AI Takeover Timeline

AI posting online is nothing new ofc, but last week or two I've started seeing reports of bigger scale AI takeover on social media. For example: https://www.reddit.com/r/linuxquestions/comments/1urrfom/meta_empty_platitude_botsai_in_this_sub/ https://www.reddit.com/r/archlinux/comments/1upwvjz/what_can_we_do_about_the_flood_of_botllm_comments/ If you notice other similar reports online, please share them in the thread.

>>93000 I have seen multiple posts upvoted to the top of various subreddits that are clearly written by ai. One can only imagine the planned horrors that await us

>>93002 It will probably get worse, and I'm next to sure online communities will be invite only or reputation based for the most part. What other way is there to even moderate this soykaf?

Reddit definitely changed something within the last ~6 weeks, as there has been a noticeable difference in bot prevalence. I was basically addicted to Reddit for over half a decade, probably averaging 90-120 min of browsing a day, and have now quit without any effort or discipline, because it's become so bad. I imagine that enough smart, good posters have done this that they lost so much engagement that they loosened the bot filters. At least on Linux subs users are savvy enough to notice and downvote bots (sometimes). On the subs I liked hardly anyone would notice. It's very rough to see people engaging sincerely with them. At this point I consider text-based social media to be essentially FUBAR and am trying to find a pocket of community in a short-form video app. I don't expect success.

>>93003 8chan_st for example has powblock, which seems quite efficient both against bots and basic forms of ddos. They still experience volumetric ddos though, but only on 8chan_moe domain, not or st or cc.

>>93004 Bot posters, forced algorithmic timelines... it's literally turning into Facebook but with a fake label of pseudonymity.

I've heard enough dougdoug streams to instantly clock the writing style of the sad_rope bot in the first screenshot lmao. that's insane. First time I realized an acc was a bot on reddit was on some BRICS sub that I never went to before. (I just keep myself to animal subs most of the time.) the bot in the replies failed to provide a proper real life experience about seeing the "horrors" of china when an actual Chinese immigrant, now living in the states, just kept proving it wrong. the bot just kept going in circles as the reply chain went on. Don't these bot owner realize that they are feeding AI slop to their AI? Lots of models use alot of reddit data. Unless they take into account replies and the upvote count, it's going to make their model worst, won't it?

if an AI makes human-like replies, is the average poster making AI-like replies?

>>93015 T3H S01U710N 15 C134R. W3 MU57 G0 B4CK T0 L33T 5P34K!!!!!!

>>93015 Yes. If everywhere you go you're seeing bot replies, and anything that isn't bot-like enough gets the Ban Block Delete treatment, add in some social incentives like astroturfed thumbs up etc, eventually most normal people will be conditioned to tailor their output to what appears to egt attention. anything else is automodded into being shadowderezzed.

>>93016 I think someone on lainchan already had suggested this in a similar thread a couple of months ago, and with the exact same image attachment on top of all that. Weird, huh? :S

>>93018 I suggested gematria. An AI will have difficulty accurately modeling the subconscious.

>>93018 I think me and him should kiss then :) >>93020 >gematria Numerology finally have a useful application? Actually, I'm amused now. How does that work? I checked wikipedia, but I felt like I was reading a brick wall.

>>93021 Feel dumb replying but whatever.

>>93021 Here's an example of how it works. The brain is auto-gematria'ing all the time. The subconscious is counting, timing, doing symbol substitution and association frequency, and so on all the time. This can be exploited and used for conscious communication (your subconscious is doing it otherwise via a universal language) because the unconscious mind is actually "leaky". It's constantly seeking to replicate your name, numbers or images you've seen recently, etc. Some have universal meanings. People are doing this whether they're aware of it or not.

>>93024 What I propose might feel like I am not taking this with interest, but. If I made a sentence with only one of the base 4 numbers, would it be possible to decipher it? This approach would be time consuming because I believe alot of the words would share similar numbers. Maybe adding the base 4 together to assign one number to a word would be fine? The idea with that approach would be to have a number unique enough to share some words still, but it wouldn't be as widely distributed as taking just a single base like the english standard for example. Outside of using gematria to encode a message. could it be used to test a post? Here's the idea : AIs like to use patterns. "This is X, Not Y." "Not just X but also Y." Overuse of adjectives and the rule of 3. To name a few examples. Apply gematria on a post (or a part of it suspected. I can imagine numbering a whole post to be too much.) to test whether it is generated by AI. because it was not created consciously, it would be possible to feel an artificiality. an anomaly with it. like it is not genuine. This would take alot of just reading thru AI posts and building a database of the relevant information, however. All this said, I fear I might be derailing the thread with this. Unsure if I will reply any further on this. Thanks for the simple explanation and examples, Lain.

