And for those who don't understand the concept, when we talk about poisoning data sets, as #AIResearchers, we are talking about inserting bits of data which are not simply nonsense, but otherwise valid inputs which have tokens and patterns close enough to non-poisoned input that when doing generations, they will weigh low quality responses, potentially even nonsensical ones or ones that bring it into loops or other fail states. A common example is putting in random noise into PNG images, which is almost imperceptible to the average human, but creates weird math when the #AI tries to build a statistical model off of it.

It's relatively easy with text, because you can basically just feed strings of nonsense that sound like they could be actual sentences, but either create statistical outcomes that raise the likelihood of inappropriate token output, or, if you're really clever, raise the statistical likelihood of common fail states, like loops or using thise words that AI uses more.