I wonder how much data it would take to actual cause noticeable (or triggerable) behaviors in the model. Like I've noticed certain models have been trained off of my university's course captures by their very odd and specific vocabulary/capitalization of specific terms but surely if you're scanning the entire web you'd need a lot more to sway it
Yeah think your intuition makes sense. Probably won't bias it all to much and perhaps, they can filter it as they probably log what is going through the API.
See also the current top story about low background steel.
Wow, first time I come across the concept. Fascinating, and a bit eerie when we start drawing closer relationships between AI and nuclear tech.