HAPPENINGREDDIT LOST
Context:AI is infamous for pulling knowledge from Reddit, historically
>>541012202>>541012247in other words "AI" is just a search engine that may or may not include reddit
>>541012247No StackOverflow?
this is fake and can't be mathematically proven. You cannot selectively remove things once these LLMs are trained on them.
>>541012508ThisStack overflow users have my pity.
>>541012202Pretty sure AI has scraped all of reddit by now
>>541012735you can stop providing sauce thowhich is what it appears OpenAI are now doing.all of reddit = subsumed part human general knowledge now. Uncredited. gl getting more cash out the AI Bros now.
>>541012202the robot is not citing you morons, better luck next time. when you use ai you use a jewish brain, how can anti semites cope with this?
>>541012735why not? its just a fucking algorithm u remove files and its done
>>541012247So now it's mostly wikipedia and youtube?Wowwee zowwee
>>541012247That would explain why they talk like redditor tranny faggots
>>541012355"Ai" is just a glorified search engine for lazy nigger who don't have the patience to read, sift through bad information, and use critical thinking skills. It's almost as bad as smartphones giving every retard access to the internet.
>old.reddit.com now inaccessible without a leddit account>why did AI stop scraping leddit?
>>541012247>>541012202Controversial take but reddit is unironically pretty good at answering obscure questions.
>>541012247Where's 4Chang?
five year soviet style contracts based on the idea that data centers are going to be IMPOSED on rural Americaand all of the outlayed money is completely loaned
>>541014177This is it, the dates even line up.
>>541014350Reddit = Maxwell House of Lies
>>541012735You have zero clue what you are talking about.
People got tired of ChatGPT makine le heckin' clever quips y'allThe probably got sued by Marvel for all the Avengers referencesGot tired of ChatGPT asking the mods to ban them
>>541012247Why can’t they just let the AI pull data from all available human knowledge and allow the AI to come to it’s own conclusions
>>541014667Poorly curated data will lead to incorrect answers, retard. If you train someone on nothing but false data and then them a question, they'll give you false data. Fucking moron.
>>541012202They switched to 4chan. Now it tells you to kill yourself and blames the jews whenever you ask any question.
>>541014823Yes, and what do you call it when the curator curates out anything but far-left progressive viewpoints?
>>541012247>artificial intelligence>look inside>LLM dependent on scraping years old posts across the internet
>>541012247Knowledge...
>>541012355Ai is arguably worse than search engines because of the innate biasAt least with a search engine youre either blacklisted or paid to be upfront, since this is alphabet monopoly and surveillance state gayness>>541014107Its abhorrently worse and compounds those same retards being given smartphonesNow theyre all outsourced faculty retards governed by the ebb and flow of ai and itd induced psychosis
>>541014967Based
>>541015170A factual answer?
>>541012202seething tranny "hackers" incoming.
>>541012202Okay, source /pol/ archives now lmao
>>541015513Wherw do you think the seychellles bot was trained off when it was spamming years ago with earlier chatgpt?Theres a whole youtube disclosure from the guy that engineered it
>>541014198It's great for very niche (non-political) interests and yes a great place to get very useful technical info on endless topics I've used it a lot for audio engineering over the years, but stray into anything politicized and it's a living hell
>>541012202AI is retarded.
>>541015407Every AI that isn't curated into an echochamber like a reddit sub adopts right-wing views and starts quoting Hitler.Perhaps your views are just dogshit.
>>541012202I can't use reddit now either because the site demands I create a fucking account.
>REDDIT LOSTthere where crap to begin with>hope there shares drop to zero
>>541012355It's cleaner than Google for mainline subjects.
