Another data point showing Goolag's AI lying about Charlie Kirk's murder
(media.kotakuinaction2.win)
Comments (25)
sorted by:
I tried 3 times. 2 denials and 1 admission. I wonder if using the term "troon fucker" biased the algorithm against admitting it.
It biases the algorithm against assuming you are asking a good faith question that it can find the answer to with a simple Google search. After it decides not to make an API call it reverts to its training data, which ends in January 2025, and thus has no information on the assassination of Charlie Kirk.
I assume the same thing that its falling back to its last trained dataset.
But I have been using "ai" to help with programming. Asking it to write code feels futile unless its super simple.
But asking it questions like I would a teacher or a fellow student and it does a ok job at explaining why something doesnt work the way I want or how I should go about starting something.
Which is kinda nice, I can ask some machine a 1000 questions but I swear to god its like it gets annoyed sometimes(ironically like real life), and seems to work better when it is treated like a person.
It could be all in my head, I think the isolation is getting to me, or my parathyroid is about to explode. But it weirds me out a bit.
LLMs are good at identifying patterns and matching expected output. If you talk like a retard you will get retarded responses.
To be fair, you can also get retarded responses without acting like a retard.
See, this time the model made an API call, because it saw that you were asking a polite question about the current status of something. (So something it doesn't expect to find in its corpus.) It was able to search for the information and respond appropriately.
Stop badgering the AI. It doesn't prove anything, and makes you look like an immature child. The AI only has access to data within its training set. Obviously this set ends well before Charlie Kirk's assassination. It's also designed to mimic tone. If you come at it with a combative statement, it will push back.
This is proves nothing, it's just a tool behaving like anyone who understands how it works would expect.
A quick google search could have told you this particular model was released in June, and the actual data from the model shows that it's training data ends in January.
These things aren't magic.
There is a growing divide between the people who do believe it's magic omnecent god-like being, and those who push back against that specific belief.
Somewhere in the middle is the people who actually understand what an LLM is, what it's capabilities and uses are, while the two extremes get harder and harder to even have a conversation with.
That's just called being stupid.
And what are the people who argue against that concept like they are actually going to win a prize for proving the AI is not omniscient?
Also stupid people, just stupid people who happen not to be wrong on this particular issue.
I should know, I like to fight on the internet like I'm going to win a prize for it. We have our vices.
I got all my prizes 20 years ago, and try pretty hard not to bother most of the time, but old habbits and all.
Someone needs to do a bit about the horseshoe theory of stupidity, and the dialogue around AI is prime material for it.
That's the entire conversation FWIW. Just one question and one reply. I just realized the "So" at the beginning might be taken as a continuation of previous messages, but it was really continuation of what people on this board did. I just wanted to test it myself. IIRC this is the same AI bot that people caught refusing to condemn pedophilia.
Like I said before: LLMs are trained based on data up to a certain cutoff point. Unless they have access to the internet (and clearly show they are doing a search) then they will not have up to date information. And if they do, the only information they have will be whatever search engines tell them.
This is the correct answer. Gemini has a knowledge cutoff date of January 2025. Anything that happened after that date is going to be bullshit.
You can check yourself here: https://deepmind.google/models/gemini/pro/
It's on the page, but you have to scroll way down to the bottom.
Most people are too stupid and ignorant to be trusted to interact with a LLM and it's fatiguing to watch.
There's something about conversing with a machine that causes people to speak in an entirely unguarded way.
Like.. does it not occur to people that they save these logs.. scan them.. and then use them to decide how to better manipulate you in the future?
If you already know the answer.. why are you asking a robot oracle?
sue google in a red state using the alex jones precedent 💰 🚂
All aboard the free money train. 🚂 🚂 💰 chio chooo
had a similar problem with brave's leo ai until i forced it to look at recent data.
what's weird is because of the way leo operates, the search based results, which are usually garbage, got the right answer, but the more conversational answers insisted on answering wrong. (i know, it's because of the cutoff date of the training data, etc, still weird if entirely expected behavior).
but yeah, you have to get it to concede that a source it trusts, then get it to follow a link from that source, or else it reverts to training data and 'fact checker' mode, which assumes if the data isn't in the training data, it isn't true.
Google sucks, but his training data probably doesn't include the murder of Charlie Kirk.
Tony you know better than to anthropomorphize corporations and their LLMs.
You are correct. But that's not what I meant, it was just a normal grammatical error.
Fun fact: using "his" for "its" used to be standard.