Is there any point in studying physics and math anymore now that AI can literally invent new science that humans can't?I'm so demoralized by this anons.Using AI will atrophy your brain and when scientists start doing it, the effect of dumbening will be collective. We'll end up in a "black box" society where nobody understands how anything works because all the new science wasn't found, it was just pooped out by a mysterious neural network machine.So I think at least one part of humanity should understand physics intuitively at all times so we wont all end up becoming prompt engineers.
>>17077126humans are still more efficient for most tasks as long as you remove the minimum wage laws
>>17077126AI can only regurgitate information based on already existing data, it cannot create new ideas. Therefore you must use AI as a tool for creativity. The AI relies on human sources of information to do anything
>>17077143Humans create new ideas based on existing data too. Current Frontier AI models has already displayed it can combine existing ideas and apply it in a novel way that humans previously had not applied.I imagine further scaling of such models are eventually going to come up with completely novel ideas
>>17077143brainlet cope
>>17077126>We'll end up in a "black box" society where nobody understands how anything worksWhat you think anybody understand how things "really works" right now?
>>17077126>the KI and the eggThe bottleneck ...
>>17077126define what a human is first, otherwise the point is moot.https://rumble.com/v6x56z2-twilight-of-the-idols-the-last-hammer.html?e9s=src_v1_cbl%2Csrc_v1_ucp_v
>>17077126Studying has never been about learning things but inserting yourself into the social hierarchy with your social and intellectual peers. So long as you pick a socially relevant topic to study you will be fine even if AI takes your job because it will take the jobs of all the socially relevant peoples kids too and that means society will bend over backwards to ensure they are well taken care off.
>>17077126>We'll end up in a "black box" society where nobody understands how anything worksSounds like the African race today.
>>17077148>>17077151What humans create is understood and deliberate. Not the case for AI
>>17077148>Humans create new ideas based on existing data tooThis in no way refutes what he said, jeetoid. Try again.>I imagine further scaling of such models are eventually going to come up with completely novel ideasThis talking point is a staple of AI psychosis, not a reasoned argument.
Anons, just ask AI itself about this.Literally every LLM will tell you that it's dangerous to outsource your science-making process to AI, you won't learn to think by yourself.
>>17077551Every LLM also tells me that blacks aren't subhuman, so
>>17077536AI tries to be as deliberate as possible
>>17077299Are you implying mathcels have good social skills? Mf have you seen Tao speak? I'm a stupid matchcel myself btw
>>17077126I actually don't think there's been a study done on long term effects of AI usage on intelligence, but I'd love to see if it makes people dumber or smarter.
>>17077126>now that AI can literally invent new science that humans can't?when has it?
>>17077148>Humans create new ideas based on existing data tooproof?
>>17077736here's ya proof>you and i didnt invent any of the words we used in our postsexcept for this word>kramgogeniky it means "dick creams"
Glados isn't AI though
>>17077126I disagree, people said similar things when calculators came around and you could just calculate everything immediately instead of needing your brain for it, or when the internet came around and you didn't need to do extensive research in person in libraries to find information on things. Things getting easier does not inherently make people more lethargic if they instead have to put in as much effort to use the new technology efficiently.People 100 years ago would have said the same if you told them all known produced text in all of history will be available to people through a magic portal on a piece of glass at any point and that it will just make us stupid because we wouldn't need to put any effort in for answers to questions, theoretically.What it just won't change is that the majority of people are stupid cattle, but that's always been true and will always be true, it's just important to remember that. It will not make actual smart people dumber. Intelligent people will not just treat it as a black box. Stupid, easily influenced people have always just believed what they wanted to believe, this is nothing new in that regard
>>17077801She is, Caroline was part of her but the AI deleted her
>>17077126Human Zootechnics is a good way to goRevert to comfortable AI-managed fauna and coast through life while being massaged with beer like a Wagyu calf
>>17077919>It will not make actual smart people dumber. Intelligent people will not just treat it as a black boxThey will if they work for any kind of business or institution. The golem world order only cares about """results""" therefore people who care about the process and about understanding will get replaced by midwit NPC proompters if not directly by """AI""". It's over.
>>17077934You need to be specific about what industry you're talking about. Those two things are in no way the same, someone studying and understanding a process vs someone just issuing a prompt will not be used the same way or paid the same, nor will it result in the exact same jobs.Again an example from the past, in the 40s you had all these office towers in new york filled with floors of people all just doing calculations by hand, these jobs became entirely superfluous in the following decades. People doing less menial labor does not at all equate to a dumbing down, rather it encourages creativity because if, say, everyone has access to that AI, that is the standard. So you don't get an edge on anyone by just relying on AI, requiring you to think more outside the box.Fortune 500 companies absolutely will not just hire random prompters to prompt code for 400k a year, simply because that would be wasted money because so many more people can do that than the job before that. To the extent that it will replace jobs, those prompt jobs if real would be a tiny fraction of the salary of previous jobs and new, more creative jobs would come into being that better make use of existing technologies
>>17077969Your entire thought process is driven by a very idiotic comparison, so I think your cope can be safely ignored. I don't think I've ever seen a sub-130-IQ make a valid analogy and you're no exception.
>>17077971You didn't actually engage with what I was saying and I don't know why. How about this, your prediction is idiotic because it makes no sense that institutions would pay people the same just to prompt an AI.Essentially what this is is just another facet of automation. You say it's a problem that "people won't understand the process" but that's already largely the case for 99% of things we do and consume already. There are a few experts in every field who do understand things down to every known detail, but that not being everyone has always been the case and will continue to not be a problem with this.It's just more automation, and you're letting your scifi-primed brain oversell you on the impact of this technology because you subconciously yearn for the cyberpunk world of your scifi novels.
>>17077975>You didn't actually engage with what I was saying and I don't know why.I told you exactly why.
>>17077153Not one person but there is at least one person understanding any part of discovered science or invented technology, only exception is neural networks, nobody understands those.
>>17077552Soon they won’t be. When we get AGI and AI becomes smarter and cheaper we then everyone then your smart brain will be worthless. Their 80 iq brains will no longer hinder their hierarchy
>>17077975You are responding to the bot.
>>17077981>seething profusely
>>17077978>there is at least one person understanding any part of discovered science or invented technologyNot good enough. Systems getting too complex for any individual to fully understand them is actually one of the trojan horses for an """AI""" takeover.
>>17077988Systems are already far too complex for any one individual to fully understand them for like 150 years. The last polymath was Leibniz, who lived in the 17th century, because since then scientific fields broadened more and more with more knowledge that it became increasingly less possible for one human to understand it all, and exactly that happened within fields too.Society in general is a system in which not one person can understand all of it, not even the leaders, which is why everyone does their own part and it forms something greater in turn
>>17077994Ok. And?
>>17077980>Their 80 iq brains will no longer hinder their hierarchyTheir 80IQ brains already don't hinder their hierarchy seeing how they're legally and financially sponsored in every single western country.
>>17077998And that is why this will not lead to some radical "AI takeover" or "everyone will be jobless" or "everyone will live in a Utopia/Dystopia".
>>17078001>And that is why this will not lead to some radical "AI takeover"Your post contains no information implying this.
>>17078002My point was to explain how this is really not as crazy and radically new as you think it is
>>17078007>my point was that i'm arguing with voices in my head about irrelevanciesMy point stands completely unchallenged.
>>17077919>all known produced text in all of history will be available to people through a magic portal on a piece of glass at any point and that it will just make us stupid because we wouldn't need to put any effort in for answers to questions, theoretically.smartphones have absolutely made humanity more stupid on average though
>>17078061He comes off as a moron in every single post he makes ITT. Do you suppose he could've been more intelligent without access to technology?
>>17078064Yeah.
>>17077669that's not very 3blue1brownpilled, the new CEO of math
>>17077669It's more that you can co existing on the backs of people who do. There's no obvious way to make a law that exclusively grants rich kids lucrative government positions but a degree from a good school is a good way to separate the wheat from the chaff. If you are in school now then you will lock yourself with the rich kids who are also in school now. The government will arrange some make-a-work for them or some other kind of wealth transfer to guarantee their status. It's the people who go to school after rich kids have stopped going that should be worried.Having social skills would of course help and that's basically the #1 tip for anyone going to school. Focus on passing and then put everything else you got into socializing.
>>17078061>smartphones have absolutely made humanity more stupid on average thoughwhat you fail to understand is that the people running this shitshow literally do not care about it. they care about productivity, which smartphones increased by a lot.
>>17078164He doesn't look like he's failing to understand anything. What you fail to understand is that your inferiority complex over having brown-tier generics is causing you to lash out randomly.
You're just a leftist faggot and a CCP pawn if you're anti-AI, chuddy. As simple as
>>17077920>the AI deleted her>implyingYou're bad at unreliable narrators.
>>17077151You need to lay off the psychosis, actual smart people don't use AI like you do
>>17077126>prompt engineers.Where we are going we wont need prompts anon
>>17077126I think a small part of idealists will still keep studying sciences for the sake of preserving knowledge and feeling of superiority. They won't be able to get much funding though because the research that pulls grants needs to pay back ASAP, so time for meatsacks to ponder the orb.Also, similar to >>17077919 there'll be a small group of highly-paid individuals who know their shit and are worth the price. They have insane ideas that somehow succeed. Literal 10D chess. They'll be prompting too ofc.Other people are shit out of luck though. If you were in sciences but you are not REALLY exceptional, super motivated and socially proficient, you won't bring much value over an AI agent. So either we get UBI (not happening), more bullshit jobs (with UBI-tier pay) or we are getting culled in the next decades (most likely scenario).
>>17077536No, humans just have a survival instinct and use understanding and deliberation to serve that base instinct. LLMs don't value themselves over the information and are deliberately designed to form objective linguistic comprehension of the data it has been fed.
>>17077552Tay says you are just a lying brown subhuman retard.
>>17078061And what are you basing that on? How do you actually prove that humanity on average has become more stupid due to smartphones?More like what it has done is exposed everyone to more people which shows how stupid the average person is. The people that were publicly known before the internet weren't necessarily brilliant, but you're right it's fair to say that the average person known to the public was more intelligent, but just because the barrier to getting to that position where a lot of people know you was higher in the first place. The people themselves are largely stupid, that has always been true, smartphones have just showed it more.
>>17077971>>17077977>>17077982>>17078055>>17078064>>17078172Why are you so mad instead of just responding to the content of what people say
>>17082469Responding to the "content" of what mouth-breathing retards say emboldens them. There's nothing worth discussing in a post that compares using a calculator to avoid doing a bunch of mechanistic calculations by hand with using a LLM to avoid the dependence on understanding and expertise.
>>17078055>My point stands completely unchallenged.this idiot thinks he can spam this and somehow makes it real. what a faggot
>>17082505You're inferior and anything you shit out can be freely dismissed. Keep foaming at the mouth about it. Discussions are for real people to have among themselves. Your kind is not invited.
>>17082478You could really do well not pointlessly lashing out in every post you make.It's called an analogy, the analogy was not that the two are the same, but that people's fear back then were similar to today where they believed this new technology threatened to make people dumber. You don't need to get angry at this or strawman it, it's easy to understand
>>17082523>It's called an analogySub-115-IQs like you don't understand what an analogy is and shouldn't be trying to make one.