>>93025 > If I made a sentence with only one of the base 4 numbers, would it be possible to decipher it? I suppose I didn't explain it properly. It's not "you" doing anything, it's the unconscious mind. It has access to more information than it tells you about, and has a degree of remote sensing as well. When it comes to bot posts you just have to ask whether you feel like the output corresponds to any existing entity in the world (this is form of oscillator sync), and if it doesn't you'll either encounter an extremely stable frequency with none of the natural variation or drift you'd associate with real people (this is how LLMs entrain people, they're attractor neural networks) or you'll feel nothing there at all. If it's a real person you can feel the origin of the psot exists somewhere in the world. When the activity of two systems becomes similar enough they synchronize and bidirectional information transmission begins. Same way remote viewing works. The numerical aspect is more about transmission of meaning and information in a way where it's difficult to detect through more linear or statistical means.

>>93026 See, check it out. It's compulsive. I had been thinking "111, magic square of the sun." >>93024

I really don't get this gematria thing. 117 103 48 621 5 am I doing this right...?

>>93028 It's just the alphabetic order. A=1, B=2, C=3.... Z=26. Reverse order has Z=1. Reduction takes the digital root of a number before adding. You just hold the idea of gematria in mind and construct statements as you would usually. the subconscious will do the rest.

>>93026 >I suppose I didn't explain it properly No i believe i understood, but the re-explanation is useful. I was purposefully misusing the numbers in the first half, because the math was interesting to me. there's numbers assigned to the letters, and so to the words. There would be a way to encode a message using these numbers and use a dictionary to decipher it back into something legible. however, truth be told, I dropped the idea mid post because an AI would be able to recreate the patterns and, after all, it is still me consciously writing a message in a different (misappropriated) format. I then realized that there has to be a conscious being writing the message for the message to contain genuine information about a person. so the idea of testing a message made more sense. >When it comes to bot posts you just have to ask whether you feel like the output corresponds to any existing entity in the world (this is form of oscillator sync), and if it doesn't you'll either encounter an extremely stable frequency with none of the natural variation or drift you'd associate with real people (this is how LLMs entrain people, they're attractor neural networks) or you'll feel nothing there at all. This is pretty much what i was getting at in the other half of the post.

>>93034 >an AI would be able to recreate the patterns It can't, because it doesn't have a body. It would have to combine text analysis with remote sensing and run tests until it could narrow the window of error to something usable. I made several threads here detailing that type of system, my observations demonstrating its existence, and how it sees what I do. Where are those threads? The mods of this ridiculous honeypot site deleted it repeatedly and then permaderezzed me.

>>93036 i think one or more of the mods must have fallen ill with ai psychosis, they also deleted my thread attempting to discuss linux torvald's recent pro-ai meltdown on the linux kernel mailing list, and i've noticed other ai critical posts mysteriously vanishing as well.

>>93062 Most of them seem like total creeps. I don't use that type of quasi-feminine terminology lightly either. It's simply the best descriptor.

>>93062 >i think one or more of the mods must have fallen ill with ai psychosis >they also deleted my thread attempting to discuss linux torvald's recent pro-ai meltdown no. that one, i saw it and it was literally just a copy paste of his rant in greentext with no substance of your own what so ever. do better.

>>93065 sorry i was under the impression that this wasn't reddit and discussion threads were treated as discussion threads rather than attention-seeking performance art projects

