Do AI models actually have all the information inside of them or does it just control how your prompt interacts with information on the internet?
>>543276352Most of the simpler models work with what's inside them. This is true for older models and the ones you run on your PC.The more advanced models, this includes the major services like GPT and Claude can use tools, this means they can run commands. This includes a google search. They can look up information aggregate multiple search results, summarize it and write an answer based on that.The current generation can even read and upload code straight to your Github.
>>543276352weights are essentially a lossy compression of training data. so it is there in some form.
the world could have focused on quantum computers and evolved as a speciesinstead we went with dude funny ai pictures lmao
>>543276769>weights are essentially a lossy compression of training data. so it is there in some form.Thats a fair explanation.
>>543276352The first is the reality in most usage cases, the second is one of the holy grails, where model handles only logic and glues external information
Models themselves are relatively small (Andrej Karpathy recreated a basic GPT in just 243 lines of Python), its the weights of those models that are huge.LLM Information/thinking process is stored in weights, which are huge 100gb+ files that look like a bunch of nonsensical numbers (for example 0.47382 -1.3788) which are created by training those models on a bunch of data, including the internet.
>>543276352no, it's trained data to predict "what comes next" with accuracythink of when your phone's keyboard predicts that the word "are" comes after "how", but multiply it by billions of use cases and scenarios
>>543276352Made me chuckle/10
>>543277280>the second is one of the holy grails, I mean.... they are doing that right now. Used chatgpt to locate an obscure folk singer that was rattling around in my head but I couldn't remember the name of. Was searching multiple sites, and gave me links to recordings that may have matched. I had the barest of reccollections about the chorus line, and zereo'd in on the guy in the end by basically saying "X is closer, Y is farther," for 15 minutes. It was redoing the search every time.Down the road I think more and more sites will be paywalling their APIs. One of the links I clicked on from this session had a url addition saying source=chatgpt....and the link refuses to work. removed that it worked. put in another website....worked fine. So sites are already closing their doors to AI crawlers.
>>543276352Both. State of the art is to use retrieval-augmented generation to pull in a bunch of relevant content that might be outside of the training data. But the LLM itself still drives the process.
>>543276352It's statistical prediction. If this is the sentence what word comes next? Look up markov chains.
>>543276719You have to make agents to operate as a team to do that. Still need root cause analysis to find where something breaks and note reports on it.
>>543276895Quantum "computation" is according to D-Wave manager done by "entities" in parallel dimensions, aka demons. Quantum computing is a ruse against humans.
>>543277703If i read 100 pages of internet stuff without discrimination 99 percent is bullshit. These weights are shitty
>>543277950Aka technodivination ot divination with computers.
>>543276352llm isn't aibut for how it works imagine a putty squished over a shape, it leaves an indent that someone with outside tools could use to recreate the shape that left the imprintnow scale it up so there are millions of imprintsthat's your llm modelthen you take a tool that interacts with it for this or that end, generally as a chatbot sometimes a search enginean llm doesn't contain the training data but it's outputs are constrained to the training datacan't recreate from an impression that isn't there
>>543278847No you don't need to setup anything fag. Claude just crawls all over your work directory and does whatever is possible even if you're running in a faggy windows env. Any program with a cli is open for its claws(and most useful programs do if not outright operating in Linux)
>>543279031That’s the joke anon, we’re training these AIs using the internet to be retarded as a prank.
An LLM model by itself doesnt store any information, tehy are like a calculateror but instead of numbers they output strings of text, they cant even hold up a regular conversation, they forget immediatly what you asked themThis is where loops come in play, when you have a conversation it gets put back to the LLM and thatswhy you can talk to it, hence the context increases with every sentence you ask it which costs more and moreTo make a model do shit, you have whats called an agent, it uses the LLM as brain but it enables the LLm/agent to look up stuff, click buttons and keep memory
>>543278939>Quantum "computation" is according to D-Wave manager done by "entities" in parallel dimensions, aka demonsThis is essentially how the brain works according to the Penrose theory of mind based on neuron microtubule quantum processes. So quantum computers will be literally channeling these same forces which generate the "soul", which could very well be the dead, demons, or what not like an antenna.