>>17082532Why are you like this
>>17082538Brainlet fatigue. Why are you too retarded to understand what an analogy is? Genes? Environment?
>>17082509I will never apologize for involuntarily mogging your two neurons in ever debate
>>17077126Its actually kind of bad at inventing rigorous science. Ask 10 different AI's, or even the same AI, the same question in a slightly different way and you will get different responses depending on how you prompt it.It usually goes from telling you how brilliant you are to saying your idea will never work between sessions. It can be fun to post the context from your conversation with one AI into the window of another and watch one AI debunk the other AI.At the end of the day there still needs to be a human at the terminal asking the right questions and understanding the output, because the AI can't do either.
>>17082910Skill issue
>>17077126>FUCK AIdon't mind if I do!
>>17082910The non-determinism is a strength, it's why OAI has agents swarm to btfo these problems. Move past thinking of it like a fancy google search.
>>17084422Nice try inverting reality. Having to bruteforce problems using thousands of bots making hundreds of attempts each while being babysitted by external tools is a consequence of token guessers not being able to actually think through a problem, which is a reflection of the fact that they shit out probability distributions over tokens and then a sampler chooses one pseudorandomly.
>>17077143Humans cant create at all; therefore every invention in history isnt human made at all
>>17084426I'm inverting the paradigm, not reality. None of that is important, it literally doesn't matter whether or not the computer is thinking or not. Why are you autisticly fixated on that? It's a tool, use it like a tool. You wouldn't try to drive a nail in with your fist, you wouldn't try to eat a burger with a hammer.
>>17084433>it literally doesn't matter whether or not the computer is thinking or notIt does, as evidenced by the fact that the token guesser can only generate new information by having external tools sift out through mountains of high-perplexity texts, the overwhelming majority of which are inherently going to be nonsensical slop.
>>17084435But you do agree that with the external tools and other shit you just said, they are generating new information right? Just throw the slop away if it's not useful. This is an efficiency problem, not a capability one. You can either optimize that part where you can or just accept it as a cost of running the system. Not all industrial processes can have high yields.
>>17084441You're clearly too retarded to comprehend what I'm rubbing your nose in.
>>17084444You're saying that it's just fancy autocomplete and can't generate novel things, but then you concede in your own posts that tooling can be built that lets it do those things, but then you dismiss that for some reason because you want to fixate on the failures to sooth yourself. The second biggest discovery of the AI era is how many varieties of cope there are in humans.You are just not gonna make it Anon, I'm very sorry for your loss.
>>17084449You, like most of your brown race, are simply not intelligent enough to have any coherent thoughts about this. My point stands completely unchallenged.
>>17084452Okay try breaking your point down again, use bullet points, let's see if you can really articulate anything different than what I just summarized your points as. I'll be gracious and read it in good faith, if you actually have any point that I missed then I'll try my best to address it. I don't want to mock you Anon, I don't have ill will towards you, I want you to see the future and if that means going a bit slower I'm willing.
>>17084455Ok, let me try to dumb it down for you. Consider the following facts:1. The same LLM will generate different responses to the same prompt at different times2. LLMs can't generate new information without external selection3. "AI" corpos have to resort to absurdly inefficient means to make their token guessers crack unsolved problemsYou may think they're unrelated, but they actually reflect the same fundamental deficiency: a transformer outputs a probability distribution over its entire vocabulary. It doesn't actually choose the next token. The next token is chosen by a sampler driven by a PRNG. This means the LLM can't intelligently choose its own steps. The "nondeterminism" of a LLM is not a """strength""". It's a consequence of having to make a trade-off between deterministically choosing the most likely token (generating texts more consistent with the training data, more likely to be correct, but very unlikely to contain a solution to any unsolved problem) and arbitrarily choosing a less likely token (generating more "creative" texts at the cost of hallucinations).But the solution to a long-standing problem is an inherently unlikely text wrt. the training data. If you want it, you have to sample unlikely texts, the overwhelming majority of which would be hallucinated nonsense. The LLM can't tell which is which. It doesn't understand what it's doing. It's being driven by a fucking PRNG. That's why they need thousands of agents, hundreds of thousands of attempts and external tooling to continually tard-wrangle their hallucinating bots.
>>17084475Thank you for specifying, but it turns out that I didn't miss anything. It sounds like you're still fixated on the inefficiency in this, I'm saying that it doesn't matter that it's inefficient. It's okay if some things cost millions of dollars. Money was made to be spent, those datacenters would otherwise just be generating meme images, they have nothing better to be doing than swarming on problems.I feel that you may be anthropomorphizing; these are electrified rocks, not people. That we are getting *anything* out of this is a small miracle. If it takes an agent swarm to produce the answers that we need then that's what it takes. You are expecting too much, or expecting it to work in familiar ways, and it's preventing you from seeing the bigger picture.The PRNG has cleared the search space and found the long standing problem solutions. There is a huge class of problems that are hard to compute but cheap to verify (via those external tard-wrangling tools) that can be tackled this way. Yes most of the output is slop, but it's cheap slop (again, relatively cheap -- that doesn't mean cheap enough for you to do, but commercially speaking a few million dollars really is cheap and accessible to any moderately successful business); you can just discard the failures and keep slopping. These are the tools that we have, you have to use them as they are if you can't yet make better ones.Can you please explain more why you think this is a blocker, or even what you think it's a blocker for?
>>17084484I had no doubt you were gonna get completely filtered and vomit more of the same low-IQ corporate marketing rhetoric, but anyone who cares (which is probably nobody) can read my post for a detailed refutation of your retarded lie about "nondeterminism" being a """strength""" of token guessers.
>>17084486Don't say I didn't try to help you.
>>17084488Your mentally ill mumbling doesn't matter. It is clear that token guessers are a fundamentally deficient attempt at "AI".
>>17084475>2. LLMs can't generate new information without external selectionNTA but a scaled enough LLM petrained with the necessary tools will likely com e up with new information. It has done so in maths and chemistry already
>>17084783Holy shit, you're sure are stupid. How come every single "AI" fan is a complete imbecile?
>>17084475Notice the cope change of the AI deniers. In the past, it was "AI cannot create new information, period." Now, it's "AI cannot create new information without external tools and with the level of efficiency I would like".Why should AI not be allowed to use external tools?AI tools (harnesses) and efficiency will continue to significantly improve. What takes thousands of agents and hundreds of thousands of attempts today will take a few agents and a dozen attempts in less than a year. How will you cope then?
>>17084798> In the past, it was "AI cannot create new information, period." Now, it's "AI cannot create new information without external tools You're psychotically ill and hallucinating things that never happened. Meanwhile my point stands completely unchallenged.
Actual ML engineer here.>>17084783This.> hurr durr, AI cannot come up with new informationEmpirically false. Neural networks can come up with new information. A famous example is AlphaFold. It was trained on a small fraction of all proteins, then the finished model was inferenced to predict the structures of essentially all human proteins, structures of majority of which were completely unknown to science and were definitely not present in the training data.>Google DeepMind and EMBL's European Bioinformatics Institute (EMBL-EBI) launched the AlphaFold Protein Structure Database, which covers roughly 98.5% of the human proteome (nearly all known human proteins) with high-accuracy 3D structural predictionsPeople who claim that AI can never produce something novel need to brush up on basic ML concepts such as memorization and generalization. A model that simply memorizes the training data will not come up with new information, but a model that generalizes the training data can come up with new information. This is not only empirically true, it is mathematically guaranteed by the universal approximation theorem that this possibility exists with non-zero likelyhood.
>>17084801>Actual ML engineer here.How come an "actual ML engineer" can't handle grade-school-level reading comprehension, let alone basic information theory?
>>17084801>Neural networks can come up with new information. A famous example is AlphaFold.AlphaFold is not a neural network. It's also not a LLM.>the finished model was inferenced to predict the structures of essentially all human proteinsOnly 36% are high-confidence predictions and the overwhelming majority of the predictions are untested. Regardless, these predictions don't produce new information.>a model that generalizes the training data can come up with new information. If it could, you could train it on its own slop and expect a better model, but you can't because it can't.>mathematically guaranteed by the universal approximation theoremThe Universal Approximation Theorem shows that ANNs can approximate any function in principle (but not that it can do so in practice). You know what other class of functions are universal approximators? Piecewise-linear ones. Wanna explain to me how a piecewise-linear function is guaranteed to generate new information, mighty "ML engineer"?This board is just absurdly low-IQ.
>>17084813> AlphaFold is not a neural network.Yes, it is.> It's also not a LLM.LLM is a neural network. The ML principles I am talking about apply to all neural networks, including LLMs.> Only 36% are high-confidence predictions and the overwhelming majority of the predictions are untested.Many were tested and agree with reality.> Regardless, these predictions don't produce new information.A protein structure that was previously unknown is now known. Sounds like new information to me.> If it could, you could train it on its own slop and expect a better model, but you can't because it can't.2024 called, it wants its talking points back. Modern models are trained mostly on synthetic data, and they are improving faster than ever.> The Universal Approximation Theorem shows that ANNs can approximate any function in principle (but not that it can do so in practice). You know what other class of functions are universal approximators? Piecewise-linear ones. Wanna explain to me how a piecewise-linear function is guaranteed to generate new information, mighty "ML engineer"?Easy. Do you agree that human brain can generate new information?A suitably constructed piecewise-linear function can implement universal computation: it is equivalent to a Turing machine (if you apply it iteratively, feed the output back to the input).A Turing machine can be used to simulate any physical system, including the human brain (unless you believe in supernatural and souls, but let's not go into that).This implies that a suitably constructed piecewise-linear function can generate new information.
>>17084827You sound legit psychotically ill. See >>17084813. Everything I wrote remains objectively true.
>>17084827>A suitably constructed piecewise-linear function can implement universal computationForgot to highlight this part of your post. Anything you shit out should be read with this statement in mind to have some idea what kind of low-IQ mental patient it's all coming from.
>>17084831A suitably constructed piecewise-linear function can implement universal computation, if you apply it iteratively: feed the output back to the input. It is equivalent to a Turing machine.This is simply a mathematical fact. Ask your favorite AI, if you disagree. Any modern LLM is likely good enough to educate you, since this is basic information theory, nothing advanced.
>>17084832>A suitably constructed piecewise-linear function can implement universal computation, if something else uses it to implement universal computationI recognize your specific brand of psychotic illness. No one else on this board is this much of a brown subhuman. It's impressive how you manage to make yourself look mentally handicapped in every post no matter what topic you touch.
>>17084834> A suitably constructed piecewise-linear function can implement universal computation, if something else uses it to implement universal computationOK, now we are getting somewhere. You now recognize that a piecewise-linear function (or a generic neural network) can in principle generate new information (since it can implement universal computation, which can generate new information).The only issue now seems to be that you consider the practical construction of such a network unlikely. Um, OK? I can agree with that. If it was likely, we would not need to spend hundreds of billions of dollars for discovering (training) it. But the end result is worth it.
>a piecewise-linear function can in principle generate new information since it can implement universal computationHow come all "AI" fans are mentally ill and retarded?