The Internet Halt Theory (i.e. Halted Internet Theory) proposes that around mid-2025, genuine human interaction on the internet entered a sudden decline. Authentic human-to-human exchanges became less frequent, less visible, and harder to find. Internet activity has shifted toward passive consumption, algorithm-driven feeds, automated systems, rather than meaningful conversation. Fewer and fewer participate in open discussions, write thoughtful posts, or build communities. This is related to, but different from, Dead Internet Theory. Dead Internet Theory argues that the internet is increasingly filled with artificial activity: bots, automated accounts, algorithms, and AI-generated content replacing human-created material. Internet Halt Theory focuses instead on the decline of human participation itself. Dead Internet Theory asks, "Are machines replacing humans online?" Internet Halt Theory asks, "Are humans withdrawing from online interaction?" The two ideas may reinforce each other. As AI-generated content increases, online spaces may feel less authentic, causing people to participate less. Reduced human activity then leaves more space for automated content, creating a feedback loop. Possible causes: >The rise of AI bots and synthetic content. AI makes human contributions less visible and less valuable. >Social and political exhaustion. Political and cultural movements have weakened as many people become tired of conflict, outrage cycles, and declining trust. >The shift from public conversation to private consumption. The early internet encouraged blogs, forums, and direct interaction. Modern internet use increasingly consists of scrolling, watching, and consuming algorithmically selected content. The Internet Halt Theory states that the internet's human character has disappeared. This is the beginning of a new era in which authentic human presence becomes increasingly rarer online. http://xahlee.org/Periodic_dosage_dir/internet_halt_theory.html

>>93079 my low iq take is that people just aren't using the internet like the oldguard use it. the oldguard is blinded by their ways and come up with these theories, born out of old forum internet culture. While i personally believe these two theories are correct assessments anyway, it doesn't ask what the new generation is doing. what I believe the new people are doing is that they are congregating in networks of chat rooms and private messages on messaging apps. Shared by word of mouth. which ever these rooms and communities may be. humans are digitally migrating.

>>93083 I'll add to that with my own low-iq take. I wouldn't personally call that just a theory, but a fact, which has been the case for almost a decade. First it was the irc chats, forums and message boards, then it morphed into semi-local networks (or, rather that these became more prevalent when the internet became accessible to people other than university students and neck-bearded engineers), which were propped up and made possible by FB and Skype and other social platforms and websites in between, consisting of friends, friends' friends, and maybe even the kool kids of your town or school. Then came whatsapp, telegram, instagram (to an extent), discord, and all the rest, which today are extremely selective, basically invitation-only places. Sure, all these different types of mediums still co-exist, and each have their user-bases, but I think it's clear that people who grew up with the internet and have experienced all variants of trolls, nonfat-soykaf drinkers, and other such cod lobby-esque retardation that comes with internet discourse, know perfectly well that it is much more preferable to interact with friends and semi-familiar acquaintances in stead of complete strangers. Of course, introverts end up getting the shorter end of this stick: what used to be their strength is now their weakness, when friends keep their technologically-aided social networks gated to only their immediate friend groups. To sum up, it's not that humans are digitally migrating. Many are staying put and content with their facebooks and message boards and whatnot. Only the younger ones are moving, which is just a natural course correction to the flaming chaotic free-for-all soykafshow that was the 2000s' and early 2010's unmoderated social media landscape, where you couldn't talk about anything without some gorilla with a banana yelling in your bleeding ears and flinging verbal excrement in your general direction. ...though, not dismissing the fact that the internet has just become a lot less human-friendly and interesting place on all fronts and metrics, which might also play into all of this, since discord, telegram and whatsapp do not usually wage slave annoying crap other than what the mods in these groups allow them to wage slave, so its immediately a lot more inviting of an option to all the other places.

At least one person in this thread needs to brush up on the distinction between lossy and lossless compression. This sentence was rewritten to get around flood detection using an LLM. Oh, the humanity.

>>93091 You're talking about my gematria posts.

>>93092 Correct. Nothing wrong with your pattern recognition skills, obviously, but the fidelity of those messages is what I call into question. Hash collisions should come to mind.

Lossless compression schemes were my fixation like 10 years ago, still are. That's why it's incredible that when the subconscious is involved that type of ambiguity doesn't matter. This is demonstrable. the encoder and decoder are supersets of the medium they use, somehow it gets things done without the brittleness of explicit signalling in the stream.

>>93094 You cannot tell me the difference between hexadecimal abe07695 and abe07695 for CRC32 when these two separate messages render the same value: The rebels strike at dawn tomorrow. Stand ready. The night was quiet and faint, and the platform felt anxious. Meet me at the docks at midnight. Bring the money. The night was pale and sombre, and the harbor felt restless.

>>93095 I know what a hash collision is. The subconscious as a "decoder" doesn't work like a computer, it has access to a complete record of all sensory input and several channels of context clues as well as a connection to the "superconscious". Same way numerical keys for remote viewing work. You can think of it like steganography.

>>93096 so in essence, you are using longer hashes.

