• dan@upvote.au
        link
        fedilink
        arrow-up
        3
        ·
        edit-2
        3 hours ago

        In the US and EU, copyright law only recognizes human authorship. Fully AI-generated (“vibe-coded”) content, including source code, can’t be licensed at all even under an open-source license, and has to be public domain. AI code is only copyrightable if it’s just used as an assistant to a human, not vibe coded.

        The fact that you wrote a prompt to output the code doesn’t mean anything. The prompt is considered an idea, and ideas aren’t copyrightable - only the actual implementation of the idea is.

        Even if that wasn’t the case, it’s risky referring to AI-generated code as FOSS, since there’s no guarantee that the output it gave you wasn’t the result of it being trained on a proprietary codebase. This also means that vibe coded clones of proprietary software are on very shaky legal ground. There’s no way to prove that the AI model didn’t use proprietary information from the commercial software.

        Clean room reverse engineering also still needs two actual people (or two teams) in order to do it in a compliant way. One to write a spec based on behavior of the original program, and another that never looks at anything from the original app and builds the new software purely based on the spec. AI can assist, but it can’t just be one person using two different agents.

        • b3nsn0w@lemmy.world
          link
          fedilink
          arrow-up
          1
          arrow-down
          7
          ·
          8 hours ago

          based af.

          tbh the only reason i mentioned claude is because afaik that’s what the slopcraft suite was coded with. but i can’t wait until foss models get to the point where current light frontier models are now. gemini flash 3.8 was already at a level where i only wrote code when i wanted to, not because i had to. if you understand the tech enough that you don’t need fable to design your app for you, we’re very soon going to have open source models that run on a pc and leave no tangible advantage to proprietary models. at least for coding.

          i’m really excited for the qwen 4 release, i hope the smaller models get on this level. it would be right on time.

          that said, if you wanna do a fully automated slopcraft, maybe you still need something like kimi to direct all your agents around. but we’re getting closer, and if the bubble indeed pops the way it’s looking like it’s doing now, hopefully the hardware is gonna be attainable soon too.

          • Natanox@discuss.tchncs.de
            link
            fedilink
            English
            arrow-up
            1
            ·
            2 hours ago

            There are no working FOSS AIs, you guys should read up on the definition. Kimi, Qwen etc. are Open-Weight, NOT Open-Source.

            The only true FOSS models are those trained on open datasets, with all settings, training routines etc. fully open as well. And there are only proof of concepts for those right now (e.g. those miniscule ones from PleIAs which always end up in a loop), but nothing that actually works let alone is useful somehow.

            we’re very soon going to have open source models that run on a pc

            No, no we won’t. We might have free models (as in free to download but a black box inside), but for FOSS models it will need decades to accumulate enough ethical data in those open data sets to train something useful.

            • silenium_dev@feddit.org
              link
              fedilink
              arrow-up
              1
              ·
              37 minutes ago

              I mean, a model licensed under GPL could afaik legally be trained on GPL-licensed code and almost every other open source project out there, if the training code and dataset is also published alongside it. So training data does exist, it’s just very expensive to train an LLM