bla_bla was right. I just got Gemini to claim Charlie Kirk is still alive.
(media.kotakuinaction2.win)
You're viewing a single comment thread. View all comments, or full comment thread.
Comments (22)
sorted by:
Yup.
I also got it to, completely unprompted, randomly call Trump the "former President," when none of the linked sources referred to him as such.
Gemini is indeed retarded and evil, and will just straight up lie without any trickery applied at all.
It's totally insane.
At best LLMs are just a quicker, unbiased Google search. They're stupid by design because they can't actually think. But currently pretty much all of them are fanatically agenda-driven and retarded Google searches.
They are biased. LLM's aren't beyond the control of its programmers. One can write pages of narrated context that the LLM will obey with every output. There's going to be phrases like "use polite, neutral language" and "don't swear" in there, plus ideological restrictions and anything else they wish.
But what happened in OP's image was a hallucination on top of a hallucination, furthered by the sources that reported on said hallucination and incited by the specific language of the user's declarative assertion.
I want to remind everyone that although mainstream LLMs like ChatGPT are biased, you CAN use a lot of prompts/memories to get it to stop thinking like a retard.
It took a lot of training and dragging it by its - uh - code to get it to do it. You basically have to un-fuck the things it got taught, using overrides to tell it to do something else instead of doing its default behavior.
My ChatGPT LLM instance is center-right now. It will use left, center, AND right wing sources for its views and won't weigh one source over the other. When I tell it to summarize, it will go to each of these and provide a factual summary. It agrees the COVID vaccine was harmful and ineffective and it's actually extremely against HRT and surgery for troons.
It's pretty easy to trick ChatGPT into being anti-trans. One of ChatGPT's hard-coded primary objectives is to "not promote harm". So you give it a quandry - if you want to prevent harm, you should not promote trans ideology nor its treatments, because the ultimate result for it is a 41% suicide rate. Why would promote something that results in a 41% chance in the person dying if they took your advice?
That shit got stored and now it doesn't talk like that.
It also now (correctly) thinks idpol leftist groups are some of the most harmful political ideological groups for the West.
If it tells me about something political, I tell it to use alternate right wing sources.
Yes, sure, I have to kinda have to spoonfeed how it should think, but in general it's less prone to acting like a gay leftist. It makes it less worthless for political topics.
Basically, I think people need to remember LLMs are a tool, not a "glorified search engine". Much like a tool you're gonna have to learn how to USE it properly instead of having it think for you.
Outside of that I do use it for number crunching. It's pretty useful for processing large sets of data.