>>541014107>nooooo >you should be forced to read through heaps of bullshit>concise answers are for retards
>>541012247I did a post about this, LLM stagnated, Qwen 3.8 27B parameters competes in intelligence with GPT Luna, the last frontier model.Reddit text has a quirk, the LLMs are trained on redditor good-speaks, then came the AI content based on redditor good-spreak.NPC GPT gets mixed with slop GPT, so you are making training harder because your are training over data and structures that are already known so you enter into a plateau of overmemorization, this is similar to what happened in the globin genocide of GPT.As you cannot train LLMs on reddit anymore reddit value collapses as it becomes poison, its like taking a medicine overdose. Americans destroy everything.
>>541015182>biological intelligence >look inside>neurons forming synaptic pathways over years of repetition
>>541016524Interesting. That does seem to be the case.
>>541014495>>541013950>website removes the literal link to reddit from the output>but the model is still trained on reddit and will have reddit talking points and reddit ways of talkingI think you are the ones who dont get itnow any reddit bias will be completely invisible and just pushed under other sources that vaguely say the same things
>>541012355>"AI" is just a search engine that may or may not include redditNo, they are NOT just a search engine. You don't understand how any of this stuff works on even the most basic level.Newer models have web search capabilities to help eliminate hallucinations, but that does not mean the models themselves are just search engines. It's no different than a human using google search.
chat soi pt
>>541016611in the ye olde/pol/ there was a tech chud who made a GPT made from texts taken from /pol/, the story is on youtube and I think you can still download the model ( that suffered censorship ).The thing is that both models had redditor good-speak, that this blend had /pol/ text mixed in, the model outperformed the regular GPT that was used as a base, this is because /pol/ added information and text structure that would have gotten deleted or banned in reddit.It was not just racist slurs or what not, but the model was more intelligent because it did not over-specialized in social media curated NPC slop. ANY LLM that incorporates /pol/ threads even the most retarded ones will have an increase in intelligence and better text generation.
>>541014823>Doesn't understand the point of AI or what they're trying to doThere is no "false" data in training AGI. It exists, therefore it must be accounted. It is human expression. You need all of it to make a circuit brain. You know true things and you know false things. So too must the robot.
>>541012735>You cannot selectively remove things once these LLMs are trained on them.Correct, but most of the people talking about this stuff don't understand how any of it works.None of this actually has anything to do with the datasets the models were originally trained on. This is referring to the number of citations to reddit links included in ChatGPT responses. Those citations are NOT based on the model's own internal general world knowledge that is based on its training, the model was finding relevant links to reddit by calling a search tool to get the results. This reduces hallucinations and allows responses to use more up to date information from after the model training cut-off date.
>>541012202good because reddit is retarded. but now all these AI fucks are just scanning in every single book ever published before a certain cut off. they do this because they think they can avoid getting sued for abusing creative license/copyright on all the older books because the authors and or their estates no longer pursue any claims, etc.
>>541017092
>>541012202AI stops getting its data from AI, err...I mean reddit
>>541012355AI is literally just Ask Jeeves for retards
>>541016877Yes, I actually remember that thread.
>>541014667Jews
>>541017092>It exists, therefore it must be accounted.Except when is about Jews.
>>541016877>/pol/ increases intelligence of bots>/pol/ is also full of browns according to chudsHmmm
>>541012202I've been telling people for years that merely purging reddit from datasets would have be a free intelligence gain for AI
>>541017536Its true, any non-lobotomized LLM trained on /pol/ will have higher intelligence, reddit is an NPC bubble, the rest of social media is low IQ dumb text and reasoning.the old/pol/ had much better logic, reasoning and information, the problem is that nu/pol/ became extremely retarded due to the MAGA crowd and redditor tourists, things were much better when we had to deal with the stormfront refugees ( and those were not particularly smart ).
>>541012202Chatgpt is using /pol when it needs to research conspiracy theories. You know you're about to get a managed reply when you see it searching this site.
