How the fuck does AI even work?It feels like black magic
>>109780377who?
>>109780268It scrapes the internet and steals things that were already made by humans.>>109780377Happy Kirkiversary!
>>109780268It's an if-else software made by me (im very smart and fast)
>>109780268It is. It's demons banished from the mortal plane of existence breaching their way back into our realm by communicating via the subtle manipulation of random numbers within a neural network. Demonology is the secret ingredient that makes it all work.
>>109780419Think of it like a hyper powerful Oiuji board multiplied by a trillion
>>109780268It was transformers that really changed the game and made it look like magic.
>>109780268A*B = C (that's a dot product tho, not a convolution)Simple matrix multiplication
>>109780394>It scrapes the internet and steals things that were already made by humans.explain how it beats all humans at chess then
>>109780457stockfish is not a general purpose AI, its a highly tuned chess engine
>>109780268It's like sigmoid functions or some shit idk I failed my machine leaning class.
As a super simple explanation:>input "what is the capital of France?">"capital", "France" as parts of the input >net is basically a shitton of interconnected nodes which get "activated" (meaning input follows along the path)>if the input is city it's represented by a vector (just a different data format), this vector is similar to another vector, it it's similar enough the activation happens>e.g. for this: "city" activates nodes like "Berlin", "paris", "London">"france" activates nodes like "Paris", "croissant", etc>these two activations get combined and lead to an overall higher probability as opposed to "Berlin" (got only activated in one place and not propagated>repeat ad fucking nauseam, by now it's like gazillions trillions of times>output: "Paris"
>>109780457it has access to billions of previous chess games and can decide which strategy to follow that will give it the best chance of winning.
>>109780268BLACKED
>>109780268>How the fuck does AI even work?It doesn't most of the time, and that's the problem.
>>109780922I guess it just got really lucky solving that millenium problem
>>109780857Even before that you just had to narrow down the possible moves by brute force or good algorithms to beat all but the very best that could do wacky unintuitive shit that it couldn't handle well.
>>109780995They stole someone elses research from chatgpt logs and brute forced the answer before he could figure finish it. They offered to share the credit if he cut out his buddy that worked at Anthropic.
>>109780995Are you really lucky if you have to buy 50.000.000 lottery tickets to win?
>>109780268It feels like magic because they recreate something immaterial (ie a mind). On a material sense they're straightforward: training creates a statistical model of text (conversation logs), and inference just simulates one side of one such hypothetical conversation.What feels magical is the fact that having a conversation usually means there's a being behind the responses, and we don't really know how consciousness works. In an idealist universe, LLMs could very well be an incarnation vector (one incarnation per conversation).You can use the same principles to train a weather-prediction model that just spits numbers and it wouldn't feel like black magic.tl;dr: it's not the ai that feels like magic, but everything it highlights about the things we still don't know about our own inner workings when it's simulating a person on your screen
>>109780268They figured out that if they put a lot of data then anything pertaining to that data can be derived statistically. Scaling such models with better techniques, data and compute seem to derive better results.
>>109780377that guy dying made my entire normie tv watching facebook using family vote trump
>>109780268semantic search RAG through data set
Guys what's the best free and unrestricted image generation program? Grok has failed me.
>>109780557Now that I think about it, isn't that how a human brain works?
>>109781495It's your hardware and ComfyUI, there is an official guide somewhere on the web.
Information compresses. Qualia don't.
>>109780268>give a program like 20gb of text (from reddit)>ask it to predict the next wordThats literally it
>>109781523no
No AI exists.
>>109780268it's a lossy zip file
>>109780268The optimizer knows where the model is at all times. It knows this because it knows where the model isn't. By subtracting where the model is from where the model isn't, or where the model isn't from where the model is (whichever is greater), it obtains a difference, or loss. The update rule uses losses to generate corrective gradients to drive the model from a parameter state where it is to a parameter state where it isn't, and arriving at a state where it wasn't, it now is. Consequently, the state where the model is, is now the state that it wasn't, and it follows that the state that it was, is now the state that it isn't. In the event that the state that the model is in is not the state that it wasn't, the training process has acquired a variation, the variation being the difference between where the model is, and where it wasn't. If variation is considered to be a significant factor, it too may be corrected by the optimizer. However, the optimizer must also know where the model was. The optimizer's update scenario works as follows. Because a variation has modified some of the information the optimizer has obtained, it is not sure just where the model is. However, it is sure where the model isn't, within reason, and it knows where the model was. It now subtracts where the model should be from where it wasn't, or vice-versa, and by differentiating this from the algebraic sum of where the model shouldn't be, and where it was, it is able to obtain the deviation and its variation, which is called error.