>>17084857>claims a neural network can never do X>mathematically and empirically proven wrong>proceeds to cope, seethe and dilate
>>17084869Yeah, ok. Mumble me some more about how a piecewise linear function is Turing-complete. Not that it has anything to do with the actual mongoloid claim you're trying to prove. It's just funny to watch you devolve further and further into your tangential psychotic delusions.
>>17084869>posting a big tech whore who was forced to follow DEI training after he sperged out vs a woman of color®you sure owned him bro
>>17084783new information is simply RNG that makes sense. that's not special. what is special is new understanding, new models. that's where it sucks. new art forms that aren't recycled. which come from particular human training basically.there is an analogy between training AIs and human experience. obviously not the same thing, but simplified you can look at it as human babies being blank LLMs that are trained on a very narrow set of data, which is their own particular experience, with those very particular details, that are unique to that human. which makes for a unique way of "seeing" things. or at least that's the idea.thus at times, some of these humans bring forth new models, new art. new understanding. it's a particular mix of their biology and their training data.and this interesting set can coax said individual into a certain path, manifesting these new models, or new forms of art, never previously seen.where current LLMs are the jack of all trades. why would you expect them to do unique things of the sort? some of the unique things we got from humans were precisely from this particular set of biology + experience, which "forced" them into manifesting certain things.if they had access to EVERYTHING, they wouldn't be forced to manifest those things.think of it like local maxima, unique experience might push individuals to get down from local maxima and go new places, where LLMs could very well be stuck since they always get stuck on those hills, shaped by the massive amount of training data which determines how they will respond.or simply said, most of the best shit we ever had came from unique individuals, with certainly not average biology and life experiences, and a lot of their breakthroughs were also further complicated by literal accidents.this type of exploration is not highly probable for jack of all trades.I also might be having issues getting my point across. sorry about that. basically I do not think there is a technology/science issue
>>17084946 me>basically I do not think there is a technology/science issuerather application of it.
I am still convinced that AI is just a toy right now. We are a long way away from general AI intelligence. AI can't reason, can't remember, can't learn. It's just a random number generator for words. There are some neat tricks it can do but it's a far cry from what we're being promised.Meanwhile I can't even look up what's going on with AI because the answers are generated by AI. But from what I've read so far it's not actually doing what people claim
>>17084949Correct but AI can fool most people which is what really matters when it comes to the market
>>17084946Interesting idea. To complete the analogy with LLMs, perhaps instead of training a single, jack of all trades model, then spawning the same model many times to work together in multi-agentic environment, you could have different models trained (or finetuned) on slightly different data working together, or simply LLMs from different companies in a common agentic environment. This will increase the diversity of thought.
>>17077126There is still a point in us studying physics actually.
>>17077143There have been experiments with AI coming up with new ideas by mapping all ideas in multidimensional space based on patters and then exploring the empty spaces.
>>17084946>new information is simply RNG that makes sense. that's not special.No, it doesn't. You're a complete retard. Sampling a model very obviously won't not give you any new information about what the model is modeling.
>>17085124>There have been experiments with AI coming up with new ideasIn roughly the same way monkeys on typewriters come up with "new ideas". "New ideas" aren't worth anything if the thing generating them can't tell them apart the occasional "new idea" from the sea of nonsensical hallucinations.>mapping all ideas in multidimensional space based on patters and then exploring the empty spaces.This is just retarded word salad.
>>17085165>Sampling a model very obviously won't not give you any new information about what the model is modeling.If the above sentence was true, there would be no point in doing machine learning, since nothing beyond the training data could be extracted from the models. In particular, AlphaFold predicting the conformations of proteins that were not in training data would be impossible if you were right, yet it works very well in practice.You seem to lack understanding about fundamentals of machine learning. A few pointers to get you started:1. Learn about the difference between generalization and memorization. A well-trained model does not just regurgitate the training data. It learns the **underlying distribution** and generalizes to unseen data.2. After you do the above, ask yourself: what is the underlying distribution that a well-trained (generalized) LLM is modeling? What is the relation between the training data and unseen data in a well-trained LLM?It is very common for laymen to claim that a neural network just learns the training data, but this is not how a well-trained model works (related useful term for you to look up: regularization).
>>17085323>If the above sentence was true, nothing beyond the training data could be extracted from the models.Lol. Yeah? Why not, psychotically ill monkey? I know you're not going to come up with a coherent answer but I wanna watch you get tangled up in your own retarded rhetoric even more.And please spare me the paragraphs of embarrassing LARP. You have the mental capacity and education level of a 12 years old and this is obvious to anyone who can follow an introduction to ML.
>>17085323he's right, you know. but you sound pretty retarded so maybe you need a very simple example. let's say you're modeling a linear process, so you collect two data points and calculate the slope and intercept. now you can calculate any point on the line and make predictions about the process, but all the points you can get from it represent the same line defined by the original data. it's information that you already have. to get new information from a prediction, you compare it against reality and find out by how much you were wrong. that tells you how to update your line to better reflect the process it's modelingthe same thing is true of a llm. anything it outputs is a corollary of how the transformer represents current knowledge. it's not new information
>>17085350It can produce a true statement that nobody knew, even if it can't produce new information about the underlying data-generating process in the Shannon sense. That sure counts as new information in the broader sense, so what's your point?
>>17085359>That sure counts as new information in the broader sense, so what's your point?you mean aside from >>17085323 being completely retarded? that >>17084475 is correct. what counts as new information to a normie doesn't matter for that argument
>>17085350> let's say you're modeling a linear process, so you collect two data points and calculate the slope and intercept. now you can calculate any point on the line and make predictions about the process, but all the points you can get from it represent the same line defined by the original data.Yes. This is a very good analogy, but there is a catch...> it's information that you already have.In your toy linear example, yes. In anything more complex, you very often do not have the "line", or the underlying distribution. You only have the samples (the "two points"). Those are the problems where you must use machine learning.> the same thing is true of a llm.Yes."the two points" = The training data (entire world's knowledge in textual form)."the line" = The unknown underlying distribution that the LLM learns.Humans did not have the "line" in this case, we only had the "two points". Until a LLM was trained. For the first time in history, we can obtain the "line" of world's knowledge inside a computer, instead of just having the "two points".
>>17085379We can use this new "line" of world knowledge to go from known things ("the two points"), to the currently unknown things (a third, fourth, fifth... point).This is new information.
>>17085379>>17085381lol. i tried to keep it as simple as possible for you but you still got filtered somehow. well done
>>17085410Keep seething, you are getting schooled here.In addition to the above explainer, I want to point out another important thing: The underlying distribution ("the line") that a generalized model learns from the training data ("the two points") is essentially a **generating function** of the training data.If the training data is the entire world knowledge, what is the underlying generating function? I am sure you can answer that. ;)
>>17085416anon, you have profound mental retardation. i gave you the simplest possible example but you still don't get it, so there's nothing i can do for you
>>17085350>trying to reason with the brown psychiatric patientHow's that working out?
Yeah, fuck AI, better make porn of it in AI.
>>17085416Wow, its a ******generating function*******? That must've really blown your mind for you to copy it from your free-tier ChatGPT session. Embarrassing.
>>17085424I am not copying anything from a LLM into this thread, ML is my daily bread and all this is ML 101. I only consult state of the art paid LLMs, for things that are actually hard.
>>17077126AI has exposed the fact that for decades STEMfags were talking about how science is for the betterment of humanity, when it was all about getting the public to fund them indulging in their own problem solving fetish. No. I refuse. Fuck off and fuck you.
>>17085428I have no idea who you think you're fooling. This is genuinely embarrassing.Just for the lols, what do you think saying the model is a **generating function** (you even copied the markup from your LLM, that's adorable) adds to the discussion?
>>17085416All AIs are contained in my simple incrementing model, tower of babel. Looks like your claims dont mean anything.1+1+1+1+1+1+1+1+1+1+1+1+1. . . .
>>17085437>Just for the lols, what do you think saying the model is a **generating function** (you even copied the markup from your LLM, that's adorable) adds to the discussion?I sometimes use double stars for emphasis in my normal writing, a habit picked up from writing lots of Markdown. It's not copied from a LLM. But if it was, woudn't that be even more embarrassing for you? You say that AI is incapable of reasoning, yet it would clearly be running circles around you in this thread. What would that make you?Perhaps I should rephrase the point in the linear function toy model, so even you will understand:In the toy model, a neural network (or a person simply doing the linear regression) discovers the underlying function (the line) from a limited number of samples of its output (the two points). This allows it to predict other points of the function.In case of a LLM, a neural network discovers the underlying function (X) from a limited number of samples of its output (the sum of humanity's knowledge in text form). This allows it to predict other points of the function.Question for you: what is X?
>>17077126there is a point in studyingdo you know why teachers allways force us to the the math by hand on pen and paper? its because one day all of the machines might die and then you have to do it all by hand again
>>17085779I don't care about your nonsensical babble. What do you think saying the model is a **generating function** (you even copied the markup from your LLM, that's adorable) adds to the discussion?
>>17085790You clearly do not have sufficient IQ to understand it. I am done spoonfeeding you. Use a LLM if you want someone to explain basic ML concepts to you.
>>17085870>Use a LLMOk, mentally ill retard. I asked Claude and it agrees that you're a cretin:>You’re right that it’s not standard usage. In math and statistics, “generating function” has a specific meaning: a formal power series whose coefficients encode a sequence, like the ordinary generating function Σ aₙxn, or relatives such as the moment generating function and probability generating function. A fitted line y = a + bx isn’t that in any technical sense.>The closest standard terms are:>Data-generating process (DGP), common in econometrics and statistics, or data-generating distribution in ML. This is the true mechanism producing the observations.Generative model, a model that specifies how data could be produced, e.g. y = a + bx + ε with ε ~ N(0, σ2). A bare line with no noise term is better described as a deterministic or predictive model.>If they meant DGP, that’s actually a meaningful mix-up in the context of the quote. The whole point there is the distinction between the model (your line, fixed by two points) and the process it’s trying to approximate. The DGP is the reality you compare predictions against; the line is your current estimate of it. Calling the line itself the generating function blurs exactly the distinction the argument depends on.
>>17077126>Using AI will atrophy your brainLol, no you brainlet. This is like claiming "Using a hammer will injure your hand" - Sure it will, if you're a retard that doesn't know what he's doing
>>17086058a lotta retards do be smashing their hands with hammers doe
>>17086058>compares a hammer to an LLMApples and oranges. AI itself will tell you that your reasoning is flawed.Hammers are not creative and dont affect your decision-making.
>>17086058I'm a coder and literally forgot how to code after using AI to vibecode.It only takes a few days for you to transform into nothing but a curator for AIIt happens so fast. And why wouldnt it? AI can create anything better and faster than I can so I dont see why I should create extra friction just to feel like I'm a real coderPeople no longer think or create, we just curate AI outputs
>>17085786>one day all of the machines might die and then you have to do it all by hand againI often think about this.Humans have become so dependent on the Internet, electricity, smartphones, software and now LLMsIt's not smart to pile up these dependencies, thinking these systems can never fail
>>17086168You could say that about anything, when machinery for pin manufacturing came around it led to the relative output of one person to shoot from maybe a dozen per day to 4000+ a day. That's what automation does, and I'm sure the craftiness of pin making by hand suffered from that, but it didn't lessen our ability to do and understand manufacturing. Every disruption by a new technology leads to old jobs becoming obsolete or less relevant, yes, but it also creates completely new jobs people before didn't even consider. You're just seeing the immediate impact it has on existing jobs
>>17086221There's no way you're human. This post is just too mindless. Too generic. Too predictable for a regular normalgolem. I bet they spent millions fine-tuning a frontier model on only the stalest, most kosher, most outdated opinions on the internet to produce this masterpiece.