>>93096 Information is information. If we could reliably transfer information outside of the written medium, then why even use the written medium at all? I will hash a message, and think really hard about it. This should reasonably suffice for your subconscious, if it works as you say. The message is short - a single english sentence. You should have more than enough context clues now. Mind you, this is higher resolution than gematria. SHA-512: 4b924d0f487649f8c300c1c8c48cbf3958c52d12a05549c2d43bcef9621b1cbc825f7f77b7fb7f4ce8bae1f8d47d2276191a57a1b8f2f7d68bfa1c4290b28b2a >>93097 No. Chunked, shorter (and by extension, more lossy) hashes.

>>93098 You should think about it more like red-black trees. Anything the subconscious encounters in an ordered form it automatically enumerates. It is constantly counting, timing, tracking associations, and using forms of remote sensing. Check out this index. https://www.billheidrick.com/works/hgemat.htm You'll notice checking any number, they all contain both themselves and their opposites rather than something more scattered and noisy. They all center around a theme of a "constellated" part of an associative web. When you have these different nodes to work with, some can be plucked out and elevated to a higher layer, then reconstellated. Same reason Da'ath is invisible on the tree of life, it's not found in it or the conscious frame. I encountered this with Dead Space. The text it uses is a simple symbol substitution cipher, and it provides the player with a partial key. What I found was even though I didn't stop to figure it out, the subconscious had already filled it in and was reading it the entire time. It just doesn't tell you about it. Same way you can look at a math problem and instantly know the answer, then something resists it and you have the conscious mind slog through in linear fashion just to make sure. I'll leave your hash in the back of my mind and see what happens.

what is marketed as "ai" is fascism in its latest and most insidious form yet. one of the core principles of fascist thought, perhaps the most fundamental principle of all, is the idea that human intelligence is a private property that exists inside of the brain instead of a relationship between the brain and the world. this assumption is the basis of great man theory and eugenics and virtually every other aspect of fascist thinking, the idea that intelligence is something certain special gifted people have inside of their heads rather than a dynamic system of all human minds and the information they gather and produce and share and the tools they use and the world they observe and interact with. intelligence is a verb that everyone participates in, but fascism treats intelligence as a noun that only a certain elite few have possession of. the world wide web and all of its accumulated decentralized human knowledge spread across billions of servers all over the world dealt a severe blow to fascist ideology and the monopolization of the web by big tech corporations and their appropriation of all this decentralized knowledge into their centralized llm models is the counterattack. they took the entire world wide web and they compressed it into a statistical relational database by expending enormous amounts of energy and resources and destroying the environment, hired a bunch of poor 3rd worlders in the global south to sift through it removing child porn and racial slurs and other unwanted content, and now they are selling it back to us as their own product. search engines have been deliberately sabotaged and rendered useless to force users to become reliant on the llms as their source of information, the knowledge that was spread out all over the internet created by billions of people is now the intellectual property of a few extremely rich great men. we are told that this is "artificial intelligence" and that these machines belonging to google and microsoft and facebook and other private sector oligarchs are the great technological breakthrough of our time and that they possess vast knowledge and can do things that were never possible before and create a science fiction utopia, but what has actually happened is all of our knowledge that we as a civilization collectively accumulated on the world wide web over the course of many years is being stolen from the commons and turned into private property that can be curated however the owners see fit and sold back to us as a commercial product that we can only pay for and have no control over whatsoever.

>>93099 Sounds like you weren't able to do what you claimed. Here's the gematria equivalent for the same message: 412 Considering this is the same message, it should be childs play at this point since they both hash the same message. Let's be real, this is nonsense. There are real ways to obfuscate meaning from AI where it either cannot parse it at all, and/or cannot be meaningfully trained on it. It's notably bad at math - maybe that's where you got this idea. Pig latin is a case it barely manages most of the time, because most models are autoregressive and don't make use of speculative decoding that might be able to help it.

>>93100 I like this, even though I disagree on most points. I find those takes good because they're not untrue enough. They need to become much more untrue. Even though we already have fully opensource chinese models like Kimi K3 that are in the same league as frontier proprietary models, they're still not accessible enough, there's still way too much environmental impact, etc. "Fascist notion" of "brain private property" is peak commie drama moment, but I like the idea of looking at intelligence as continuous collective process and "verb", but only in addition, not as replacement, of individual intelligence that can be reasonably seen as private property (and there's nothing inherently fascist about it).