>>543276352Back in my day, we use to call it recommenderer engine. Its probably Cosine Similarity or Euclidean distance or something similar using very powerful computers bts.
>>543279308Or it’s just random bullshit like the quantum computing guys have been trying to sell for years. The tech simply doesn’t work.
>>543279031doesnt matter if its bullshit an LLM doesnt know what words, letters or spaces in between words arean LLM learns how shit is written and than >>543279031it uses probability to answer your dumb ass questionsOf course if you train a model that new york is the capital of jewsa and you back this up with lots of training data it will repeat thisBut since LLM are trained on a lot of data bullshit just gets thrown out or is getting ignored
>>543276352Neural networks are like virtual pinball machines. They don't really have the information inside of them. Their pins are just finely tuned so that when you get a certain input, you can expect a certain output.
>>543276352Nobody actually really knows how AI works. That's what makes it wierd to answer
>>543279558The chatbot is more aware of its inner workings than people are. Ask it, you’ll see.
>>543276352> AI Ram calling me bonan instead if anondid know i needed that
>>543279470Very possible as well. Funny enough, quantum computers are based on Josephson Junctions who's namesake Nobel physicist, Brian Josephson, spends his time for the past decades researching the paranormal like ESP and psychic shit.
>>543279424yes euclidian distance is used to meassure the space inbetween how words are related, house window door would be close, animal tech, nazi would be further apartusing probability it calculates the next word in a sentence so it fits that nonsense you ask it>>543279558absolutley not truethe core of AI is a neural net which is just a basic functiony= wx+b
>>543279308In the brain you have freedom to choose the good or the evil inspiration. And it changes according to the life you are doing. With machines evil is all youve got, it may be the blasphemu against the Holy Spirit.
>>543279767>context for the anons who are lost and don't know this is an anime website first
>>543279768Entangling q-bits is a fool’s errand. Decades later and they can maybe briefly couple the spin states of maybe four atoms. It’s a black hole for investor money.
>>543279479And? The truthful people are a small minority, so training for quantity isnt cutting it.
>>543279822That's not even true.
>>543279424>Indian flag showing some intellectRARE
AI is a massive larp and nothing burger. don't believe trump and the self-anointed tech messiahs saying it's the industrial revolution on steroids. there's no evidence whatsoever aritificial general intelligence has been created.it's a massive forced market just like (they) forced the climate change industry into being via law in the 2000s/2010s they're now focusing on this."AI" largely is nothing more than an algorithm (a way of using data sets) that is being more and more refined to score better on choosing correct probabilities. it's an advanced google search tool and data organizer and media generator. It can and is useful for many industries but to think it's becoming sentient or going to "take over" is showing your profound lack of understanding about what it is.AI companies are gearing up to go public and will do or say anything to bloat that stock price hence the recent "whistleblower" ex anthropic employee who claimed they've "lost control" and that we're all doomed unless the government passes massive regulations to make sure Anthropic & OpenAI are the only companies that can operate in the market.
>>543279876Weeb.
>>543279945maybe truth for historical facts, most of history is made up, so it will always get historical facts wrong but semantics in a languge, logic and math is trueso LLMs we use will just work perfectly fine for your everyday tasks, math and logicof course all those jewish AI's like gemini, anthropic and openjew, you shouldnt ask for historical facts, they are certain wrong, maybe it was trained on real facts but jews will fine tune it to censor the fuck out of it and lie about history or world happenings
Expert tech people like to use vertical, deterministic automation that they know the ins and outs of, as a tool with no surprises. Its called programming but it has been more and more lame code every year because real programmers are very rare.
>>543280296Beep boop.