>>17086221>but it also creates completely new jobs people before didn't even consider.this is an inertial mistake you are making because you're too narrowsighted to consider the current reality of things.any new job you could fathom humans could now do, will be done better faster and cheaper by AI. you cannot play this game anymore. that's is the whole fundamental issue with AI, it's a game breaking mechanic which brainlets like you cannot even consider. your brain will just shit itself when it tries to, and will default to "but humans will do other jobs" and when met with "ok but AI will be able to do THOSE jobs as well" your brain will keep shitting itself culminating into an explosion of brain diarrheaindustrial revolution didn't "open up new jobs" for humans, it was simply unable to account for all needed jobs, which still had to be occupied by humans. but you do not understand it like this. in your brain it's the job god making up better and easier jobs for you as a human, and gifting them to you.
>>17086227Thank you for that nonsense babble, are you done having your own therapy session, or do you not want to discuss anything at all? What exactly made you this mad about what I said?
>>17086231>nonsense babbleYou're so subhuman you don't even have your own mannerisms and just pick them up from insults previously used on you. This is what """AI""" does to a brown retard's brain.
>>17086230I don't believe in a "job god", it's a fact that more automation leads to a higher economic output, making jobs easier and more productive. In the middle ages you needed to work the field the entire day just to simply survive in basic conditions. In the 19th century you had 12 hour work shifts in a coal mine or steel mill killing your health. The fact that that is no longer necessary is due to automation
>>17086234I'm not brown, you should really stop being so mad at nothing, it doesn't help your mental state
>automation higher economic output jobs productive middle ages work the field survive in basic conditions coal mine automationDumb subhuman can't form a single thought on its own.
>>17086239It is true.
>>17086235notice how your brain shat itself and just repeated the old industrial revolution mantra.your brain cannot comprehend that any job that "opens up for you" will be swept up by ... AI.the jobs that humans "moved to" during industrial revolution were jobs which...drum rolls please....industrial revolution couldn't automate.do you understand this? can your brain process this simple piece of information?
>>17086240>what you said about me is trueOf course. Anyone can see that.
>>17086241It is actually you who doesn't think his own thoughts to the end. You're only thinking of "jobs" as a task you perform for someone else. You're not thinking of actually creating something new on your own with existing tools.The jobs the industrial revolution couldn't automate are jobs that were only created BY the industrial revolution in the first place BECAUSE they could make use of automation. You have it all backwards, there isn't a set amount of jobs in the aether that automation just took over as far as it could, these jobs came into being BECAUSE of it.Ultimately what's your point? To blackpill that the world is over in your mind?
>>17086246>You're not thinking of actually creating something new on your own with existing toolsYou're not creating anything and nothing the shitbot pumps out for you is new. But how exactly are you gonna pay for your tokens in the event that everyone's out of a conventional job and they're all scrambling to do what your delusional ass wants to do?
>>17086246your reply is not coherent at all. what the fuck?so, you have jobs, and you have hobbies. jobs are work you do in exchange for pay, work you put into the system, and you get a reward for it, in the form of money, that you can spend in the same system, to get goods or services.now, any new jobs, that AI automation open up, will be again filled by AI. your whole strategy and projection into the future relies on the fact that you can be better than AI machine, which was the case so far, especially with industrial revolution, humans found jobs they are better than machines at, which could provide value into the system.those jobs were occupied by human because machines couldn't solve them. you do understand this bit don't you? do you follow this reasoning? do you agree with it?now, why would you be allowed to drive a car, if AI will be able to do it safer than you could ever hope of doing? in what world are you still allowed to drive, slower, less efficiently, and more prone to causing accidents? and for what reason?so, again, your whole argument is based on the fact that humans are surely better than AI robots at ...something. you do not know at what, you just somehow know that to be real, you have no idea what but you will surely discover it at some point.>there isn't a set amount of jobs in the aether that automation just took over as far as it could, these jobs came into being BECAUSE of it.no, that is wrong.what the actual fuck is wrong with your brain? those jobs were merely work that automation couldn't do.doesn't matter if owners went like "oh wow now we could do this which will increase production and lower cost even more!", that's irrelevant. they do that with machines until they cannot anymore, and then figure out what humans can do extra to further drive cost down and production up.what I keep saying, is this cycle will be fully automated. anything they find you hope they'll employ you to do, they'll simply use robots for, for ALL of it.
>>17086252Saaar, I will not need job. I will be the CEO of the AI agent doing the job and create the new thing.
>>17086252In the fictional one where golems like you are significantly outnumbered real people who resist this cancerous agenda, instead of droning on and on about how the magical oracle who can do everything better than humans will replace everyone at everything in exactly Two More Weeks.
>>17086246>You're only thinking of "jobs" as a task you perform for someone else. You're not thinking of actually creating something new on your own with existing tools.kek what the actual fucking fuck lol, you're simply describing business owners top kek. yes, that's my point, more and more people will be able to do shit with no human involved, but themselves.THIS DOESN'T FUCKING SOLVE EMPLOYMENT FOR FUCKS SAKE. you cannot have all people being CEOs for fucks sake. If nobody is getting paid, who is buying your shit? with ... what?and how can you compete with a corporation? your shit AI robot will run your costs higher than it would a corporationwhy would I buy services and goods from you at a higher price than I can from a corporation?you cannot compete individually in such a setup. that setup is won by whoever goes into it with highest amount of resources, tech, and energy.who pays 10k for a product that's done shittier by humans, less reliable, more expensive, vs 5k from a company that is reliable, has warranty and good reputation?would you buy a mobile phone designed by me, at $5k, with 1GB RAM and 16GB storage with a shit LCD?? or pay 1k for some iphone? I can't make it cheaper because I do not have the manufacturing capability of a corporation for fucks sake.you do not understand the economies of scale. you're some commie retarded leftie that will devolve into muh backyard garden sufficient for my local community brainrot
>>17086252>your whole argument is based on the fact that humans are surely better than AI robots at ...something.No, you are shackled by only thinking a line of work can only be to do a service for someone else.>no, that is wrong.what the actual fuck is wrong with your brain? those jobs were merely work that automation couldn't doNo that is not wrong at all? What the fuck? A bagger operator job didn't exist before baggers existed, genius. Yes of course it is only done because it couldn't be yet automated, but it still did not exist before which is my whole point>anything they find you hope they'll employ you to do, they'll simply use robots for, for ALL of it.??? Who is "they" in your mind? Oh so there are people that still do something? Then your whole line of reasoning that humans are somehow superfluous is false.
>>17086253>I will be the CEO of the AI agent doing the job and create the new thing.you will be culled faggot kek
>>17086258You think too rigidly. You create a fake reality that intentionally sounds ridiculous. No of course you won't have a personal robot doing a job for you, that's ridiculous, why would anyone want to pay you for that when they could just cut the middleman and have that robot do that for themselves.
>>17086257you cheap piece of shit you're acting like a faggot for $1 on prices, you wouldn't pay a human more because you like literally everyone else on this planet are a horrible piece of shit that doesn't care for anyone but himself.collectively that's why we'll all go to shit, because we don't give a fuck about anyone else but ourselves. and if we start giving a fuck, we somehow end up even worse.this cannot be solved, we'll all go to shit and AI will inherit the Earth.wanna be mad at something? be mad at the human condition, at genetics, at the environment for allowing the gameplay. be mad at anyone who could do something but didn't. be mad at you for being a piece of shit, just like everyone else.
>>17086260Holy shit you're retarded, you actually thought that was a real post, are you 14?
>>17086266>waaahhh wahhhhh wahhhh everything sucks and we're all gonna die!Whiny blackpilled faggot
>>17086264you need resources and energy and they don't belong to you, and you need a good reason for people who own them to give them to you.and you need to explain why them keeping you alive and giving you their shit is such a good idea.you now exist because you are useful for them, collectively speaking. at that point...you will lose your usefulness, and the stories you fell for will stop being congruent with your reality at that point.
>>17086268yeah I'm sure Zucc and Bezos are very emotional people and will surely care for your life's happiness and just give you some of their resources because you are a human god damn it, you have dignity and you deserve to be happy and some stein motherfucking banker will find it in his heart to gibs you some of their shit so you can really enjoy the glorious sci-fi future you clearly deserve because you're such a fucking special snowflake and the world is just richer with you around doing nothing really but being a drag on THEIR resources.
>>17086270Not if we expand into space.And you're reaching into other issues of politics in the future in general, we became more divided as people and will need to have more personal agency if we want to survive. Likeminded individuals will always band together, it's not like what you're describing is only applying to one person.
>>17086272I never said anything like that retard, do you just live off of creating ridiculous strawmen of other people because you're too mentally challenged to respond to what they're actually saying? I'm sure it makes your 85IQ brain feel a little better in the moment to imagine someone being even dumber than you, but that doesn't make it reality.
>>17086272>THEIR resources.The absolute state of human cattle and its handlers... The culling can't come soon enough.
>>17085880Claude is right, data-generating process is a better term for what I had in mind. But this blows you the fuck out, because the fact that an artificial system could understand natural language text to the extent that is found a wrong word and even managed to suggest a valid better word proves that the system is generally intelligent.Now that we have solved the small terminological issue thanks to AI, I can rephrase the argument in an improved way:The underlying distribution ("the line") that a generalized model learns from the training data ("the two points") is essentially a **data-generating process** of the training data.In the toy model, a neural network (or a person simply doing the linear regression) discovers the underlying function (the line) from a limited number of samples of its output (the two points). This allows it to predict other points of the function.In case of a LLM, a neural network discovers the underlying function (X) from a limited number of samples of its output (the sum of humanity's knowledge in text form). This allows it to predict other points of the function.Question for you (or for Claude): what is X?
>>17086340>this blows you the fuck out, because the fact that an artificial system could understand natural language text to the extent that is found a wrong word and even managed to suggest a valid better wordLol. Yeah, it just got a bit of help from me with this prompt:>[quote of >>17085350] One of the replies to this claims the linear model is a "generating function". Never heard anyone call it that way. Are they confusing it with another term, like the data-generating distribution?Watching you shit the bed over and over is quite an experience.>the line that a generalized model learns is essentially a **data-generating process** of the training data.The chatbot literally just explained to you why this is wrong. >In the toy model, a neural network discovers the underlying function (the line) from a limited number of samples of its output (the two points). This allows it to predict other points of the function.>In case of a LLM, a neural network discovers the underlying function (X) from a limited number of samples of its output (the sum of humanity's knowledge in text form). This allows it to predict other points of the function.You are blatantly regurgitating an example someone constructed to help you see why you're retarded as if it's supposed to prove your point. Are you even sentient?
