65
AI test that worked on a scam bot, or someone who doesn't speak English well: In the novel Blindsight, they're discussing whether an alien they've encountered is sentient, so the linguist sends this message:
"Our cousins lie about the family tree, with nieces and nephews and Neanderthals. We do not like annoying cousins."
"We'd like to know about this tree."
Sascha muted the channel and gave us a look that said Could it be any more obvious? "It couldn't have parsed that. There were three linguistic ambiguities in there. It just ignored them."
The scammer ignored this and ChatGPT whiffed it. Claude gave a similar response in that it tried to explain a sentence that can't possibly be explained. Neither caught the ambiguities. Happy hunting!
That's the point. A human would ask for clarification, but AI trys to continue the conversation.
It's similar to this guy doing a job interview with an AI where he just says things like "Freight cadence Richard walnut and sprucewood. Six years of sprucewood, actually seven if you count Richard." And the AI interviewer responds "It sounds like you've built up a long tenure with Richard at Sprucewood."
https://www.youtube.com/watch?v=Kcp1VBaCvEY
I see. One could ask for an AI capable of saying "Wait a minute, what the fuck are you talking about" when it needs to, but it would probably give away how fucking dumb it is.
The training is one part but its also how they work on a technical level.
They're trained to always give an answer and never give responses like "I don't know" and "I didn't understand". They could be trained to give these types of responses more frequently, but I'm not sure it fixes anything because of the technical aspect.
Technically, LLMs have no understanding of words, let alone sentence structure, because they're mathematical models. Input words are tokenized (turned into numbers, essentially), run through a large equation that was crafted and tuned during training, and the resulting output tokens are turned back into words. They can't be confused by ambiguous sentence structure because they don't parse sentence structure. They can only ever be trained to appear like they don't understand in the same way they are trained to appear like they do.
AI is designed for maximum usage loop. It’s why most of its answers end with questions.