution yourself, how the pieces fit together, and what the end result looks like, create tests for the functionality, maybe even files with stubs for the functions you want if you're working with a small model. Then you run an agentic harness like opencode where the model can iterate on the solution, run tests, and get feedback on what it's doing wrong. You create the scaffolding for the solution and you own the contract, while the model fills in the details for it.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 1 point 2 days ago (1 child)

    Trust is relevant in the same way it's relevant with old school blueprints, dependencies, and with other automation.

    In my experience the LLMs don't consistently get closer.

    A puzzle generator doesn't need to be thousands of lines of code by any stretch.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] [S] 2 points 2 days ago

    Not really because trust was never how we ensured software quality in the first place. What matters is the process, doing things like code reviews, having a good test harness, coming up with specifications, doing manual testing, and so on. LLMs don't really change anything here.

    I'd also love to hear how you actually use LLMs and what your workflow is, because they absolutely do get closer when you set up constraints correctly and run them in the agentic loop in the harness. I also don't know what you refer to by puzzle generator here, clearly you mean something simpler than I assumed. If you want to write out what the specific requirements are, I can show you what Qwen 3.8 comes up with on my machine for them.

  • permalink
  • fedilink
  • source
  • parent
  • [–] 0 points 4 days ago (9 children)

    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?

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 9 child comments
    •
  • [–] [S] 4 points 4 days ago (8 children)

    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.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 8 child comments
    •
  • [–] 0 points 3 days ago* (last edited 3 days ago) (7 children)

    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

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 7 child comments
    •
  • [–] [S] 5 points 3 days ago (6 children)

    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.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 6 child comments
    •
  • [–] 0 points 3 days ago (5 children)

    Okay, but can you please explain how it helps you instead of just saying it does?

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 5 child comments
    •
  • [–] [S] 5 points 3 days ago (4 children)

    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.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 4 child comments
    •
  • [–] 1 point 2 days ago* (3 children)

    Thank you very much. I really had not seen a full explanation anywhere before from someone who's gotten genuine benefits from it, because as you know there are piles of viral marketing everywhere extolling the big-name models for things they are not (like automatic hackers, real human personalities, etc)

    One senior engineer I know very well who works for a Big Corporate Retail company told me they had a ChatGPT subscription thrust on the entire software department, and he removeded about it quite a lot. Another I know works in a company that makes medical software and he said they bought everybody a Claude subscription and demanded they use it. He is not a fan and had the experience I mentioned above, having to proofread through every bit of output and constantly correct it. To the point that its raw output would not even pass the compiler without erroring out most of the time.

    What models have you been using, and what have you been feeding them as training data?

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 3 child comments
    •
  • [–] [S] 5 points 2 days ago (2 children)

    Right, there is a lot of insanity going on with this tech right now, and companies are doing nonsensical things with it. But the mania has to be separated from what the tech can actually do, and whether it's useful or not. Eventually the bubble will pop and a lot of this nonsense will go away with it. So, I'm not really worrying too much over that aspect.

    In a way, I'm also lucky that it's not being pushed on me at work, and I can decide when and how to use it. Which makes a big difference of itself. I can very much understand how people who are mandate to use LLMs for everything and are then have their timelines for projects adjusted based on that would be very frustrated.

    I've been mostly using DeepSeek and GLM, and my personal projects are open source, so I'm not to worried about the models getting fed my data.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 3 points 2 days ago (1 child)

    Chinese open-source models which keep undercutting the American techno-fascists by performing better and more efficiently for free? Awesome. I'm fine with those.

    I probably hear so little about their real-world use cases because none of the American press want to upset their capitalist masters, and there's not a ton of info on my usual Lemmy instances either.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] [S] 3 points 2 days ago

    Yeah, western media isn't a great source on this tech for the most part. And Lemmy tends to mostly attract liberal left which predictably attracted a reactionary position blaming LLMs instead of the underlying capitalist relations. I find Cory Doctorow has been doing a fairly good job covering this in a sober way.

  • permalink
  • fedilink
  • source
  • parent