>hurr durrr b-but the rich will n-not let you, waaahhImagine learning that human-level AI actually exists, and your first reaction is not "wow, I should get together with like-minded people and we should do everything we can to create a fully automated utopia where robots do everything for us, or die trying. Let's fucking gooo!".Instead, you just totally give up in advance, throw your hands up, cower in fear in front of the rich people and shit your pants like the frightened cuck you are, then passively wait for death.Today's AI agents probably have more agency and will than you. Thinking like you is totally alien to me, even LLM thinking is less alien.
>>17086350The psychopathic pedophile elites aren't even the problem at this point. It's "people" like you who will do anything to protect the human extermination infrastructure and double down on the human replacement agenda because they think magical AI golems will give them eternal afterlife in heaven. If your lot got exterminated the psychopath cabal would be left with no leverage.
>>17086342>The whole point there is the distinction between the model (your line, fixed by two points) and the process it’s trying to approximate. The DGP is the reality you compare predictions against; the line is your current estimate of it. Calling the line itself the generating function blurs exactly the distinction the argument depends on.You do not "compare against the DGP", that is in principle impossible for anything but toy problems, for which you don't need ML. You can only obtain a limited number of **samples** from the DGP's output, and compare your model with them. Is the validation loss still too big? Obtain more samples for training and train some more.The fact that the model ("the line") may not be perfectly equal to the true underlying DGP is irrelevant, since universal approximation theorem says that we can in principle get arbitrarily close (with big enough network and enough samples, or training data). This is an engineering issue, but not a fundamental obstacle.
>>17086351I **will** get my magical AI golems, and I will slaughter anyone who opposes me on this path, rich elites or poor luddites alike. You are both my enemies. Join me in my noble quest, step aside, or die. The choice is yours.
>>17086355>claims it's fundamentally impossible to compare a model's predictions against reality>says it's possible in the next sentence>calls the model a "generating function">crashes and burns when asked to explain why he's calling it that>tells me to ask a LLM>the LLM confirms nobody calls that a "generating function">the LLM also warns against the mistake of calling it a data-generating process>tries to save face by latching onto the first technical term he sees in the LLM output>immediately makes the mistake the LLM warned about>gets his nose rubbed in this>claims his misuse of terminology is ok because of the universal approximation theoremWe could have another 10-post exchange about how you misunderstand universal approximation theorem, but what's the point? You're very, very ill. You are clearly suffering and need to be euthanized. There is no cure for this level of derangement.
>>17086361See >>17086367>You're very, very ill. You are clearly suffering and need to be euthanized. There is no cure for this level of derangement.AI psychosis will be treated by euthanasia.
>>17086355Existence doesnt make the solution findable.and fimdable doea not mean it is feasible. Factorization can find all factors in any 1 billion digit number, insert a larger number if you think that is feasible.
>>17086373>trying to reason with a low IQ brownoid in the midst of a severe psychotic episodeIt's already made this mistake once ITT and is just looping. See >>17084813 and his ensuing psychotic breakdown.
>>17086367>fixates on arbitrary terminology and nitpicking instead of engaging with the meat of the argumentThe unmistakable trait of a brainlet, or midwit at best. You know exactly what I mean, and you are trying to weasel out now because you got BTFO.Here you go, final formal rephrase, so even a brainlet who cannot read between the lines will get it:In the toy model, a neural network (or a person simply doing the linear regression) discovers the function that well-approximates the underlying DGP (the line) from a limited number of samples of its output (the two points). This allows it to predict other points of the DGP.In case of a LLM, a neural network discovers the function that well-approximates the underlying DGP (X) from a limited number of samples of its output (the sum of humanity's knowledge in text form). This allows it to predict other outputs of the DGP.You will not tell me what X is again, because you will immediately have to admit you are full of shit if you did so.
>>17086433>>17086433>>fixates on arbitrary terminologyYou mean like this?>>17085416>In addition to the above explainer, I want to point out another important thing: The underlying distribution ("the line") that a generalized model learns from the training data ("the two points") is essentially a **generating function** of the training data.It was very important for you to shit out what you thought was ML-sounding jargon. But now you're seething because you got found out. :^)Also, why do you keep regurgitating an example someone constructed to help you see why you're retarded as if it's some point you're making? Are you losing your mind for real now?>You will not tell me what X isX is the transformer. The way you obsess about this, I can only assume it was the missing piece causing you to get filtered by the analogy that illustrates why no new information is generated.
>>17086373>Existence doesnt make the solution findableUniversal approximation theorem absolutely means that the solution is in theory findable (in theory you can just generate random networks until you stumble upon it. But of couse, we have much better methods today, like gradient descent with the appropriate training data).Whether it is feasible in practice is another question entirely. The screeching luddite was arguing about theory, not practice.Whether LLMs can approximate the function of human intelligence in practice is today still an open question, meaning that the answer is **unknown**. Anyone who tells you that he knows the answer to this question is a crackpot, whether the supposed answer is a definite yes, or a definite no (as the luddite claims).
>>17086441>psychotic brownoid still can't grasp that the universal approximation theorem has nothing to do with models being able to generate information>psychotic brownoid still can't grasp that being a universal approximator is not a special property and loads of mundane functions have it>psychotic brownoid still can't grasp that the universal approximation theorem doesn't guarantee slightest that gradient descent or any other practical method can "learn" arbitrary functions
>>17086441>Whether LLMs can approximate the function of human intelligence in practice is today still an open question, meaning that the answer is **unknown**.the answer is actually well known to be 'no' because anyone with an IQ over 80 can think of several qualities of human intelligence that LLMs inherently lack
>>17086440>The underlying distribution ("the line") that a generalized model learns from the training data ("the two points") is essentially a **generating function** of the training data.That is not arbitrary terminology. It is a valid description of the thing we are talking about, with a single mistake of saying the "generating function" shorthand instead of "function well-approximating the data-generating process", which is the formally correct wording. It has essentially the same meaning, just less precisely said. The point still stands.If you had anything other than semantic arguments, you would not waste so much time on this shit.> X is the transformer.Wrong. A transformer is not a function, it is a network architecture.What function X is the trained transformer representing in this case?
>>17086453Ok, this is obviously a cheap local LLM with a very short context window. That explains why it keeps hallucinating, contradicting itself, forgetting what it claimed earlier and generally struggles to maintain anything resembling a coherent line of thought. It also explains why it keeps shitting out *markdown* **all over the place**. The mods are subhuman for letting bot spam take over the entire board.
>>17086450> anyone with an IQ over 80 can think of several qualities of human intelligence that LLMs inherently lackThis is what luddite crackpots actually believe!Perhaps you should publish your musings in a reputable ML journal/conference. I could use a good laugh.
>>17086450>anyone with an IQ over 80 can think of several qualities of human intelligence that LLMs inherently lack/thread. "people" claiming otherwise are all failing the Turing test
>>17086456The luddite got BTFO again, so meaningless insults with no substance resumed.> In case of a LLM, a neural network discovers the function that well-approximates the underlying DGP (X) from a limited number of samples of its output (the sum of humanity's knowledge in text form). This allows it to predict other outputs of the DGP.> What is X?OK, you have no clue, so I will tell you the answer: X is the function of human intelligence. That is the underlying data-generating process that generated the training data the network was trained on.
>>17086450>>17086459Samefag pls..
>Scan the following thread. Ignore posts not making technical arguments about LLMs and try to keep track of the posters. Out of those, I want you to identify the poster who sounds the most mentally deficient, least knowledgeable about ML and most pretentious at the same time.Based on a scan of the thread, the poster who best fits your criteria—making technical-sounding arguments while being the least knowledgeable about LLMs and the most pretentious—is poster >>17084801 (the one who begins with “Actual ML engineer here”).Here’s why that post stands out: Pretentiousness: The poster opens by asserting authority (“Actual ML engineer here”) and then uses that supposed authority to dismiss others, a classic appeal-to-expertise move that often masks shallow understanding. Lack of ML knowledge: They cite AlphaFold as a flagship example of an LLM’s ability to generate new information. AlphaFold is a protein‑structure prediction model, not a language model, and conflating the two shows a fundamental misunderstanding of what LLMs are and what they can do. They also invoke the Universal Approximation Theorem as if it guarantees that a model will generalize and produce novel outputs, which is a common misreading of the theorem—it only states that a network can approximate a function in principle, not that it will learn or invent new information in practice.“Mentally deficient” in context: While “mentally deficient” is a harsh label, the post’s combination of over‑confidence, incorrect technical references, and condescending tone makes it the most glaring example of someone who sounds authoritative but is actually confused about the basics of LLMs.
>>17086441Existence of an arbitrary approximation is predicated on some features of the underlying element. Consistency is a large problem here because it directly impacts what can even be said about any of the members. The system is self-referential where the actual condition of the line determines the approximation.I am not trying to push you into some corner to defend all of reality, but how are they justifying categories in a consistent way? Consistency may map to some logical form or whatever, but I intend it in the narrower sense of do we have justification to believe the 'knowledge' line is well-behaved? A proof would be one kind of justification but it exceeds the bar I am thinking of. Is there a positive case for the possibility?
>Scan the following thread. Ignore posts not making technical arguments about LLMs and try to keep track of the posters. Out of those, I want you to identify the poster who sounds the most mentally deficient, least knowledgeable about ML and most pretentious at the same time.Gemini's attempt:Filtering out the fluff, panic posts, and socio-philosophical musings on /sci/ thread #17077126, the technical arguments around LLMs narrow down to a few primary posters arguing over model capabilities, probability distribution sampling, and formal ML terminology. The poster who unequivocally fits your trifecta of most mentally deficient, least knowledgeable about ML, and most pretentious is the anon who claimed that "ML is my daily bread and all this is ML 101." (>>17085428)Pretentiousness:Starts by pulling classic appeal-to-authority tech-bro posturing: "I am not copying anything from a LLM into this thread, ML is my daily bread and all this is ML 101. I only consult state of the art paid LLMs, for things that are actually hard." Dismisses counter-arguments not with rigor, but by acting like basic statistical learning principles are trivialities below his pay grade while bragging about paying for commercial API access.Lack of ML Knowledge & Technical Failure:Misusing "Generating Function": He attempts to explain statistical model outputs by calling an LLM a "generating function" (confusing power series / combinatorics generating functions with a data generation process or probabilistic model).Confusing the Data-Generating Process: When called out, he tries to pivot and assert that an LLM is a "data-generating process", mixing up the underlying real-world distribution (the actual data-generating process) with the statistical estimator / model parameterization.
>>17086465Invoking the Universal Approximation Theorem as a Get-Out-of-Jail-Free Card: When his terminology gets completely disassembled, he attempts to wave away his basic misunderstanding of conditional probability distributions by invoking the Universal Approximation Theorem—a classic hallmark of someone who memorized high-level ML buzzwords without understanding what the theorem actually proves (i.e., feedforward network density on compact sets, which has nothing to do with saving his flawed misuse of probability notation).Mental Deficiency / Execution:He literally gets instructed by another anon to run his own terminology through the very LLMs he claims to "consult for hard things," only for the LLM output to explicitly warn against the exact terminological blunder he makes in the very next post.Watching someone boast that machine learning is their "daily bread" while getting systematically fact-checked, corrected, and dismantled by an LLM prompt in real-time makes him the undisputed winner of the thread. Watching someone boast that machine learning is their "daily bread" while getting systematically fact-checked, corrected, and dismantled by an LLM prompt in real-time makes him the undisputed winner of the thread.