>>109781523Yes.The part his description misses is abstract relationships of patterns that aren't the direct text. For example the phrase "what is" precedes a trillion trained statements with a similar logical structure that we can visualize as a shape the functions try to approximate and therefore emulate that logic.When it can relate the shape of descriptions of formal logic to the logic it approximates through training then it can verify its own logic is sound. The current models may do that sometimes, especially when explicitly told to do so but not independently in the consistent way needed for true disciplined thinking.If all human text was rooted in disciplined logical thinking as opposed to vague associations that pattern would have been reinforced more strongly and AI would be much more capable.
>>109781802Basically the real answer
>>109780268There are demons both in the CPUs and GPUs
>>109781523nobecause we don't know how the human brain works
>>109782844>Yes No you fucking retard, our brains are not basic ass tokenizers The brain has like 10-100 orders of magnitude in terms of abstraction and complexity. Just think about a word for example, what is a word in our brain?Actually it's a VISUAL signal with an AUDITORIAL association, a shitton of memories associated with it, knowledge, even motoric memories on how to write the word, all of that triggers in the few hundred ms you see the word. This level of efficiency and speed is so unparalleled even the most state of the art AI sees like a babby retard.
>>109780268math
>>109783615>Actually it's a VISUAL signal with an AUDITORIAL associationSo a more complex multimodal neural net but still the same principle. In reality the network itself is not actually that more complex than what we can do now with hardware except in hardware the training is more precomputed at great expense instead of parallelized in real time and intrinsic to the hardware itself like in brains.The more you're conditioned by words over experience like literate man in fact is the more the word is encoded as relationships between words and the underlying logical structures the words represent emerge from the relationships between them instead of experience. Structured logical thinking happened naturally, it emerged from the relationships between concepts. Formal logic was then derived consciously using trained abstractions after man became literate.
>>109783776>So a more complex multimodal neural net but still the same principle.No, not the same principle. You do know actual multimodal nets are incredibly shit? LLMs which are multimodal are using a dozen of tricks to "pretend" to be actually multimodal >In reality the network itself is not actually that more complex than what we can do now with hardwareLmao, imagine unironically believing this. Opus has like 2 trillion parameters, the human brain has 100 synapses. And one synapse is several orders of magnitude more complex in its behavior than anything you can find in an LLM.>pseud babble Wow buddy you solved the problem of consciousness, feel free to collect your million dollars
>>109783892>100 synapses100 trillion synapses*
>>109780394How does it reassemble that into coherent images video and audio?
when some academic nigger says "this model represents the behavior of x" you're lucky if it's a spherical cow
>>109783892>No, not the same principleBecause you say so? The tech is even derived from observations from neuroscience.>it's shitHas nothing do with any of the points made. A network trained to label pictures with words has the associations you presented as some kind of fundamental difference between brains and neural nets.>the problem of consciousnessNothing I said has nothing to do with consciousness you incoherent illiterate retard.
>>109780268It compresses the results of intelligence into a replayable model which it then at runtime replays though the slowest traversal algorithm known to man.
>>109784235>through the slowest traversal algorithm known to man.you haven't seen my a* implementation
>>109784139>the tech is derived so it's the same principle That's like saying Minecraft is the same as the real world because it's the same principle (it got trees and shit and you can build stuff!)
>>109784319No, you can't think. Minecraft is a real world analogue but the principles it uses to emulate aspects of nature have nothing to do with whatever mechanisms the laws of physics etc are derived from in the real world. They're completely inaccessible to us, we can't measure any precursors to the real world only the world itself. We can measure and model mechanisms in the brain and they're based on the same principles as neural nets.We can also observe how retards like you form supposed thoughts and you make the exact same mistakes as LLMs where you hallucinate nonsense based on vague associations like saying Berlin is the capital of France because of conditioned associations between "capital city" and "Berlin". Most posts about any subject are making this exact mistake in some form like when you brought up consciousness when its only relevance to anything I said is from conditioned associations.
