I'm not sure if this is a problem with overfitting. I'm ok with the model knowing what Indiana Jones or the Predator looks like with well remembered details, it just seems that it's generating images from that knowledge in cases where that isn't appropriate.
I wonder if it's a fine tuning issue where people have overly provided archetypes of the thing that they were training towards. That would be the fastest way for the model to learn the idea but it may also mean the model has implicitly learned to provide not just an instance of a thing but a known archetype of a thing. I'm guessing in most RLHF tests archetypes (regardless of IP status) score quite highly.
What I'm kind of concerned about is that these images will persist and will be reinforced by positive feedback. Meaning, an adventurous archeologist will be the same very image, forever. We're entering the epitome of dogmatic ages. (And it will be the same corporate images and narratives, over and over again.)
Granted, but not the best example, red and green are the emblematic colours elves wore in northern european cultures.
Santa is somewhat syncretic with Robert Goodfellow or Robin Redbreast, Puck, Puca, etc etc. it wasn’t really a cola invention.
> I'm ok with the model knowing what Indiana Jones or the Predator looks like with well remembered details,
ClosedAI doesn't seem to be OK with it, because they are explicitly censoring characters of more popular IPs. Presumably as a fig leaf against accusations of theft.
If you define feeding of copyrighted material into a non-human learning machine as theft, then sure. Anything that mitigates legal consequences will be a fig leaf.
In this case the output wasn't filtered. They are just producing images of Harrison Ford, and I don't think they are allowed to use his likeness in that way.
The fact that they have guardrails to try and prevent it means OpenAI themselves thinks it is at least shady or outright illegal in someway. Otherwise why bother?
I wonder if it's a fine tuning issue where people have overly provided archetypes of the thing that they were training towards. That would be the fastest way for the model to learn the idea but it may also mean the model has implicitly learned to provide not just an instance of a thing but a known archetype of a thing. I'm guessing in most RLHF tests archetypes (regardless of IP status) score quite highly.