Who’s ignoring hallucinations? It gets brought up in basically every conversation about LLMs.
People who suggest, let’s say, firing employees of crisis intervention hotline and replacing them with llms…
Less horrifying conceptually, but in Canada a major airline tried to replace their support services with a chatbot. The chatbot then invented discounts that didn’t actually exist, and the courts ruled that the airline had to honour them. The chatbot was, for all intents and purposes, no more or less official a source of data than any other information they put out, such as their website and other documentation.
It really needs to be a disqualifying factor for generative AI. Even using it for my hobbies is useless when I can’t trust it knows dick about fuck. Every time I test the new version out it gets things so blatantly wrong and contradictory that I give up; it’s not worth the effort. It’s no surprise everywhere I’ve worked has outright banned its use for official work.
I agree. The only application that is fine for this in my opinion is using it solely for entertainment, as a toy.
The problem is of course that everyone and their mothers are pouring billions into what clearly should only be used as a toy, expecting it to perform miracles it currently can not and might never be able to pull off.
The part that’s being ignored is that it’s a problem, not the existence of the hallucinations themselves. Currently a lot of enthusiasts are just brushing it off with the equivalent of boys will be boys AIs will be AIs, which is fine until an AI, say, gets someone jailed by providing garbage caselaw citations.
And, um, you’re greatly overestimating what someone like my technophobic mother knows about AI ( xkcd 2501: Average Familiarity seems apropos). There are a lot of people out there who never get into a conversation about LLMs.
I was talking to a friend recently about this. They studied medieval English and aren’t especially techy, besides being a Millennial with techy friends; I said that merely knowing and using the term LLM correctly puts their AI knowledge above the vast majority of people (including a decent chunk of people trying to make a quick buck off of AI hype)