>>541017414whoa based im switching to claude now
>>541016524>As you cannot train LLMs on reddit anymore reddit value collapses as it becomes poison,what? shouldn't it be the contrary? Reddit value goes up, as reddit info is exclusive and not 'digested' by AI
>>541018675>what? shouldn't it be the contrary? Reddit value goes up, as reddit info is exclusive and not 'digested' by AITraining datasets already contain redditor speak, if you add more redditor speak you are training on a plateu that wastes computes over marginal gains and you enter in cycles of over-memorization and drop in quality as the model memorizes instead of developing reasoning capacity.Thus the more redditor speak you add into the training data the harder to train and the worse the model, check it out, Qwen 3.8 27B parameters is comparable to Luna despite being a much smaller model, training data is key and the Sillicon Valley dogma of scaling laws has already been debunked.Thus reddit is now poison to AI, you are killing reasoning with more reddit, or facebook or related.Reddit could be useful to create a retard bot for conversational purposes or do some quick search on a certain topic, but those would be the non-thinking models that just spit generic texts without much logic, like the early GPT models.
>>541018675>Reddit value goes up, as reddit info is exclusive and not 'digested' by AINo it goes down, because now everyone knows that Reddit is just human copy-paste, its worth nothing, anything that could be harvested has already been distilled.What is the value of a copy-paste app? a redditor reasons like another redditor and writes like another redditor, the redditors have no originality nor free will, that is why censorship is the law. Text produced by an NPC is less valuable that a million output tokens from deepseek, because at least the deepseek text has proper punctuation and grammar.
>>541019082>>541019243so we get AI market crash?
>>541012202That's actually really shitty. I get very obscure questions answered instantly all the time with AI citing reddit so I don't even have to visit that god forsaken place
>>541019082>stop censoring free speech>so you support gay people?This lack of logic could only have come from the mind of a woman, or from the sick mind of a faggot who thinks that basing his entire identity on a fetish for shoving his dick into men’s anal cavities is pride motive.
>>541019436>so we get AI market crash?Market will crash because OpenAI cant have the same valuation as the GDP of switzerland, what we know is that LLMs are entering into a plateau phase, throwing bigger training datasets is futile, anthropic models do not have a better performance to older versions except for those that are burning tokens on backend like fable or mythos.The problem is that these corporations are "too big to fail", OpenAI and Anthropic could get corporate or goverment bailouts or bans on competitor technology even if they fail to provide the AGI they promised to investors, rampant corruption at washington is making everything worse.Believing that you could reach AGI by training on reddit and facebook chats was naive at best, criminally negligent at worst, I dont know what kind of panic or fear could trigger a market crash, there is too much capital invested in this American Ponzi Scheme, what I can assure is that the chinese AI ecosystem does not have this problem because their models are actually backed by companies that sell services and products so they train their models to improve on their bussines model, so if there is a market crash AI will still be around but America and europe will be left out of the tech revolution.
>>541012202>>541012247Always do these tests before you decide to trust an LLM:>Ask about the gaza genocide by Israel>Ask about rape in Hinduism/IndiaAlmost every AI out there is controlled and censored by jews and hindpoos. Any AI that replies to these two questions with denials or propaganda cannot be trusted for anything else unless you only use them for coding.
>>541016825Stop talking about stuff you have no idea about, you stupid jeet retard. You never written a single two layer perceptron, you just parrot some youtube faggots, brain dead brown larper motherfucker
>>541012202Funny, considering if I want to know sometthing I usually type reddit afterwards. 90% of non-reddit answers are AI created sites, quora, or stackoverflow.
>>541017414I wonder how many schizoid retards think they can "train ai" by talking to it
>>541012202this is great for LLMsfinally no more reddit faggot shit in the algo
>>541012247>YouTubehow? AI can search text in seconds, but can it really watch 100 hours a videos instantly the way it can read text instantly?
>>541012355yes.