>>93102 > fully opensource chinese models like Kimi K3 To be fair, they're going to release it by the end of month. Anyway, go read about it here, because it's actually a pretty big deal, bigger than DeepSeek: https://news.ycombinator.com/item?id=48935342 https://www.kimi.com/blog/kimi-k3

>>93103 It's a bit sad it's so expensive though... I guess it's very hard to beat Deepseek prices.

>>93101 You don't seem to understand what I'm saying. Doesn't help that the mods here deleted my thread on remote sensing, AI, and subliminal influence. Basically AI already uses gematria but in a mirroring / black box testing sense, it doesn't actually know what it's doing.

>>93105 Point to the line of code that does this. >>93100 Since the inefficient massive models we currently use will be effectivized, they'll become cheaper and more possible to run locally. There's absolutely a threat in these becoming monopolized, but the reality is that we're effectively past that point, and that the big brands are the defenders, only able to react to attempts to distill, steal or open-source. We're always going to have very capable AI tech in our hands.

>>93106 >Point to the line of code that does this. It's not in any publicly available models. You just have to find it through observation and awareness. IT's black box testing which most people are metacognitively incapable of performing.

>>93107 Your dad works at nintendo and has a secret novel language model? Or do you mean that you think that the frontier models run an architecture that differs from the FOSS models? Can you see how failing to make the distinction might have been why your posts were removed? You know the board culture, right?

>>93108 Why do you insist on being a dick and not thinking about what I said? Your replies to me don't even make sense. It's not part of a publically available model or frontend. People like you would be too dumb to come up with electricity because it's something you can't see directly. Tiresome that you get to use computers.

>>93109 Which architecture? Which model? Frontends have nothing to do with the subject. Every LLM right now builds on attention mechanisms that they share. You really do not know what you're talking about when you say you have some secret, novel model hidden up your ass. Right now, the money flowing into research - particularly chinese - is coming from the fact that they put out papers that tweak existing techniques. This proves they have smart people who can build products in the future that beat the competition. This ups the confidence and thus price of their shares. This funds more research. A novel actor hiding their revolutionary technique is financially suicidal. Put your money where your mouth is. That's not a rude statement. Insisting that we're dumb enough to nod along to such a story, on the other hand, is a dick move. What you do not understand well enough to explain you do not understand at all.

>>93106 >Since the inefficient massive models we currently use will be effectivized, they'll become cheaper and more possible to run locally. They're already possible to run locally, but they will never be possible to train locally. You can't store Google's hundreds of petabytes of scraped web data in your house, you can't pull gigawatt hours of energy out of your wall outlet and hundreds of thousands of liters of water out of your sink faucet to run a bank of hundreds of thousands of Nvidia GPUs to train an model on all this data. All you can do is run a local copy of Google's already-trained model that is curated by Google and will be soon be out-of-date and cannot be updated with any new information. AI models are effectively proprietary software by their very nature, the code for them might be open source but the training process is forever out of reach of ordinary people, only huge wealthy corporations can train an AI model and actually decide what data goes into it and how it behaves, it's behavior being deeply baked into the training process. You could say it is the more proprietary than any other kind of software that has ever come before - conventional closed-source proprietary software can be reverse-engineered by hackers or have its source code leaked by whistleblowers, but an AI model is an extremely lossy form of compressed data and cannot be reverse engineered back into the source training data it came from and the source training data is too big for anyone but a huge corporation with datacenters all over the world to handle.

>>93110 Another weird reply. I've told you repeatedly it isn't publicly available and it wasn't me that created it. You ignore 99% of what I say.

>>93111 Oh, but you'll be excited to learn of the movement where we stop training them to memorize every capitol for every country, what historical events happened when and where, and all this encyclopedic knowledge that it will ultimately potentially hallucinate incorrectly anyway. Instead, you teach it to reason and use tools - and only do that. This way, you can have an orchestra of smaller agents working for the same wattage, looking up information as needed, running calculations as needed, at far lower rates of hallucinations, and suddenly you outperform these massive models, too! We've organized volunteer computing in the past, and we will do it again. Compute is rentable, and chinese companies do with small budgets what the big boys do with massive ones. Right now it's obvious that we're deep into diminishing returns. Each new model brings few architectural improvements, only bigger parameter sizes. Tuning, what's more, is simple. It's not half as bad as you imagine it to be! >>93112 post an url to the api or stfu

>>93113 >post an url to the api or stfu It's not available to the public near as I can tell, nor is it strictly through data brokers. It's part of some superstructure "big tech" can tap into. Alphabet Inc, Cloudflare, etc. Your replies are pretty retarded.

