So my theory is that with the help of telemetry or something else, AI can learn from data stored on users’ computers, meaning AI can steal your completed work, as well as your edits and corrections to your work, etc., even offline if you’re a Windows user for example.

In short, AI will be able to learn from you even when you edit your articles, edit your drawings, improve your music, etc. In other words, AI will literally steal your soul.

What do you think about it?

  • zlatiah@lemmy.world
    3·
    17 hours ago

    I mean that is pretty much what AI bros want to do… and/or maybe already doing

    From a researcher/developer perspective: the biggest bottleneck that affects current-gen AI is the lack of high quality training data; the more high-quality (a.k.a. human-generated and not complete shitposts) training data, the better. What people write on their computers would probably overwhelmingly be high quality. That means, without major technological advancements… if AI companies have access to the types of contents you just described, it is very much in their interests to use them

    I don’t 100% agree with this view, but if you subscribe to Prof. Emily M. Bender’s thought of seeing AI models as plagiarism machines, maybe you can say that AI is “stealing your soul”

    • brucethemoose@lemmy.world
      21·
      15 hours ago

      From a researcher/developer perspective: the biggest bottleneck that affects current-gen AI is the lack of high quality training data

      I don’t completely agree with this. Recent papers have been working miracles with synthetic data generation and smaller datasets (eg, Phi).

      Meanwhile, there’s a lot of speculation that Llama4 failed because Meta’s ‘real’ data was vast but not ‘smart,’ with hints via lines like this:

      In order to maximize performance, we had to prune 95% of the SFT data, as opposed to 50% for smaller models, to achieve the necessary focus on quality and efficiency.

      Whereas Deepseek, with a very similar architecture and size, wrote about how well synthetic data worked in their GRPO paper.

      And this keeps happening. As an example, Kimi Linear is (subjectively) performing very well in spite of its ‘small’ training dataset: https://huggingface.co/moonshotai/Kimi-Linear-48B-A3B-Instruct

      IMO the limiting factor seems to be GPU time, dev time, and willingness to ‘experiment’ with exotic architectures, optimizations, and more specialized models (including the burnt time/cash on experiments that don’t work).