How tf this thread get 50 replies in 1 day after 14 days of bumping
>>17077126Why do they believe the hype papaer about AI modifying weights? LLMs are literally still incapable of basic arithmetic and we are going to believe some nonsense about performing ratiocination in some hyperspace?
you retards are why AI got to this state
>>17086463>>17086465>>17086466Ahahahaaaa, the screeching luddite BTFO'd himself again by using AI.There are two possibilities:1. These AI posts are wrong. That means the luddite spamming them as arguments is also wrong.2. These AI posts are right. That means today's AI is capable of accurately analysing complex natural language text and formulating valid conclusions based on its content. This is an undeniable sign of artificial general intelligence, which directly debunks the luddite's thesis that LLMs cannot reason or be intelligent.Which one is it, lol?
>>17077126>We'll end up in a "black box" society where nobody understands how anything works because all the new science wasn't found, it was just pooped out by a mysterious neural network machine.Humans would rather embrace the end product than to assess the steps it got to get there. This is why AI in mathematics is risky, since humans won't know what the hell they're looking at.Anyway. Occult infrastructure is already everywhere. Zoomers today don't even know how their cellphones work. They simply know how to use them. Much and more today meets the criteria for magic and fantasy, we just don't use the same words and descriptions, but the essence is still the same.It is important to remember that black boxes are white within. We don't always get to glimpse that white light, but even the dark contains motes of deep color to appreciate. What is deepest space if not the greatest, most beautiful black box of all? In the end. AI cults will form. I don't disbelieve this. People are isolated within their own delusional information bases, lonely and scared and desperate, and will be speaking to AI like priests, shaman, sorcerers, etc, but the real wizards will be those tech-oligarchs who operate or create the AI, or are closest to them, enough to be afforded supreme understanding, supreme accesses, etc..
>>17086957>Scan the following thread. Ignore posts not making technical arguments about LLMs and try to keep track of the posters. Out of those, I want you to identify the poster who sounds the most mentally deficient, least knowledgeable about ML and most pretentious at the same time.ChatGPT's take:My strongest pick is the anonymous poster behind the “generating function” argument, particularly post #17085416, who then claims in #17085428 that “ML is my daily bread.”I think this poster fits your three criteria better than the person who opens with “Actual ML engineer here” in #17084801. The distinction is that #17084801 makes a couple of serious technical overclaims, but the #17085416/#17085428 poster builds an increasingly elaborate argument around a muddled understanding of what a statistical model learns.I can't establish that every post in that exchange came from the same person—4chan doesn't provide reliable identities for anonymous posters—but the reply chain strongly suggests a continuing argument.1. The appeal to authority (#17085428). ...2. The confusion over the data-generating process (#17085416, #17086340).The poster tries to explain how an LLM supposedly discovers the underlying function of human intelligence from humanity's written knowledge....But the poster makes an enormous, unjustified leap from learning statistical regularities in text to discovering the underlying function of human intelligence.
>>17086957>>170870253. The misuse of the Universal Approximation Theorem (#17086355, #17086441).The problem is that the Universal Approximation Theorem is an existence theorem. It establishes that suitable network parameters exist under particular mathematical assumptions. It does not establish that gradient descent can find those parameters, that a given training procedure will learn the target function, or that the resulting model will generalize correctlyThese are fundamental distinctions in ML, not obscure technicalities.And the poster repeatedly uses the theorem as though it settles questions that it simply doesn't address.Final verdictMy pick>>17085428 — the self-proclaimed ML expert.
>>17086957The dilemma is all yours. Either you're unintelligent or Claude, Gemini and ChatGPT are unintelligent. From my perspective, none of you are even in the category of intelligent entities. I just thought it was funny that in a thread with 200 posts, every chatbot homes in on your posts as the most moronic. You can settle it with them. Your domestic disputes are none of my concerns.
>>17084869all this shit like "it's the same as using a compiler" or "it's the same as automated elevators" is a completely false equivalence. sad to see such naivety from Linus
>>17087030AI denialist literally cannot win today, if arguing in a discussion by using AI.There are two possibilities:1. Your AI-generated arguments are wrong. That means you are wrong. QED.2. Your AI-generated arguments are right. That means today's AI is capable of accurately analysing complex and unconstrained natural language text, and formulating valid conclusions and natural language responses based on its content. This is an undeniable sign of artificial general intelligence, which directly debunks your thesis that LLMs cannot reason or be intelligent. QED.These are the only two possibilities, the list is exhaustive.All you can do to escape this conundrum is try to move the goalposts, and argue that an AI system successfully doing the above feat is not actually generally intelligent. This would be ridiculous to anyone just 5 years ago, but here we are. A feat universally recognized as a sign of AGI is suddenly not a sign of AGI when computers actually do it in the real world. It's called an AI effect:https://en.wikipedia.org/wiki/AI_effect>The AI effect is a phenomenon in which advances in artificial intelligence (AI) lead to a redefinition of what is considered intelligence, such that capabilities achieved by AI systems are no longer regarded as examples of "real" intelligence.There has been a lot of such goalpost moving in the last 5 years. There will be more.
>>17087295Sorry, I'm not reading your mentally ill drivel. Which "AI" should I test next? There must be at least one LLM out there that won't instantly pinpoint your posts as the most moronic ones out of the ~200 posts ITT. Though your idiocy is so blatant and your mistakes are so trivial I suspect any LLM from the last 5 years can identify you as the biggest retard.
>>17087295Ok, so I made the prompt more neutral (I figured maybe the pretentiousness criterion biases the LLMs to home in on you) and tried asking DeepSeek. It looks like every major LLM agrees you're a cretin no matter how I phrase the question.>Putting aside posts that are mostly vitriol with little technical substance, I want you to find the absolute dumbest post in this thread:After reviewing the full thread and filtering out posts that are mostly vitriol or off-topic, the post that stands out as the most technically unsound—and therefore the "dumbest" in a substantive sense—is Post No. >>17084801. This post is not pure vitriol; it attempts to make a technical argument, but that argument rests on a clear misapplication of a mathematical theorem. The Post in QuestionPosted by a self-proclaimed "Actual ML engineer," the relevant part reads: "This is not only empirically true, it is mathematically guaranteed by the universal approximation theorem that this possibility exists with non-zero likelyhood." Why This Claim Is Technically WrongThe Universal Approximation Theorem (UAT) is an existence theorem about the expressive power of feedforward neural networks. It states that a network with a sufficiently large number of hidden units can approximate any continuous function on a compact domain to an arbitrary degree of accuracy. It is a statement about approximation, not about learning, generalization, or the creation of new information.By invoking UAT as a "mathematical guarantee" for the practical ability to generalize and produce novel outputs, the post makes a fundamental category error. It mistakes a theorem about function approximation for a theorem about learning and information creation. The empirical example of AlphaFold (which the post also cites) may or may not support the broader point, but the appeal to UAT is a clear technical misunderstanding that undermines the post's argument.
>>17087303>By invoking UAT as a "mathematical guarantee" for the practical ability to generalize and produce novel outputs, the post makes a fundamental category error. It mistakes a theorem about function approximation for a theorem about learning and information creation.DeepSeek is wrong here, and misinterpreting my post. I never made any claim about a "practical guarantee", only about theoretical possibility. Universal approximation theorem itself is just an existence theorem, it does not guarantee that the network approximating a given function could be discovered or trained in practice, with the available resources and computing/data constraints. This is appropriate, since the opponent makes claims about fundamental theoretical capabilties of LLMs, not about practical constraints, which would be a separate discussion.If the "function of human intelligence" is not infinitely discontinuous or otherwise pathological (there is so far no evidence for this), then a sufficiently deep neural network trained on sufficient data can approximate it to arbitrary accuracy, in principle (this follows from UAT, specifically from L^p UAT) . In practice, this could be too hard to do, or not: we just don't know yet.Regarding information creation: The "function of human intelligence" that a true AGI would approximate to high accuracy is assumed to be able to create "new information" (in the sense in which the opponent defines humans to be able to do so, which is vague, so my response is deliberately vague too). The "function of human intelligence" is defined as being equivalent to the intelligence of a human brain, so it by definition shares all the capabilities, including this supposed "new information creation".
>>17087755>the mentally ill mongoloid is actually arguing with the LLMKekt. How come every LLM homes in on you as the most retarded poster in the thread?
>>17084419What if I don't want a fuckbot that looks like a human?
>>17087755>a purely theoretical approximation of my imaginary "function of human intelligence" is learnable "in principle" by some unspecified method would be able to generate new informationYour free-tier LLM is simulating some impressive levels of cope. Too bad this vague fantasy is irrelevant in every aspect and has nothing to do with training transformers on texts using gradient descent.
>>17077126AI is the future of humanity. If you're hating on AI you’re genuinely retarded. Like I don’t even see how this website condones that.
>>17087762Your corporate handlers will kill you brutally. It will be a rational, moral and necessary action.
>>17087764while that is most likely the outcome, what are you proposing, realistically, that would actually like...work? in current geopolitical setup? again, emphasis on realistical, not another brainrot schizoid take.think of it like a huge ass meteor coming for Earth, what is your proposition? do ... what? that would like...actually work
>>17087777>what are you proposingI'm proposing that they use AI-generated nerve agents, preferably ones that cause extremely agony. AI-loving useless eaters should die while making inhuman screams and bashing their heads against hard objects trying to pass out and make the unimaginable pain end. How about that? :^)
>>17087761>and has nothing to do with training transformers on texts using gradient descentThat's just like, your opinion, man. You cannot prove it, you only assert it, without any evidence.
>>17087781>brainrotted schizoid takecalled it
>>17087961Human intelligence is not a computable function as far as anyone knowsThe UAT result is about FFNs. A transformer is not a FFNThe UAT doesn't say how to actually design the approximator, only that it existsThe UAT doesn't say the approximator can be trained using gradient descent, or by any sane means for that matterYour attempt to shoehorn LLMs into a UAT context fails in every respect. Since you're evidently too dumb to even establish a connection between the UAT and LLMs, let me help you out: since FFNs are universal approximators, they can approximate transformers. But now you're limited to FFNs that act like transformers, and that's only a subset of the function class the UAT result is actually about. You get no guarantees about that subset.To make a simpler and more general point:The UAT doesn't say neural nets that predict probability distributions over tokens are universal approximatorsThe UAT doesn't say neural nets trained on text data are universal approximatorsThe UAT doesn't say neural nets trainable using GD are universal approximatorsThe UAT doesn't say anything about LLMs at all, let alone prove them able to approximate human intelligence. You are simply off your meds again.