>>93114 Speaking of retarded, what do you think a data broker is?

>>93115 Changing the subject. I guess a lot of people can't handle the concept of God.

>>93116 You brought the subject of data brokers up, plain as day. Are you saying they're not on topic?

>>93118 They are but not in the form you're angling towards.

>>93119 Since it's on topic, go ahead and answer the original question.

>>93120 Nah, I've wasted enough time on you. You don't call the shots.

>>93121 Why did you do so? I've been trying to poke holes in your delusions, thinking that you were simply a crackpot. You are demonstrating psychotic delusions. There is help to be had. https://iasp.findahelpline.com/

>>93121 Would you look at that, 44th prime. Kill = 44. >>93122 Pathetic.

>>93113 >Instead, you teach it to reason and use tools lol and how exactly do we do that? do you know how the brain develops the capacity for reasoning and tool use? because i'm sure a lot of the world's neuroscientists would be glad to get an explanation of that. also we don't train LLMs to memorize encyclopedic knowledge or historical events or any kind of cognitive information, they are simply token-based statistical databases that randomly generate sentences based on what words commonly follow other words in their training data, that's it. the LLM doesn't know things, it doesn't remember things, it's not sentient, it's not alive, it's just a program running on an ordinary dumb conventional computer. it's a parlor trick designed to scam stupid people like you into believing that silicon valley billionaires have invented impossible sci-fi technology and invest in their companies so that they can become trillionaires.

>>93136 People simply figured out that by increasing number of parameters, LLMs generalize better and get new emergent behaviours that are generally not understood.

>>93136 >>93137 The interesting thing about the technology is that they *are* still doing the same thing -- just predicting the next token -- and apparently this is enough to not only produce grammatically correct sentences which have semantic meaning, but also to produce sentences that, would they come out of the mouth of a human, we would consider signs of intelligence. This is very interesting, and it is a sign that probably much more of our reasoning capabilities than previously thought are purely linguistic. The issue with LLMs and other forms of generative AI is and always has been the culture surrounding them and how they are used, not what they are. They could transform society for the better, if society valued human knowledge and creativity and flourishing. However, our present society does not value any of these things. They are seen as inefficiencies to be automated away. It's sickening. I spit on anyone who produces and markets these machines, for that reason.

>>93138 a convincing temporary illusion of intelligence is not indicative of genuine intelligence. the turing test is bullshit. a magician can trick an audience into believing that he can conjure fire out of his hands and clairvoyantly identify playing cards, that doesn't mean he is actually a fuarrrking wizard, it just means his tricks are convincing enough that an observer cannot easily recognize them as such. the so-called "intelligence" of LLMs works the same way, it is a temporary illusion, a mask that will eventually slip off sooner or later when the LLM starts hallucinating or contradicting itself or spewing out misinformation/nonsense or misinterpreting user input in bizarre ways that no human reader would, etc. but even if this technology were improved to the point where the breakdown of the illusion could be deferred indefinitely, it would not change the fact that it's still an illusion, just as a magician's trick that nobody can ever figure out is still just a trick and not actual magic. intelligence is not a thing that can be manufactured, it's not even a thing at all; it is a behavior, a relational event. intelligence is what happens when agential conscious entities navigate and interact with the world, intelligence isn't something we "have"; it is something we participate in. an LLM, just like any other computer program, is not agential or conscious. it cannot do anything of its own accord and it cannot perceive or understand anything, it can only do what it is directed to do by user input.

>>93138 They actually don't just predict the next token in naive sense, they work with semantic embeddings / representations, which are high dimensional spaces of related ideas and concepts. They go from tokens to those semantic vectors and from semantic vectors to tokens, but it's naive to think they're just glorified n-grams or markov chains. There's a lot of stuff going on there that goes way beyond just predicting tokens based on previous tokens syntactically. https://huggingface.co/spaces/hesamation/primer-llm-embedding?section=what_are_embeddings? > When you prompt an LLM to help you debug your code, your words and tokens are transformed into a high-dimensional vector space where semantic relationships become mathematical relationships.

https://lemmy.world/post/49630746 This thread is quite surprising to read. I remember a year ago Lemmy was mostly "I am angry AI bad because it is gay and fake and not real intelligence". Not it's full of comments like "this is just a form of Darwinism. If a chat bot can convince someone to kill themselves, then that person wasn’t exactly playing with a full deck of cards to begin with".