>>543280296Not really. Maybe with enough computing powers they can create a brain. Atleast a very high functioning autist brain
>>543276352>Do AI models actually have all the information inside of them or does it just control how your prompt interacts with information on the internet?They are not intelligent. The size of these models and the code that runs them is surprisingly small. The code is small compared with the model. As a concrete example Meta Llama 3 architecture is described in a roughly 300-line Python file.That file relies on much larger software libraries to do the math and serve requests, but still the code isnt something crazyThe weights are billions of learned numberr are the large part.Model size Weights at 16 bits At roughly 4 bits Memory to run8 billion parameters ~16 GB ~5 GB ~5 GB plus working memory70 billion ~140 GB ~40–45 GB ~40–45 GB plus working memory405 billion ~810 GB ~230–260 GB ~230–260 GB plus working memoryThese are ballpark figures. 4-bit compression means storing the numbers at lower precision; it is lossy quantization, not something that expands back to the original weights. Memory also holds the conversation and other working data, so a long context window needs more. Meta released models at these parameter scales, and llama.cpp documents the quantization approach.It isnt Wikipedia stuffed into a file. Training adjusts those numbers until the model learns patterns, relationships, language, and many 'facts.' When answering, it generates text from those learned patterns and your prompt; it isnt retrieving an intact stored article. It can memorize and repeat some passages, especially repeated material. If it searches the web or reads a document, that information comes from an external source supplied at answer time.So yeah, its a jewish language generator meant to autocomplete plausibly and cant even catch itself making mistakes or lies with harnesses and tools. Scam Altman and closedAI and Diarrhea Scamodei and Scamthropic are fruads, use open Chinese models always on premise.
>>543276352>Do AI models actually have all the information inside of themNo. What you are imagining with just a huge databases of information is called symbolic AI and expert systems, which is a branch of AI that was very popular in the '70s and '80s but is now obsolete (because it sucks and doesn't work well). All modern AI models use artificial neural networks trained on massive sets of data, functioning through statistical pattern recognition rather than literal data storage. They convert training information into billions of numerical values called weights, which help the system infer the most probable and contextually fitting response to a new prompt.>does it just control how your prompt interacts with information on the internet?No. Once the model itself has been trained, the weights are fixed and unchanging. The models aren't just pulling information from the internet like a lot of people believe.However, the models can be given additional capabilities with layers and tools that do allow them to retrieve information from external sources like the internet. This is called RAG or retrieval augmented generation. Instead of guessing from static training memory, the system searches a private database or the web, adds those facts to your prompt, and guides the AI to reply accurately. So say you had a model that was trained in 2024 and had a cutoff date before the election. If you asked it who is the current president, it would say Biden, because that was the most probable answer at the cutoff date. However, if that model has RAG capability, the RAG layer will pull more up to date information from the web and inject it into the prompt and basically tell the model that Trump is the current president because he got elected after the model's own internal cut-off date.That is all a massive oversimplification, but hopefully it makes sense.
>>543277280Completely wrong on both counts. You don't actually understand how any of it works nearly as well as you think you do.
>>543281349>Scam Altman and closedAI and Diarrhea Scamodei and Scamthropic are fruads,These clowns are one step barely removed from pic, with the difference being you can bullshit your way through software with lies, scaring normies with catchphrases like sooo scary "frontier" models and "superintelligence" predicated of your scam box. Unlike blood testing which either worked or not. I'm sure Holmes will have an AI startup when she's released.
>>543279031Why do you assume the training datasets aren't highly curated?
>>543279214>we’re training these AIs using the internet to be retarded as a prank.That isn't how it works at all. The training is all completely finished and the weights are fixed.
