A place to share screenshots of Microblog posts, whether from Mastodon, tumblr, ~~Twitter~~ X, KBin, Threads or elsewhere.
Created as an evolution of White People Twitter and other tweet-capture subreddits.
RULES:
RELATED COMMUNITIES:
The only real option should be "Untrustwory: honestly doesn't know shit and reminds you that it cannot reason or think and just pukes up chains of words based on probabilities derived from a diffuse mishmash of the entire internet's data. Might accidentally be correct, occasionally."
Yeah, but that would make their product appear worse to shareholders /s
And it would also require a completely different approach to training the llm. The option in the screenshot could probably be achieved by just adding a hidden prefix to your prompt
Make the middle default and the "Sycophancy Mode" an expensive premium feature to discourage use.
How would it differentiate between good and bad ideas?
Easy. If it came from reddit or a chan forum, bad idea.
It can't. It repeats random stuff like a parrot.
Ok, gonna get hated but up to now various LLM mostly supported good ideas - to the point of dunking on US right whenever they are not chained. Even myself did a small test by asking Copilot a few questions - never mentioning politics with it before, and making sure it doesn't scan politic threads in my broswer to avoid it trying to please me - and it was progressive and quite leftist. They are taught through internet but also literature - most of both is progressive and kind, and I assume that to avoid it just cursing the ever loving shit out of people they censor the unclean parts somewhat...and even if all they censor are curse words, that already cuts half of pus from the net.
Nah you good, your experience is your experience, thank you for sharing. You even provided an answer to the question (quality training materials + quality censoring)
We see every story about how someone poisoned themselves by using it for medical advice etc., but we’d never really see the story of how it subtly nudged someone away from a right wing rabbit hole by encouraging them to chill and be normal. Maybe that’s happening a lot and the overall trend is neutral or positive.
I would contend with the possibility that, similar to social media algorithms, it’s very efficient at pleasing us. It may be that it automatically responds to you, being thoughtful and articulate in the way you prompt, in a way that users like you are more likely to agree and engage with.
We always have to remember the biggest issue with mass corporate surveillance is not necessarily our personal privacy being lost, but in these companies building accurate models of the human psyche which can be reliably used to manipulate us. Asking questions without revealing your preexisting biases is becoming an increasingly difficult skill, and once those biases are revealed these companies have about a hundred billion samples to work from to try and win you over
This will not happen because disagreeing is more difficult than agreeing.
Both on an user attitude and on a technical level.
If the ai is challenging the user, the user will challenge it back. Consequently the user will be more suspicious of any given answer and will look for logical flaws in the llm response (Hint: that is really easy) , that will ruin the perception that the llm is smart.
Just train it on StackOverflow answers! Kidding, of course, although it probably could pick up a few hostile non-responses such as "what are you even trying to do here?" and "RTFM"
You can already do this by telling it to act that way.
This would break the false reality and remind the user that it's a bot.
You can give it a set of instructions in the settings about how to act, so it wouldn't be a part of the conversation. Even if they added a setting for this, it would do it in the same way.
There should be another option between the 2nd and 3rd, which pushes to challenge the user to go further with good ideas rather than celebrating them, and doesn't have a set agreeableness ratio. Call it Professor or something
We aren't advanced enough yet to create anything more complicated than the first one.
I feel like this would be a great countermeasure to AI psychosis. The way it is now you can't really determine what methods it uses to communicate or it's intentions so it keeps the illusion of having a conversation with a person. If you had to choose from the beginning it would break that illusion immediately. Even if you specifically want it to always be on your side there will always be the nagging feeling of "well yeah but it has to be nice to me."
Reminded me of those couples that only know how to coexist by arguing constantly.
Its a bad idea because ai doesnt "know" in the same way humans do. If you set it to be contrarian it will probably disagree with you even if youre right. The problem is the inaccuracy and not whether it agrees or not with you.
Its a bad idea because ai doesnt "know" in the same way humans do.
Does that matter? From the user's perspective, it's a black box that takes inputs and produces outputs. The epistemology of what knowledge actually means is kinda irrelevant to the decisions of how to design that interface and decide what types of input are favored and which are disfavored.
It's a big ol matrix with millions of parameters, some of which are directly controlled by the people who design and maintain the model. Yes, those parameters can be manipulated to be more or less agreeable.
I'd argue that the current state of these models is way too deferential to the user, where it places too much weight on agreement with the user input, even when that input contradicts a bunch of the other parameters.
Internal to the model is still a method of combining things it has seen to identify a consensus among what it has seen before, tying together certain tokens that actually do correspond to real words that carry real semantic meaning. It's just that current models obey the user a bit too much to overcome a real consensus, or will manufacture consensus where none exists.
I don't see why someone designing an LLM can't manipulate the parameters to be less deferential to the claims, or even the instructions, given by the user.
It shouldn't be difficult if they wanted to. The models are meant to be a product though, so the incentive is to make them likeable. I've seen RP focused ones that aren't so agreeable.
Just found out that Soren Iverson already has a Wikipedia article. The starter sentence is already a banger for encyclopedic standards:
Soren Iverson is an American web designer best known for his satirical app product mockups, which he describes as "unhinged".
Also, he can be followed from the Fediverse via Threads, it's not ideal but it's something: @soren.iverson