|
|
|
|
|
by msp26
845 days ago
|
|
I want image generators to generate what I ask them and not alter my query into something else. It's deeply shameful that billions of dollars and the hard work of incredibly smart people is mangled for a 'feature' that most end users don't even want and can't turn off. This is not a one off, it keeps happening with generative AI all the time. Silent prompt injections are visible for now with jailbreaks but who knows what level of stupidity goes on during training? Look at this example from the Würstchen paper (which stable cascade is based on): >This work uses the LAION 5-B dataset... >As an additional precaution, we aggressively filter the dataset to 1.76% of its original size, to reduce
the risk of harmful content being accidentally present (see Appendix G). |
|
That’s the crux of what’s so off-putting about this whole thing. If Google or OpenAI told you your query was to be prepended with XYZ instructions, you could calibrate your expectations correctly. But they don’t want you to know they’re doing that.