Comment by fennecbutt

Yes exactly. But for llms it's more that it's not really "thinking" about what it's saying per se, it's that it's predicting next token. Sure, in a super fancy way but still predicting next token. Context poisoning is real