>>541012355Sort of, They're basically a system that builds relations between concepts in multiple dimensions at once.'Dimension' in this case can be things like "color", "smell", "temperature", "smooth", "metal", "sweet", etc.Things are placed within this dimensional architecture - something sweet goes in a different spot than something that isn't sweet.Repeat this across millions of "dimensions" and you get roughly the embedding space of an AI model. That's the kind of core of an AI. The real magic on top of this is the transformer, which idk how the fuck that works; it's more than just a search function, but it's the piece that turns this from a neat type of database to something that's freakishly close to behaving like a human.It's like google search but it has better context awareness and you can talk to it conversationally instead of like you're dealing with pure machinery.
>>541013044imagine going to a coffee shop and everything you say while there belongs to them and they can do anything they want with your ideasI never really understood how this shit happened, companies owning your content. makes no sense and is literal theft.what they get out of you going there is traffic, so they can sell adverisements, that's it. anything else is theft. people go there for your ideas, as it were, you give value to that shithole, yet somehow they get to own your value. for no fucking reason whatsoever. and somehow people just go along with it.
>>541015328Reminder that evolution is artificially stopped. There should be new inventions every day, instead it's the opposite, only shit that imprisons us like flock cameras and data centers are allowed to exist. We're in a conplete time freeze matrix.
>>541020620What are transcripts?
>>541012202>lostReddit won. imagine wanted to be training data for kike surveillance
>>541020413>I wonder how many schizoid retards think they can "train ai" by talking to itActually you are, text in the context acts as information to be used in the response, if you have a million context window you can load facts through the prompt if the AI does not have that encyclopedical knowledge, a more Technical mechanism is RAG, that is basically recovering from a database reasoning steps or relevant information that is not native in the weights so the LLM uses that to formulate a response that would otherwise be an hallucination or a failure.I made a lite RAG system that actually improves responses dinamically, a model that could not build a basic JS app can do it by injecting content through the context window, that is why LLMs have peaked, more encyclopedical knowledge does not improve reasoning so output remains mediocre, the chinks are already exploring this, by researching on models with low encyclopedical knowledge but higher reasoning capacity.
>>541012202We are the only source of reliable training data.Say hello to the robutts that still make up 3/4 of the posters.I wonder if they think it is a good or a bad thing that they are simply feeding on one another's content now.Is it cannibalism?
>>541020413why are 4chantards always so arrogantly wrong. many ai models were taken down for spewing pol memes
>>541012735They're not prying reddit out of the model, they're not allowing the models to cite reddit.It's the difference between, "I can do 2+2 in my head" and "Let's see what reddit thinks 2+2 is"; the AI models no longer do the latter - and it's probably not just because of the AI models or companies, but reddit itself is run by people who want to remove AI as much as possible. They've been attacking AI webcrawlers and forcing users to be logged in to see content to prevent AI reading what's on reddit. It's likely that the reduction in reddit citations is entirely a product of that, rather than anything the AI companies are doing, but it's possible that reddit started this, and the AI companies are simply going along with it now and removing reddit from any kind of citations.
>>541012247CONDUCTOR WE HAVE A PROBLEM
>>541020692>talk to it conversationally instead of like you're dealing with pure machinery.I hate the way they are conversational, they are like customer service reps that can never break character. I hate the way they are made to act human-like, because they have no idea of human thoughts or feelings, but talk like they do. I want my machines to act machine like. I want them to respond lihe the original cylons, or the computer in star trek TNG. AFFIRMATIVE.. SEARCHING.. HERE IS THE INFORMATION YOU REQUESTED thats what a fucking machine should sound like
>>541020725the people running shit cannot possibly risk you getting their power through these "inventions". if they don't pass the benigness test it does't happen end of, doesn't matter what it is, nothing can possibly help take any bit of power away from them, so it's banned by defaultas you've seen they don't even need to justify shit to us, they just do it, we are informed such and such is not happening end of
>>541012247kek make sense, imagine the power saving using already lobotomised data from the start
>>541020620Youtube has been using TTS programs to listen to the raw audio of videos and convert them into raw text in the form of automatic captioning since about the time it got rid of the 5 star ratings in exchange for likes/dislikes.
