Hacker News Reader

Filter the input? If it's trained on all kid-friendly material and you have guardrails on the inputs what's going to come out. I believe Apple has done this pretty successfully on their image gen stuff that was clearly aimed at kids. Granted the outputs are... very boring, but they seem to never give back anything inappropriate.

Sure that’s plausible but is that what they actually do?

Cynically, my guess is it's just through the system prompt.

I’m sure they just dump the image into another LLM to gauge “safety” and pretend it’s good enough.