Twonks | Bluesky

Transcript

TW😶NKS

A comic in four panels:

Panel 1. White text on black

AI Design Logic

Panel 2. A guy sits in a restaurant at a table with a checkered table cloth. A waiter stands near, hands behind back waiting attentively.

Guy: Get me a cheese pizza

Panel 3. The waiter returns with a pizza in hand.

Panel 4. The guy gestures proudly at the pizza. The waiter looks less than amused.

Guy: Wow, look what I made!

  • lime!@feddit.nu
    633·
    2 天前

    you know, i’ve tried to defend some usage in the past, explaining my processes and the many steps of manual refinement, masking, and layerwork i put in to things, how i only run local models with open weights, how all my power comes from hydro etc etc

    but as the tools keep evolving i’ve realised nobody else seems to actually care about the process. the pro-people just want as much slop as possible. someone likened it to a slot machine, where you keep pulling just because. that’s where we are now.

    • Flames5123@sh.itjust.worksEnglish
      122·
      2 天前

      I fully get where you’re coming from. I fully believe that you can’t vibe code correctly unless you already know how to code correctly. I’m against the shifting paradigm of “who cares what the code looks like aa long as it works properly and the LLM can read it quickly” bullshit that’s coming out of it. I want to read the code and understand it too. I want it to be object oriented and not just dumb ad hoc methods everywhere that’s 1,000 lines when it could’ve been 100 lines.

      Now anecdotally, as someone who uses it for my main work and side project, I am still getting a lot of use out of it. I’m learning new things at a faster rate than I would have before. For my side project, I am trying to optimize gear sets for a game and there’s hundreds of millions of different configurations. The LLM I’m using knows about my code and the project and what I need and is able to suggest other algorithms, like I was able to learn about Dinkelbach’s algorithm. I have it write up design docs with formulas and pseudocode implementation and I review that and it takes my comments into account. I treat it like a junior developer and ask for questions to make sure I understand what it’s doing. I think a lot of people aren’t treating it like a junior developer or intern and that’s where problems come from.

      Now, I wouldn’t be using it if it wasn’t free with my company, as this is more of “learning/research” for my job.

      And for my job, we have semantic memory and a ton of MCP servers setup that guide it through the right code and can do internal documentation search so it’s way more powerful than just using base Claude Code or whatever. It has helped me stay more on track with my projects as an ADHD person (even though I’m medicated) by documenting what it does after it does it in a shared doc rather than me forgetting to do that because I run a command and log off for the day or something.

      I do hate the water usage and energy usage though…

      • Holytimes@sh.itjust.works
        2·
        19 小时前

        As a min maxer in a lot of games, I used to make spreadsheets in little tiny janky python programs do all sorts of crazy s*** to optimize stuff.

        Would spend hours on it! A little local model has basically reduced my effort. 10. Fold and saved me hours upon hours of work.

        Probably the single best thing is I can just yell at it to tell me what’s on the f****** to-do list and it will just tell me. And it helps my ADHD ass actually decide on what to start working on for the day and actually get something done instead of doing 20 things at once.

        Fantastic little tool.

        Now I have more time to play my games and I have something to yell at me so I don’t get distracted

      • oce 🐆@jlai.lu
        8·
        1 天前

        Same, but you forgot, the massive stealing of intellectual property that no government fights against because they are scared to be left behind by big tech, the impact on learning (especially students), and the impact on little hands that had to moderate the NSFL content.

      • lime!@feddit.nu
        21·
        1 天前

        i refuse to use ml models for code. the copyright issues alone should be enough to keep them away from every public code base until the matter is settled. but also because local tooling is, frankly, shit. i have a bit of hope for text diffusion models, but i have a hard time seeing the situation improving because everyote is full in on cloud models now.

      • lime!@feddit.nu
        21·
        1 天前

        i do. i experiment with transformer and diffusion models like a few hours a month, tops. the result isn’t interesting enough. the process of bending and breaking the models is the fun part.

          • lime!@feddit.nu
            1·
            1 天前

            yeah, it’s a great way to see the limitations of these systems. just like ctf and ioccc.

              • lime!@feddit.nu
                4·
                1 天前

                i will always take the time to explain things i find interesting to people. the benefit here is that i can now much more efficiently break large models as well when i come across them. helps me add anti-ai clauses to websites, cv’s, and repos i publish.

                  • lime!@feddit.nu
                    1·
                    1 天前

                    it’s the curse of getting into something early and it becoming a global issue; i did a course on machine learning in 2014 and have tinkered since then.

                    like imagine getting into blockchain-based systems because of git in 2007. can’t talk about that stuff at all now.