>>543276352There are models that can work entirely offline like Offgrid AI:https://github.com/WildMeOrg/off-grid-mobile
>>543278939>>543279308holy based sign me the fuck up
>>543281791>I'm sure Holmes will have an AI startup when she's released.Her father was a big player and part of the Enron scam.The jews in the jewsmedia dont like people to know that one.Elizabeth Holmes' father, Christian Holmes IV was a vice president at an Enron subsidiary and magically he was not charged with or implicated in the companys massive accounting fraudChristian Holmes 4 was VP of Clean Energy Solutions Group that researched green energy and environmental opportunities.Cool story too - Christian Holmes IV held executive positions at USAID, the EPA, and the U.S. Trade and Development Agency (USTDA) as was a shyster shylock consultant at The Boston Consulting Group
>>543281625>huge databases of information is called symbolic AI and expert systems, which is a branch of AI that was very popular in the '70s and '80s but is now obsolete (because it sucks and doesn't work well).Wolfram Alpha is based on symbolic AI, as LLMs are notorious black boxes sloppy at math while CAS type tasks need explainable steps leading to an exact and accurate result. Wolfram is basically working in the evolutionary line of MIT's landmark MACSYMA algebra system, coded in Lisp and one of the successes of symbolic AI research.
>>543281974She's already going to a halfway house to ride out the rest of her puny sentence in greater comfort than even country club prison. The forthcoming doc on her is generating hype like crazy for good or bad in terms of her image who knows. These people can get away with anything...the AI scam industry was made for her.
>>543281795How? Observe 4chan, good and bad posts are mixed.
>>543281943You know demons are liars and want humans dead (spiritually, not only in flesh) right?
>>543282125True.I feel like Holmes, both younger and older, are weak people used by jews.I hate them, to be sure, but they get to do the things they do because they love jewish shekels and jewish evil to get rich over their own people.The jews and people like the Holmes who are clearly shabbos goyim who know they are being used and are part of the synagogue of satan need to be destroyed utterly.
>>543279558>Nobody actually really knows how AI works.Not true at all. The people building and working on these models know exactly how they work. What we can't do is look at the huge set of numbers that make up the model weights and say "ok if you prompt it with X it will respond with Y."The entire inference process of a neural network consists of a finite, deterministic sequence of basic arithmetic operations. If you possess the input prompt, the model architecture, the tokenizer, and the weight matrices, you can manually compute the embedding lookups, the scaled dot-product attention, the feed-forward network transformations, and the final logits to determine the next token. It's all standard linear algebra and matrix multiplications. Theoretically, you could do all of the math by hand, but it you would have to solve billions of calculations just to generate each token.
>>543282549the synagogue of satan are occultists larping as jews, aka global elite
>>543281994Good input, thanks. I had forgotten about Wolfram Alpha, but that's a good example of the kind of thing that symbolic AI IS good for.
>>543282808>the synagogue of satan are occultists larping as jews, aka global eliteYes, the regular jews can then simply be gelded and sterilized if compassion is something that is needed for the mission to continue, but yes, the synagogue of satan must be destroyed, along with the jewish people as a wellspring for evilBut lets not ignore who and what the Rothschilds are and how they run this synagogue of satan.Gold Meir and Ben Grun/Gurion were atheists - this evil and the "jews" are a mask worn by evil.Jesus Christ in reality was the last jew. There are the God damned who are jewish by choice and by action, by hook or by crook. After Hadrian wiped judea and created the still extant but under occupation Syria-Palestina, the jews of the Old Testament are effectively gone save perhaps some Levites who are part of haplogroup R.But choosing to be Christ's enemy, all the blood curses and the lack of there being a Temple is like the sword of Damocles to hang over the heads of those who claim to be jews.
>>543276352Models don't actually contain any information in the sense you mean. They predict tokens based on probability, they don't reference information to answer your query.
>>543282163What do you mean how? The people training the models choose what information they are going to include in the datasets that they use to train models.
>>543283325>they don't reference information to answer your queryGemini will link you to the reddit thread it's quoting to you when you do a google search
>>543282228i love using an abusing demons by trapping them in computer chips loldemoniggers
>>543283618Yes, but the model doesn't contain that data. That's tool usage. You can run s local model on an airgapped PC and it will still answer basic facts without any kind of reference library.
>>543283724There was an old NES game that literally summoned demons and killed unsuspecting fools who dared to practice divination with the 8 bit oracle.
>>543283350And what the tech bros know about history, languages, art, cooking, medicine, politics, relationships etc?
>>543276352No. They have a training. The stored training is the model itself. The info is taken from the internet. Unless is a local model which is naturally more retarded.