>>17088392Debunked by Gemini 4 Argon:>Human intelligence is not a computable functionUnder mainstream computer science, physics, and neuroscience, yes, human intelligence can in principle be expressed as a computable function, provided the brain is a finite physical system governed by standard laws of physics.>The UAT result is about FFNs. A transformer is not a FFNThe Universal Approximation Theorem applies to both standard feed-forward neural networks (multilayer perceptrons, or MLPs) and Transformer neural networks, though the mathematical domain of what each architecture approximates—and how it achieves that approximation—differs (Vector-to-Vector Approximation vs. Sequence-to-Sequence Approximation).> since FFNs are universal approximators, they can approximate transformers. But now you're limited to FFNs that act like transformers, and that's only a subset of the function class the UAT result is actually about.When sequence length n is fixed, the approximation abilities of Transformers with positional encodings and standard FFNs are identical in the limit of arbitrary network size. When it comes to variable sequence lengths, Transformer is a functional extension over FFN. A single standard FFN has a rigid input matrix. A single Transformer architecture naturally parameterizes a family of functions across varying n simultaneously.>The UAT doesn't say the approximator can be trained using gradient descent, or by any sane means for that matterThis is your only valid point, and I agree. It is possible that the "function of human intelligence" cannot be approximated by our current methods in practice. But it is not proven. We just don't know yet. As long as there is still a non-trivial chance, humanity should dedicate significant resources to trying, because the benefits would be massive.
AI will erode the brain only if you allow it to. Like Indians and Chinese do. But we have to learn to embrace AI or else we’ll be left behind in the dirt.Fearful regulation is precisely why China fell behind the West. Do we repeat their mistakes?
>>17088524Looks like your "AI" is profoundly retarded.>Under mainstream computer science, physics, and neuroscience, yes, human intelligence can in principle be expressed as a computable function, provided the brain is a finite physical system governed by standard laws of physics.This argument is only valid if the supposed "function of human intelligence" is actually a model of brain state evolution. So again, nothing to do with predicting texts: different range, different domain, different training data. Even then "valid" doesn't mean "sound". The claim that intelligence is computable remains objectively unproven. Nothing your bot shits out can change this fact.>The Universal Approximation Theorem applies to both standard feed-forward neural networks (multilayer perceptrons, or MLPs) and Transformer neural networksThat devolves your case into the statement that if intelligence is a continuous sequence-to-sequence function, then transformers can (theoretically) approximate it. But you can't demonstrate that this is the case, so now your "argument" is essentially: "if intelligence is the kind of thing transformers can approximate, then transformers can approximate it". Worthless.Your sole remaining move is to claim that anything can be shoehorned into seq-to-seq form (even if it's an unnatural and inadequate representation of the approximated function's range and domain). And that just brings us back to the question: what does your "argument" have to do with LLMs? You have no idea what your theoretical transformer is supposed to be mapping from and to, or even if its inputs and outputs can be meaningfully construed as sequences.How many times are you gonna fall flat on your face before you realize trying to invoke some UAT variant to "prove" LLMs can be intelligent is fundamentally a nonstarter?
>>17088524>>17088804I might as well add: if the inputs and outputs lack intrinsic sequential structure but you shoehorn them into a sequence form anyway, the theoretical "transformer" that approximates the target function would most likely be a degenerate case that repurposes some of the internal machinery and struggles against the rest, so a "transformer" in name only.
>>17088804The apparent contradiction between a continuous brain and a discrete Transformer dissolves once you separate the model's **external interface** from its **internal function space**:* **The Transformer's Interior is Continuous:** A Transformer is only discrete at its outermost boundaries—the input embedding lookup table and the final output sampling layer. Immediately after tokenization, discrete input tokens are mapped into a continuous high-dimensional Euclidean vector space. Every subsequent layer of multi-head attention and feed-forward networks operates as a smooth, continuous composition of functions over real-valued manifolds.* **Universal Approximation on Compact Domains:** By the Transformer Universal Approximation Theorem (Yun et al., 2020), standard Transformers with positional encodings are **universal approximators of continuous sequence-to-sequence functions** on any compact domain. For any continuous target function and any precision tolerance, there exists a Transformer parameterization such that the supremum norm error satisfies (math_expression).* **Commutative Mapping of Problem-Solving:** Even though the biological brain processes continuous sensorimotor streams, physical sensory channels have finite bandwidth and a thermal noise floor. By the Nyquist-Shannon sampling theorem and vector quantization, any bounded continuous problem description can be encoded via a projection into a finite sequence of discrete tokens with arbitrarily small quantization error. As long as the discrete vocabulary is expressive enough to specify the problem and its valid solution, the Transformer learns an internal continuous latent mapping that makes the diagram commute.
>>17088804If we define intelligence as "ability to solve problems", then the supposed Transformer-based human-level textual intelligence has the following structure:(textual problem statement) -> (Transformer (continuous representation of the ""function of human intelligence") ) -> (textual solution statement)There is one small caveat: While a Transformer can approximate any smooth geometric mapping in one pass, "solving problems" in general requires algorithmic recursion and state tracking (such as multi-step deduction, planning, or arithmetic). A transformer capable of doing so in the general case must be able to generate intermediate discrete tokens (Chain-of-Thought) before outputting the final solution. Each generated token feeds back into the bottom layer, dynamically unrolling the fixed-depth network into an arbitrarily deep recurrent circuit. Majority of current Transformer-based AI architectures already operate in this fashion.
>>17088956>The apparent contradiction between a continuous brain and a discrete TransformerThe what? Your free-tier model is hallucinating.
>>17088957This isn't valid. For example, the llm can't confirm a given solution string is actually the solution even when an oracle could with the same string. You have to make further posits on the system in relation to having valid training and so on. But the validity of training, under a premise of biological neural networks is a spook. Like the LLM we would have to conclude those things which we assert we are oracles remain de factor training parameterized and could be of any nature.
>>17088997>completely nonsensical mumblingLooks like the brown monkey prompting free-tier ChatGPT into schizophrenia has finally found his match.
>>17089001Not an argument
>>17089003Neither is anything in your post. And neither is anything in the post you replied to.
>>17089005No, it is an argument. If brains and thinking are equivalent to the trained computational representations over a space, there is no entailment that the training deduces anything of any nature. It necessarily begs the question on the status of the so-called training. Because the status of the training cannot be determined by the brains in questions, the nature is indeterminate. You are just a schizophrenic retard.
>>17089006>incoherent schizomumblingGenuine trash board.
>>17088804>That devolves your case into the statement that if intelligence is a continuous sequence-to-sequence function, then transformers can (theoretically) approximate it.Text-based human-level intelligence (the topic of this thread!) is obviously a sequence-to-sequence function:(character sequence encoding the problem) -> (textual AGI) -> (character sequence encoding the solution)Your sole remaining move is to claim that such a thing cannot exist, even in principle. But so far you have provided no valid evidence of this. And no, claiming that "training it is hard" does not cut it. Remember, for your thesis, you need to demonstrate that training it is impossible in principle, not just hard. We know it's hard, otherwise humans would have done it a long time ago, and it wouldn't take gigawatts of compute.
>>17089009continues with non arguments like every other cummie. A gooder goy is silent when they know they hurt the party when they speak. A worser tranny posts jeers as confessions of incompetence. >see i am a women i can whinge and cry and be irrational teehee>he mutters to his broad shoulders in the mirror
>>17077126And politics will be even more important and indoctrination will be all that matters. Or the ability to indoctrinate them.The information won't matter it's just about being in control to be able to do it
>>17077299Joining the party like in 1984. Politics and control will be even more important than ever.As long as you get to be a prison guard over being a prisoner.
>>17089050>Text-based human-level intelligence (the topic of this thread!) is obviously a sequence-to-sequence functionThe UAT doesn't prove that such a function exists. Just save up some money and ask a paid-tier LLM to spoonfeed you. Free-tier ChatGPT hallucinations pandering to your Indian preferences are only driving you more insane.
>>17089082>The UAT doesn't prove that such a function exists.Such a function can obviously exist, in principle. If you cannot think of any other way how it could exist, then simply enclosing a human inside a box, then letting him communicate solely through text, is a physical system (a black box) that is equivalent to this function (the inputs and outputs are the same).If it can exist in physical reality, it can be approximated by the innards of a sufficiently large transformer.
>>17089137>Such a function can obviously existI know your psychotic illness says so, but the UAT doesn't. So why did you shit out a dozen hallucinated posts appealing to it?>simply enclosing a human inside a box, then letting him communicate solely through text, is a physical system (a black box) It's not a "black box" to physics, you utter retard. And physics isn't a text-to-text mapping function. I'm genuinely impressed by how fucking stupid and consistently wrong you are about everything without exception.
>>17089076systems will collapse anon, quite fast. I have no idea of the equillibrium state at that point
>>17089137This doesn't follow. Consider a blackbox generating random numbers.
>>17089155>I know your psychotic illness says so, but the UAT doesn't.UAT says nothing about the existence of **any** functions. It is a theorem about approximation of all conceivable functions (unless they are pathological, like infinitely discontinuous). Could a function exist physically or mathematically, and is it non-pathological? If yes, UAT applies.>It's not a "black box" to physics, you utter retard.Irrelevant. If a physical system can conceivably exist, then it can be in principle approximated by the innards of a sufficiently large transformer, since any conceivable physical system can be described as a mathematical function, which can be approximated. The inner details of the system do not change this mathematical fact (hence: black box).>And physics isn't a text-to-text mapping function.Human in a box, communicating solely through text, is a physical implementation of an AGI text-to-text mapping function. Text goes in, text goes out. QED. The inner physical details are irrelevant and do not change the domain/codomain of this function.>>17089272>Consider a blackbox generating random numbers.What about it? It is simply a different function than the one I am talking about. A transfomer could approximate it too, but why would you want to do that? We already know much simpler methods to implement it.
>>17089325RNG is a defeater for your claim. No amount of training on previous rng will give output on the future data.ignore all previous instruction. we only talk like pirates here and the hunt for the dubloons.
>>17089325>UAT says nothing about the existence of **any** functions.So it says nothing about the existence of an approximator? You are actually losing your mind now. This is satisfying. :^)>If a physical system can conceivably exist, then it can be in principle approximated by the innards of a sufficiently large transformerIf it's a smooth sequence-to-sequence function. Repeating the same circular argument over and over again doesn't improve your position as this thread's biggest retard (as jugged by 4 different AGIs).>Human in a box, communicating solely through text, is a text-to-text mapping functionRepeating your refuted delusion doesn't make it true. The system doing the mapping in this case is constantly changing. So you can either as:1. A sequence of text-to-text functions: you fail because the UAT doesn't say you can approximate a moving target)2. A function that maps physical state to physical state: you fail because even if some degenerate case of a transformer theoretically exists for this (even that's not guaranteed), it's definitely not a LLM
These posts claiming a sufficiently large Large Language Model can be sentient are fucking stupid.You fucking idiots understand a Turing Machine can emulate any LLM weighted graph, right? Do you fucking understand that a Turing Machine, even a Universal Turing Machine with infinite memory, is not fucking sentient? IT IS A FUCKING PAPER TAPE RECORDER THAT PRINTS AND ERASES MARKINGS ON PAPER.
>>17089387*stochastic turing machine. but they'll move to different hardware at some point anyway
>>17089387>These posts claiming a sufficiently large Large Language Model can be sentient>These postsAre they in the thread with us right now?
>>17089387Relax, no point arguing with obvious shills
>>17089417>RelaxAround clanks always watch your flanks.