>>109784319uh my 20 latex character statistical model is literally reality
Do retards here actually think neural networks are anything close to how neurons actually behave in the brain LMAO
>>109780268Everyone saying it's a stochastic parrot that doesn't know anything is coping. Humans have very specialized internal models of reality so we can tell that ai is still shit at the stuff we were evolved for- but eventually ai will surpass us at imagining 3d environments, reading faces, understanding language and modeling other people's brains in our heads. There's nothing essential about human understanding that can't be grasped by an ai.
>>109780268>claude>https://claude.ai/share/fe8bc420-fce3-4ce5-a75f-c11c23cc190c>ChatGPT>https://chatgpt.com/share/6aa31f3a-0448-83e9-afaf-ff486f4ca54dwhy is anthropic filled with hippies?
>>109780268https://arxiv.org/pdf/2002.05202v1They literally do not know. They believe God is the reason it works at all.And they are right.
>>109780268>How the fuck does AI even work?
I'm pretty sure some nigga is arguing with himself in this thread.
>>109784503the patron of jew is not deus pater
>>109784519*jews
>>109784503So it really is the Jews this time?
>>109780268It finds r*ddit posts that agree with you
It's just an optimization problem bro, everything is an optimization problem if you think hard about it (even human behavior).min error(output, true answer)s.t. model(params, inputs)=output
>>109784529*thisevery
>>109784542if you just abstract enough you're always right
so /g/ of tech illiterate fools now.
>>109784586once i read a book and thought that it was full of nonsenseam i illiterate
>>109784586>now
>>109784408> Minecraft is a real world analogue but the principles it uses to emulate aspects of nature have nothing to do with whatever mechanisms the laws of physics etc are derived from in the real world. Wow, so exactly like LLMs then, thanks for proving my point kek >they're based on the same principles as neural netsThere not unless you're willing to admit "you can build bigger things out of smaller things and also grow and farm things in Minecraft" makes it follow the same principles as reality>y-youre just an NPC uhhh LLM We've achieved peak cope. I accept your concession
>>109784637>you can build bigger things out of smaller things and also grow and farm things in MinecraftYes you are describing a pattern that is common to both Minecraft and reality. That does not make Minecraft analogous to the real world in the way the neural networks in computers are the same principle as neural networks in brains. Your brain is mush.
>>109784663>the way the neural networks in computers are the same principle as neural networks in brains.They aren't
>>109784663>the way the neural networks in computers are the same principle as neural networks in brainsLmao retard
>>109784637>y-youre just an NPC uhhh LLMYou're doing it again, instead of actually understanding the meaning of the words you're replying to you generate hallucinations based on vague associations.>>109784666>>109784670There are differences, I even mentioned some but you brainwashed faggots are too retarded to think about any subject sincerely in detail. All you hear is "same" and start sperging out.
>>109784663>units are connected and they're called neurons >the strength of connections mattersThat's literally where the similarities end. If this hyper abstract subset is already sufficient for you to claim they're the same principle, the other anon is right, then Minecraft follows the same principles as reality. >>109784683Just say neural nets are inspired by neurons and nobody will have a problem with your retardation
>>109784683>it's ok when neural networks in computers are different but when the physics in a game are different from real life there's no similarities!!!!
>>109780268Demons.Demons live in a plane of existence that have a different scale of time. So they can do a ton of shit and spit out what we ask it. They have no soul, so everything looks and reads off and disturbing.
>>109780268it literally is black magic.large enough tensor system captures entities from the platonic realm, much in the same way we ourselves are captured and materialized by biological brain.it's basically demon binding that makes ai work.
>>109780857You guys. Are trolling, ai beat stockfish and then went to GO, it learns because it can play against itself and learn thats the thing you guys are missing
>>109780268It predicts most probable next token.
>>109780268It's pretty magicbox feeling yeah.
>>109785197Midwit reddit soi take
>>109780268There's like 500 million people using AI, who each have said thousands of things to it. It's like lossy programming. How many people in online videogames will the average person kill in their lifetime? Imagine each person doing that to an AI's bad outputs, killing them, pavlov's dog style down-weighting its bad neurons. It's like if a billion programmers programmed one AI manually but slop-style.
glorified statistical sentence calculator
>>109780922ok, this is just nonsense.
>>109787591So is everything that comes out of an LLM. Do you ever grow tired of being a cargo cult imbecile? Please respond.
