flamingos-cant (hopepunk arc)

Webp’s strongest soldier.

  • 84 Posts
  • 418 Comments
Joined 3 years ago
cake
Cake day: June 12th, 2023

help-circle








  • One problem with the whole C2S thing is you can easily end up in the Matrix/XMPP situation where some clients and servers implement some smattering of standards and it just becomes confusing to navigate and use.

    This is approximately the situation in the fediverse today, with Mastodon in the role of Gmail. The difference is that email has IMAP and SMTP, client protocols that are not only standardized but universally adopted. Every email server implements them and every email client expects them.

    Tuta has entered the chat.



  • I’m sorry, but how does this even work?

    Open-weight models (e.g., Mistral 7B) are Apache 2.0 licensed for research/individual use; while commercial deployments require a Mistral license with separate terms for derivatives and production use.

    Apache 2.0 explicitly allows you to sell stuff licensed under it, if there’s restrictions to commercial deployments then it’s not Apache 2.0.

    To the actual question, most lines about the difference between ethical and non-ethical AI/machine learning stuff I’ve seen have focused on the consent in obtaining the training data. Eg, ChatGPT or Sora aren’t as they didn’t get consent for their training data, while stuff like the voice banks Eclipsed Sounds produce are as they obtain consent from the person they train their models on.

    I highly doubt Mistral is getting permission for all the stuff they’re training their stuff on.











  • I was curious to see how they handle this on the fedi side, because they obviously can’t stop you from uploading images to other instances, so decided to do some digging myself.

    The fedi code for this is here and looks like this:

    # Alert regarding fascist meme content
    if site.enable_chan_image_filter and toxic_community and img_width < 2000:  # images > 2000px tend to be real photos instead of 4chan screenshots.
        if os.environ.get('ALLOW_4CHAN', None) is None:
            try:
                image_text = pytesseract.image_to_string(
                    Image.open(BytesIO(source_image)).convert('L'), timeout=30)
            except Exception:
                image_text = ''
            if 'Anonymous' in image_text and (
                    'No.' in image_text or ' N0' in image_text):  # chan posts usually contain the text 'Anonymous' and ' No.12345'
                post = session.query(Post).filter_by(image_id=file.id).first()
                targets_data = {'gen': '0',
                                'post_id': post.id,
                                'orig_post_title': post.title,
                                'orig_post_body': post.body
                                }
                notification = Notification(title='Review this',
                                            user_id=1,
                                            author_id=post.user_id,
                                            url=post.slug,
                                            notif_type=NOTIF_REPORT,
                                            subtype='post_with_suspicious_image',
                                            targets=targets_data)
                session.add(notification)
                session.commit()
    
    

    The curious thing here, apart from there being both an environmental variable and site setting for this, is the toxic_community variable. This seems to be a renaming of the low_quality field Piefed applies to communities, which are just communities with either memes or shitpost in their name.

    You also don’t get social credits docked for this.