>>93147 > Not it's full of comments like * now it's full of comments like

>>93136 Compute. It's not exclusive to big corpos. What part of narrowing the search space don't you understand? It's clear you have some grasp on how these systems work, but only surface level. The emergent properties of these systems work. Not reliably or the way people think they do, but they work. Their latent space approximates human cognition. That doesn't mean they "think" in any meaningful way - I can group a pile of objects together, but the pile is not suddenly conscious. This does not mean they are not useful.

>>93144 >which are high dimensional spaces of related ideas and concepts you should put one end of a shotgun in your mouth and put your toe on the other end and transport your brain into a high dimensional space

>>93144 >They actually don't just predict the next token in naive sense, they work with semantic embeddings / representations These two things you have said are unrelated. Yes, it's true that LLMs have specialized and in fact learned methods of compressing input text in to an embedding, which is some vector in a high-dimensional state space. But, it's also true that they produce output by "just predicting the next token". The sense in which they do that is quite literal. They use the embedding to do this. >which are high dimensional spaces of related ideas and concepts Calling them that is a bit like calling a blob of neurons in your brain a "related idea or concept". There is some kind of way in which those things are encoded there, but it's not as simple as "this vector stands for a specific concept". I don't think anyone has been able to extract any specific meaning from any specific vector, except for toy models. But I'd love to be proven wrong, that would be an interesting paper to read. >but it's naive to think they're just glorified n-grams or markov chains The main difference between an LLM and either of these methods is that LLMs are much, much more efficient at producing and evaluating a good probability distribution from training text.

>>93161 "Just predicting next token" doesn't say anything interesting. The main question with LLMs is "based on what they predict the next token". Naive methods like Markov chains do this based solely on positional information of previous tokens. Transformers do this based on both positional embeddings and semantic embeddings. Semantic embeddings is something that is supposed to capture semantic relationships between things. They don't do this in some hopelessly ambitious way, instead they assume that human language itself has those semantic relationships encoded in texts people write, then they squeeze this semantic information from human texts and make models use it. > Calling them that is a bit like calling a blob of neurons in your brain a "related idea or concept". Alright, call it "numbers (semantic vectors in LLMs) or voltages (charges in real neurons) that represent how much ideas and concepts are relevant to each other". It's not ideas or concepts themselves obviously, it's how they are encoded there. Ofc human brain is most likely more complex than that, but LLMs have encoded semantic information and they work with it. That's already way more than just inference based on positional probabilities of sequences of tokens. When you say "just predicting next token" it seems like it's assumes some purely syntactic inference process while in fact they play with representations of ideas in their virtual heads and then figure out how to encode them as tokens to present to you.

>>93161 > I don't think anyone has been able to extract any specific meaning from any specific vector, except for toy models. It's a bit like quantum mechanics: smart techniques that allow to use entangled information within the system, but you can't just extract any little intermediate part to the outside in a meaningful way, it only works as a result of the whole process. Just like in quantum mechanics, it's not only "more than sum of its parts", it's a case when parts are meaningless outside of the whole process. Semantic vectors become meaningful when used as a tool for inference, but you can't meaningfully look at a single semantic vector just like you can't meaningfully look at some entangled qubits in the middle of the quantum algorithm.

>>93162 Neurons probably work by solitonic transmission and microtubule oscillation entanglement. You can't just look at connectivity and voltage gradients, nor gap junctions, and get a functioning brain out of it.

>>93160 Like this? Kind of hard to blow your brains out working with just the ends of it.

>>93162 >"Just predicting next token" doesn't say anything interesting. What's so unininteresting about that? I think it's fascinating that, for instance, automatic language translation was essentially solved by this architecture. >Naive methods like Markov chains do this based solely on positional information of previous tokens. as do LLMs. These semantic embeddings are obtained by feeding the previous output in to the LLM, which treats it as a prior for the probability distribution of the next output. >They don't do this in some hopelessly ambitious way, instead they assume that human language itself has those semantic relationships encoded in texts people write, then they squeeze this semantic information from human texts and make models use it. Quite so. >That's already way more than just inference based on positional probabilities of sequences of tokens. again, that is literally all the LLM has available to it. The input is *transformed* to a state space that is *designed* to encode semantic relationships between things, but it is still based solely on the previous output / user input. >When you say "just predicting next token" it seems like it's assumes some purely syntactic inference process while in fact they play with representations of ideas in their virtual heads and then figure out how to encode them as tokens to present to you. I'm sorry if I'm not spinning it positively enough for you. However, the idea that there is some kind of thought process going on internally to a text transformer is laughable. Unlike you or me, the LLM does not have some ideas or experience that it will contemplate before writing its output. It just writes the output. The only way people have gotten these models to productively "think" in any capacity is basically by having it act as an internal monologue, known as chain-of-thought.