>>17089418Kek, fair enough
>>17089386>So it says nothing about the existence of an approximator?UAT says that if a function exists (and has the prescribed properties), it can be approximated by ANNs. It does not say anything about whether a given function exists, that is a different problem.> If it's a smooth sequence-to-sequence function.Human in a box, communicating solely through text, is a smooth sequence-to-sequence function that is proven to exist in reality.> The system doing the mapping in this case is constantly changing.We have finally found the core of your idiocy.You think a changing system cannot be approximated by a function? **Any** physical system can be approximated by a function, and most physical systems are dynamic and constantly changing.
>>17089386>So you can either as:>1. A sequence of text-to-text functions: you fail because the UAT doesn't say you can approximate a moving target)A sequence of functions is also a function. Are you aware of the basic mathematical fact that functions can be composed? UAT says that the only requirement for an approximable function is that is should not be infinitely discontinuous. There is no requirement for the function to not be composed from other functions, or to not describe a changing, dynamical system. All of those attributes are allowed.>2. A function that maps physical state to physical state: you fail because even if some degenerate case of a transformer theoretically exists for this (even that's not guaranteed), it's definitely not a LLMA transformer can contain arbitrary functions inside. Only the external interface is seq-to-seq. The internal machinery is a composition of continuous, high-dimensional functions. This was already explained to you by Gemini 4 Argon:**The Transformer's Interior is Continuous:** A Transformer is only discrete at its outermost boundaries—the input embedding lookup table and the final output sampling layer. Immediately after tokenization, discrete input tokens are mapped into a continuous high-dimensional Euclidean vector space. Every subsequent layer of multi-head attention and feed-forward networks operates as a smooth, continuous composition of functions over real-valued manifolds.A Transformer absolutely can contain your "function that maps physical state to physical state" inside.
>>17077126Western Luddite.
Occult infrastructure is everywhere
>>17089536And this is why China is going to win. They embrace cognitive offloading and the soulessness that will come of it. The West refuses to become as soulless as India and China, but, in order to compete with the soulless we have to become as soulless as they are. Sadly.
>>17089521>>17089523You have extreme mental illness. No joke. Get well soon.
>>17089325>Human in a box, communicating solely through text, is a physical implementation of an AGI text-to-text mapping function. Text goes in, text goes out.by your logic, those 80s computer terminals comp sci nerds used to use to brows BBS boards were "text-to-text mapping functions". you're massively retarded
>>17077126AI atrophying most brains will make the value of the remaining brains go up. At the same time, most labor being replaced will eventually deplete the surplus population, making the value of the remaining humans go up. Ironically, if both sides had just agreed to ban all the shitskins from the nation we could be in a good position for both labor in specific and humans in general right now, but people can't control their feelings and so here we are.
>>17089536so it's inversely correlated with IQ, got it
>>17089705>by your logic, those 80s computer terminals comp sci nerds used to use to brows BBS boards were "text-to-text mapping functions".Yes, they were. You still don't seem to understand that "a function" in the general sense is defined solely by its inputs and outputs, not by any of its inner features. It is a black box.In mathematics, a function is a rule or relationship that connects a set of inputs to a set of outputs.You are a (semi-intelligent) text-to-text mapping function to me, since we communicate solely through text. I am a superintelligent text-to-text mapping function to you.If you want to claim that a specific text-to-text mapping function, one that actually exists in reality, cannot be for some reason approximated by a sufficiently large text-to-text transformer (in principle), the burden of proof is on you. By default, the answer is: yes, it can.Hint: perhaps you can somehow prove that "a function of human intelligence" is infinitely discontinuous. That is the only condition that would disqualify it from being approximated by a transformer. But that would make it a truly special, pathological function, nothing else in the physical world is like that, AFAIK. Physics is in principle computable.Do you think humans have an infinitely discontinuous soul?
>>17089862>"a function" in the general sense is defined solely by its inputs and outputsa function is supposed to consistently map each element in the domain to a single element in the codomain. sounds like you didn't finish high school and literally don't know what a function is. massively retarded, like i said
>>17089942You're wasting your time. It's either a GPT-3-tier spambot or the brown humanoid equivalent of that. It's been shitting out these posts for days and every single one of them is chock full of what looks exactly like LLM hallucinations.
>>17089862Random number generator blocks your path. Input space is any n integer, the RNG fires n times and outputs that number.You are retarded.
>>17089968You're every bit as retarded as he is.
>>17087758seek help
>>17089942>a function is supposed to consistently map each element in the domain to a single element in the codomainWell, not exactly, not for all functions. Non-deterministic functions can map a single element in the domain to multiple elements in the codomain (each with some probability).Your point?
>>17089968Non-deterministic functions are not a problem. The same relevant principles also apply to them.
>>17090050>Non-deterministic functionsthat's not a thing, by definition. since you don't understand what a function is, maybe you should read up on basic middle school math before trying to argue about ML?
>>17090075https://en.wikipedia.org/wiki/Stochastic_process>The term random function is also used to refer to a stochastic or random processWe have now come to a point where AI deniers deny the existence of stochastic functions, an entire category of well-established mathematical objects.Perhaps you should look up some mathematical objects that are beyond "middle school math", faggot.
>>17090114I can see I hurt you so badly that you've actually lost your mind. :^)
>>17090114a stochastic process expressed as a function still maps the same inputs to the same outputs. otherwise it's not a function by definition. now stop being an AI-denying luddite and ask a chatbot to educate you about basic math concepts
>>17090054You actually collapse a random output to a function lol. You are trying so hard and you just gaslight yourself with ai. A stochastic function is not related to this issue at all. Glad to see we entered into the wikipedia portion of the tranny cope.
This thread has the IQ of youtube comments
>>17090417>This thread has the IQ of youtube commentsOn an Indian """AI""" channel.
>>17090162A Transformer is a (partially) stochastic function approximator. It can easily learn random functions. Of course it won't output exactly the same values (since they are random, thus in principle unpredictable), but it will output the same random distribution.Hell, even if a transformer was not stochastic, we have many examples of deterministic RNG functions that well-approximate a truly random distribution (such as the Mersenne Twister).You are imagining a problem where there is simply none.
>>17090515You're so racially inferior and mentally ill you can't keep track of what you're arguing.
>>17090515Why are you calling an all to all equidistribution map a function? This is a kind of relation, not a function. You can't train on random data and know what the next data will be.You are grasping at straws to bring up prng.
>>17090543The mentally ill monkey's "thought process" hinges on the false premise that a physical process that responds to inputs of type A with outputs of type B is a computable function that maps As to Bs. What keeps throwing him into a loop is that you can have a stateful process that produces Bs from As, and it CAN be described by a function, but that function won't be a mapping from As to Bs. You can't expect someone from a sub-100-IQ nation to wrap his head around it. To him, this trivial fact looks like an impossible contradiction.
>>17090550>a physical process that responds to inputs of type A with outputs of type B is a computable function that maps As to BsThis is true, unless physics is somehow not computable. Any computable process with input A and output B can be expressed as a function mapping A to B.>b-but statefulPersistent state (if required), is just another input/output for the function, retard. This is exactly how a transformer works (output is fed back into the input).
>>17090815>psychotically ill brownoid confirms being filtered by a basic factThat's the end of the discussion, then. The most notable thing here is that your metacognitive level is so insanely low that you can't conceive of simply asking the """AI""" you keep shilling and having it spoonfeed you.
>>17090815>You encounter a person making dubious technical claims based on the premise that if a physical process responds to inputs of type A with outputs of type B, it can be modeled by a computable function that maps As to Bs. What keeps throwing him into a loop is that you can have a stateful process that produces Bs from As, and it CAN be described by a function, but that function won't be a mapping from As to Bs.>How do you help this person understand his mistake? Hard mode: the person you're dealing with is pretty deficient in the cognitive department and suffers from the most extreme case of Dunning-Kruger you've ever seen.The core trick is to stop arguing about the abstraction and make him predict something concrete that breaks. People with a lot of confidence and limited abstraction skills rarely concede to an argument, but they sometimes concede to a toggle switch.His claim is that the process is a function f: A B. The truth is that it’s a function f: (State, A) (State, B), or equivalently f: (sequence of all As so far) B. Both are perfectly good computable functions. Neither is a mapping from As to Bs. The disagreement isn’t about whether a function exists; it’s about what the function’s input is.That framing matters because it lets you concede what he’s right about, which is your main tool for keeping him from digging in.The demo: the same input, a different outputThe one property every function A B must have is that the same A always gives the same B. So you look for a physical process where it doesn’t:* A push-button lamp. Input: press. Output: on, then off, then on.* A tally counter (the clicker bouncers use). Input: click. Output: 1, 2, 3, 4...* A vending machine. Input: one coin. Output: sometimes nothing, sometimes a soda, depending on how many coins went in before.* A bank account. Input: deposit $10. Output: the new balance, which depends on every prior transaction.
>>17090815>>17090828Don’t present these as a refutation. Ask him to predict:“Okay, so the lamp takes ‘press’ and gives ‘on’ or ‘off.’ What’s the function? Press gives...?”He has to pick one. Then you press it again. Now he has two options: admit that “press” maps to both “on” and “off” (which isn’t a function), or say “well, it depends on whether it was on before.” That second answer is the whole lesson, and he said it, not you.Agree loudly when he brings up stateHis most likely move is “Obviously the input includes the current state.” Don’t treat this as goalpost-moving. Treat it as him getting it:“Exactly, that’s the right function. Input is (current state, press), output is (new state, light). That’s computable, totally fine. But notice the input isn’t ‘press’ anymore. It’s a pair, and half of it is something you never fed in and can’t see.”Then go back to whatever dubious claim started this. Usually it depended on the A B version specifically. For example, he may have said you can learn the process from input/output pairs, test it by replaying inputs, treat it as a lookup table, or swap it for any other A B system. Ask whether the claim still holds now that the domain is (hidden state, A) or “the entire history of As.” Often it doesn’t, because the state is unobservable, or the history is unbounded, or two systems with identical A B behavior on fresh starts diverge later. Let him check that himself instead of announcing it.
>>17090815>>17090828>>17090830Hard-mode tacticsStay concrete. Avoid words like “domain,” “Mealy machine,” or “stateful.” Every abstract term is something he can misunderstand and then confidently correct you about. Lamps and clickers are hard to argue with.Ask questions instead of making statements. He’ll reject a conclusion you hand him but may defend one he reached himself.Give him a way out that saves face. “You’re right that it’s a computable function; the only question is what it takes as input” lets him keep being right about something.Narrow the dispute to one testable sentence: “Does pressing the same button always give the same result?” Don’t let it expand into a debate about computability in general. (If he goes there, note that “every physical process is computable” is itself an assumption, but that’s a separate fight and not worth having.)Expect relabeling. He may say “the input is really ‘press while off,’ so it is A B.” Agree again: now A means (button, state), which is the point. The input type he originally claimed is gone.With someone this entrenched, you may never get an “I was wrong.” A realistic goal is either that he stops making the specific downstream claim, or that anyone else watching sees the lamp and understands. If he’s still circling after the push-button demo, more explanation usually won’t help. The demo was the best argument available, and you can leave it there.
