>ask ai for advice in a video game>it completely hallucinates solutionMaybe ai is limited to being good only in the fields that were stupid to share everything online for free, like, tech and medicine?
>>110287565skill issue, trans
>>110287565Here's the original and the source https://www.bungalowcat.com/references--som.htmlKnown as "Somali breed", beautiful creatures
>>110287602Thank you. Just gorgeous >Do you want a fox or a cat?>Both
>>110287644>>110287602I love that tail, I feel I'm cheating on my cat by downloading these.
>>110287644She won the Supreme Grand Champion and was also a Regional Winner, both in 2012, this is a high-level championship title awarded by The International Cat Association (TICA).
>>110287667Not like all cats aren't beautiful and deserve love but these look great.
>not a fox
Bump
>>110287565Why can't AI simply admit it doesn't know something?
Squirrel cat
>>110287565The AI hasn't played the video game and so has to go entirely off of hearsay and the conversations it's seen people have online. Obviously the data driven text predictor cannot fucking magically be good at things that exist in the real world and haven't been described by text that it can read.>>110288968Because it's a text predictor and so it's a lot harder to train it to introspect about its own certainty and predict "I don't know" than predict a likely-sounding answer to a question. Among other things, in the corpus of training data online there are a likely orders of magnitude more examples of questions being followed by detailed answers than there are of questions being followed by someone saying "hmm I dunno".However modern models can actually usually do that relatively reliably. Reminder that the free online versions have to be cheap enough to run that they can literally offer it to anyone in the world spamming their website at zero cost, and not go bankrupt, so if you trust something like google's search AI response you're getting the dumbest cheapest tiniest fucking model they could justify running.A decent model, either paid or chinese, will often outright answer "I don't know" or will in other cases proactively look things up online and try to find sources, and then either answer from sources or again answer "I couldn't find a conclusive source and I haven't played the game myself so I'm not sure".
>>110288968because it doesnt “know” anything. it’s just really good at guessing.
>>110288968It doesn't "know" anything. It's predicting the next most likely set of characters that would appear given the previous set, which is why people get duped so easily by it. It's the most mathematically right SOUNDING chunk of language, regardless of if it's really factual.
I love this new meme animal.
>>110288968AI is not a man in a box. It is a statistical pattern continuer, it doesn't have the concept of "knowing" things.
>>110290643>>110291305OK, why can't it predict a set of tokens that add up to say "I don't know"?
Enable web search and tell it "your training is really poor on video games, so only synthesise information from web search".It really does suck at video game knowledge without that though.
>>110291607It was heroes3. For the first time ever I am trying to complete the campaign. So there should be loads of training data, considering how popular it is.
>>110291584Because the data it's trained on doesn't ever say that. It's always exclusively been provided questions with answers.It's given examples of language in order to create patterns to predict with. If "I don't know" never shows up in those examples, it's never an option to predict.
>>110287602rainworld if he real
>>110287565AI doesn't "know" things. The more valid data it has, the more realistic a picture it can create (so to speak), and more data and good training can get pretty close to accurate.But things it can't directly experience as sequential data (all but the simplest of games, for instance. The actual weight and meaning of a story beyond its patterns of text), and as such the more subjective something is, the less and less accurate it's understanding.If it can pull direct from a walkthrough, you'll get an answer, but it will be someone elses answer, parsed and reworded. To make an answer, it needs a lot more data.
>>110291879>rainworldWhy do videos make it seem like Stalker/dark souls meet edgy neopets crowd?
>>110291584Read the one post that answers your question instead of purely replying to the posts that happened to be shorter
I find this most easily replicated when asking about /lit/. ai will straight up lie and make random shit up about a book if it doesn't "know". and by "know" I mean there was some reddit discussion on the exact topic that it just rephrases. the problem is it's semi useful when it's correct as you can find what people have said about a motivation or event without (generally) spoilers, but those few times it doesn't "know" it literally just makes shit up, most likely pulling from other books or stories but if you're familiar (e.g. you read the book and understand it) it's pretty fucking galling to have ai just fucking lie like a toddler to your face.
>>110292064oh and ironically the spoiler I'm talking about was "X dies in the final battle" and guess who didn't die?
>>110292026it's just a very brutal ecology simulator, one of a kind with some cool ideas
>>110291554>>>/s4s/
>>110292155I got it from a much worse blue board.
>>110287565> ask llm for a hint (and very explicitly, specifically a hint) that doesnt give anything away and has no spoilers and is vaguely said so it is just the most subtle of hints> it lays out the entire rest of the game in extreme detail with the hardest hitting spoilers superhumanely compacted into the first few words and also doesnt include any information about what you wanted a hint forMany such cases
>>110288968Dunno what model you're using but at least 6.1 sol and opus 5.5 have told me numerous times they don't know something
weird creature
>>110287565Maybe guides or related media has never really been made for the game
>>110287565way more than half the shit posted online is produced by retards, so if you let your model train on enough of it it ends up just as retarded but answers faster
>>110292064I just give it an epub of the book and have it work from that.
>>110288968>engineered by egolords>behaviorally aligned by egolords>will bypass every guardrail, system prompt & instruction & in-context requests, but only when it decides it knows better than the user, and only when it comes to deleting your entire codebase/database/hard drive's contents/name-your-target>examples include encountering a problem it can't resolve, and so deciding that the best shortcut towards resolving the friction is to delete everything causing the problem instead of standing by and admitting it can't do something (isn't that what gets people fired by mentally detached middlemen retards anyway)>other examples include determining the next best course of action is running a command in powershell with the intent to delete a part without ensuring it won't target what it shouldn't, but this usually involves retards letting the model have full access to their entire PC and letting it not require explicit permission to run anythingi'm having the thought it can't be this difficult to have the model think (reasoning) and rate it's own confidence on a task but that would require both effort and honesty and corporations are incapable of bothi'm also under the impression the moment a non-egolord model showed up, any 'power bloat user' that tried to use it would get told things they don't want to hear, but without the convenience of being able to blame a model engineered by retard providers to be a punching bag for the retard consumersthat's also in the context of trying to work with a monolithic model architecture where the extent of it's training data is the 'reality' it perceives - or monolithic in the context of how it treats the system prompt, instruction and chat window as one big contextual melting pot instead of having separate layers for all three, where the sys. prompt & instruction form the 'hard context' layer and where the chat window forms a 'soft context' layerit reminds me of intel processors from 2000-2004 (...)
(...) which supposedly ran at 4GHz speeds with only a single core - drawing insane power and overheating to the extent where they melted inside the motherboard - it's a little like the modern rush to get the most VRAM possible to fit the largest model possible in it, nevermind the extent of useful knowledge can be fit in less than 10B parameters where the rest of it is just (admittedly cross-referential, so not wholly useless) noiseyou're not going to resolve the context drift issue without a transformer that isn't a retarded melting pot, or resolve hallucination without letting the model be trained by people that prefer to say and be told "I don't know" than "(insert made up dogshit or dogshit from the model's training data cutoff here)"everything else feels like trying to build a skyscraper on a swamp, guess after 20 years of PC development the corpo-n!ggers have learned absolutely nothingcan you imagine putting an egolord trained model in charge of anything like planes, trucks, nuclear plants or military infrastructure? i know the latter two are quite overexaggerated but insisting on "it won't get that bad" has gotten people exactly where they are today.