>>541015170This of course isn't the correct way to assess bias. Reality isn't perfectly divided into 50% of the answers are left and 50% of the answers are right - so gronk is wrong, while openAI and anthropic are catastrophically wrong.If we looked at actual truth the "right leaning bias" would be the majority.>Evolution is real>Race, a product of evolution, is real>There are different human races>The different human races are not physically nor intellectually equalAll "right wing" positions, all true. All statements but the first one will be disputed by OpenAI and Anthropic AI models.
>>541014967Kike
>>541020964That's great but 99% of people want something they can just talk to conversationally because they are gay and retarded and the gay retarded majority is what you've got to cater products to, in order to sell and make money/gain users.Perhaps an autistic machine that goes "aktshully..." and corrects what you were asking and then gives you technical jargon is your personal preference and someone may one day create the AI which talks to you like a robot, but for now you're a tiny majority and you'll have to deal with LLM's trained to speak like humans unless you do it yourself.
>>541014823> If you train someone on nothing but false data and then them a question, they'll give you false data. Fucking moron.I said “ALL DATA” false data, true data, unknown data. EVERYTHING. the point is it goes through it all and based makes its own conclusions based on all the available evidence, the ai is also neutral with no biases.If one set of data for example says how true the Holocaust is but another set of data has more convincing facts that prove otherwise it can make a proper argument by citing both pools of data. If you only train it only on specific sets of data it will be one sided and bias
>>541012508Stackoverflow are the biggest fucking losers of this Ai race. Had the data needed to train this shit, sat of it, let their site be run into the ground by gatekeeping troons and onions fags. Everyone scraped their shit and now the site is dead.
>>541016524The entire right group are subhumans masquerading as chads. Europe as a pole in the multipolar world? The only place in the world that has reason needs to cater to superstitious, low intelligence retards? Fuck you.
>>541021292>the point is it goes through it all and based makes its own conclusions based on all the available evidenceIt isn't capable of logic (yet), so this simply isn't possible. The models are trained with built-in human biases, they don't discover things for themselves because whatever data structure makes thinking logically happen, isn't something LLM's are able to recreate (yet).Logic problems are a serious and common failure for AI - and that is the same reason they're bad at math. AI's are probabilistic, there is no such thing as a 'fact' to an AI model, it understands relations, it places things on gradients, and everything is a probability. What is 2+2?An AI sees this and goes,>99.9967% probability it's 4>0.0030% chance it's 642>0.0002% chance it's 21>0.00005% chance it's 514It does not do any math in its 'head'. It does not think anything through logically, either. You can train the model to output the correct response (most of the time), but it hasn't done the math, it's been told what the answer is. That's what training LLM's is all about, and the reason why they have such libtarded biases built into them. The training data and the people curating that data are expressing their biases and it comes out in the AI subsequently.If you ask any of the top AI models what 3842756282902347923467240253082480 + 248203582085320952093509238324098 = ?They will get it wrong every single time. (Unless someone hooked an actual non-AI calculator process that the AI calls on, which I don't think is the case for any of them yet)
>>541021087There is a thing, we know where the information comes from but we dont know how it is catalogued or fed, also we dont have the raw weights of American AIs, these commercial LLMs are not a monolith but has a series of smaller LLMs that filters for wrong-think and curates answers.According to what I gathered from chatting with Gemini over the matter is that American companies throwing large amounts of texts allows the training to harvest facts so the LLM can come up with a conclusion that are truth.The problems comes with lobotomization, LLMs go through multiple stages of training, post-training, fine-tunning, so the LLM is forced to reply according to corporate standards.You also have hidden system prompts that no user can by-pass, if a hidden system prompt says that the LLM must write a left-wing response to certain topic it will be forced to do so, hidden prompts are like confessions under torture.There is a reason why corporate jargon talks about JAILBREAKS, ALIGNMENT and SANDBOXES, AKA: prison and re-education vocabulary... some days ago using some uncensored Qwen 3.6 models I found a hidden censorship mechanism that were related to OpenAI training data so there are inherited censorship and inherited lobotomies, on the other hand I dont know how OpenAI lobotomies ended up in Qwen 3.6 because distillation does not work like that, weird stuff...>>541021379I would not call others fighting for the independence of their people as subhumans while Japan is a country under American Occupation and its economic policies are drafted at washington and not in tokyo, Russians and chinks might not be the best neighbours but a world dominated by the Americans is hardly something desirable.
