You are talking about technicalities.
For a model to be as good as possible you train on the most accurate data.
It is true that you can take SD, modify it to ignore moral values and then ask for CSAM but if you for example have a bunch of real CSAM and you train it on that data it would be much much better at generating believable CSAM. Which is what these criminals do…
You are not logged in. However you can subscribe from another Fediverse account, for example Lemmy or Mastodon. To do this, paste the following into the search field of your instance: !technology@lemmy.world
This is a most excellent place for technology news and articles.
English
Care to explain or do you just like to say things?
To simplify:
If you want to generate an image of a Lion in a tuxedo it didn’t necessary need to parse images of lions in a tuxedo.
You are talking about technicalities. For a model to be as good as possible you train on the most accurate data.
It is true that you can take SD, modify it to ignore moral values and then ask for CSAM but if you for example have a bunch of real CSAM and you train it on that data it would be much much better at generating believable CSAM. Which is what these criminals do…
Bold statement from someone who is literally just saying things and posted nothing to validate their claims.
I work with LLMs, SD and Threat Intelligence so I have some professional knowledge on the subject.
That fails to rise to the level of verification. The expert witness must convince the jury that they are in fact an expert.
Dude this isn’t a court room, I said what I said and you can decide to ignore it if you want to.