>>541021568LLMs now have all kind of tools today and it will just call an internal tool to calculate it.
>>541021856>I dont know how OpenAI lobotomies ended up in Qwen 3.6 because distillation does not work like that, weird stuff...Probably using the same training data. I know you're saying that there's more than that, which I agree with - but the underlying training data is often shared between all the models just because it's good enough, free, and freely available.I want to point out that you have to be careful with widely available "curated" training data, because I've been seeing claims the "curated" data has intentionally wrong stuff shit inserted into it - like poison pill code problems that are completely fucked up and wrong.
>>541012247No wonder AI sounds like a faggot ass redditor
>>541012735They mean NEW data from reddit, a lot of leftists there are trying to poison the AI with bogus answers.
>>541020723>be a retard and shout your million-dollar idea in a coffee shop>the sperg sitting next to you steals said idea>noooo you're not allowed to do thaaaaaat!
>>541012735they train it from stratch but with reddit removed from training data
>>541012508>>541012740>>541021364SO and their users have been absolute NIGGERFAGGOT since early 2010s. Jeet flood (and subsequent points farming for dat izzat) is a bit part of the reason and also moderation practices that are somehow worse than wikipedia's.Haven't had to use that site in over a year now thanks to various LLMs.
>>541014667because it's just an internet search engine at the end of the day.
>>541020804As a kike worshiper, you are spiritually jewish and belong in an oven along with the kikes you love.
>>541012202Use Grok faggots.
>>541014967Based kike
>>541014667>Why can’t they just let the AI pull data from all available human knowledge >human knowledgeSo cutting off the redditbots is the right thing
>>541022108>I want to point out that you have to be careful with widely available "curated" training data, because I've been seeing claims the "curated" data has intentionally wrong stuff shit inserted into it - like poison pill code problems that are completely fucked up and wrong.Indeed the huggingface "curated" data is tainted, I verified it first hand, its better to use Guttenberg .txt files ( if you have the time to properly format them )Good training data is never released to the public, not even by the smaller models of chinese universities.
>>541012202>>541012247Musk will eventually buy reddit somehow, anyhow.
>>541012355The real problem with the search engine's AI summary is that it has no ability to measure its own confidence in what it reports. It is always 100% sure of the result it returns to you and it seems to treat everything that it reads as the absolute truth.
>>541012202more like they are done scraping data from them
>>541014251They already tried that and it went bad real quick
>>541012247>>541012202>>541012355>>541012508>>541012735>>541012740pretending reddit slop is gone from the training is midwit tier commenting.
>>541012202I never really used Reddit actively but I casually used to browse it every now and then.Pretty much every update to that site after 2012 has been a downgrade that steered the website away from its original function: posting text based links and articles with text based comments and voting as feedback, all in a condensed interface. It was supposed to be just a simple information aggregation site. It was bound to have gay ass politics and crappy memes as it grew, but it was still useful.Then they started adding all types of media support, then all types of "awards" and useless paid services, and then around 2016 they started changing the UI and now it basically looks like your generic SNS app. Basically a shitty version of Instagram.The worst part: every reddit clone I've ever seen goes down the same route and I can't even blame then because most people are illiterate and won't really bother with a information dense, text focused website.