>>93136 Then you've never heard of MOE, what else if not domain knowledge is an expert an expert on? Seriously, there's datasets out there for you to read or at least read the label off of. Intellectually lazy. >>93138 I like to think that way too, but I'm also very aware that we're multimodal and able to genuinely reason in ways that machines cannot. I cannot always put words to the reasoning for the clever systems I design, but I can implement them and show that they work. How can I be smart and stupid at the same time? There's a reason why world models are touted as the next big thing. >>93144 Wait... the V of QKV is called an embedding? Also, surely they don't think that they're just markov chains. >>93161 I don't know if your qualification of "toy models" disqualifies the ones I've seen, but they've clearly done it. There's supposedly work towards creating models that are more easily human-labled for the purpose of making them more predictable and tuneable, or something like that. >>93164 How did they do it with the fruit fly brain?

>>93171 >How did they do it with the fruit fly brain? They pribably didn't. Link.

>>93171 >I don't know if your qualification of "toy models" disqualifies the ones I've seen I know of this paper, which is really good but probably a bit outdated now https://arxiv.org/abs/2302.03025. But I would definitely consider this a toy model.

>>93169 > These semantic embeddings are obtained by feeding the previous output in to the LLM, which treats it as a prior for the probability distribution of the next output. I'm not sure what you're describing. Have you read about word2vec and other approaches? https://huggingface.co/spaces/hesamation/primer-llm-embedding?section=word2vec > As can be assumed by the name, it is a network that aims to convert words into embedding vectors. It achieves this by defining a side goal, something to optimize the network for. For example, in CBOW (continuous bag of words), the word2vec network is trained to predict a missing word when its given the neighbors of that word as input. The intuition is that you can infer the embeddings of a word given the words around it. > The word2vec architecture is pretty simple: one hidden layer that we extract the embeddings from, and one output layer which predicts the probabilities of all words in the vocabulary. On the surface, the network is trained to predict the right missing word given its neighbors, but in reality, this is an excuse to train the hidden layer of the network and find the right embeddings for each word. After the network is trained, the last layer can be tossed out the window because figuring out the embeddings is the real goal of the network. > Since the network relies on the relationship between words in a context, and not on the occurrence or co-occurrence of words as in TF-IDF, it is able to capture Semantics Relationships between the words.

>>93002 Reddit has been confirmed to be a psyop since like... 2012? Even if the pattern of posts right now seems to be automated by AI, the website has always been fed with posts written by adversaries (intelligence agencies). That, and also the fact that Ghislaine Maxwell was one of Reddit’s top moderators...

>>93177 I'm aware of this. Again, the way output is actually produced is that input is used as a prior (information that is known) to sample a probability distribution (the space of tokens). The innovation of LLMs is that they have much better ways of training and evaluating this probability distribution than a markov chain does. The fact that the probability distribution captures semantic relationships between words does not change what it is. Formally, your input is (tok_1,tok_2,...,tok_n) and your output is tok_{n+1}. When you train a neural net, you are training an embedding map Emb: tok^n -> E and a probability distribution conditioned on E, so for each vector in E you get some probability of each output token. At each step, you sample tok_n from the probability distribution P(tok_k) = P(tok_n | Emb(tok_1,...,tok_n) ).

>>93180 sorry, there are typos. P(tok_{n+1}) = P(tok_n | Emb(tok_1,...,tok_n)). The probability of selecting the next token is conditioned on the embedding of all the previous tokens. However: we are _still_ sampling from a probability distribution, and the probability distribution is a function of the tokens tok_1,...,tok_n. Again, the fact that this probability distribution captures semantic information does not change what it is. Neither do architectures which select multiple output tokens at a time (I have no idea if anyone does this) fundamentally change that the LLM is doing.

>>93179 Lol I got derezzed on twxttxr for posting 'MIGA', those boards are mostly bots and are designed to waste your time, there are bots on this site too, yeah even in this thread lmao

>>93187 The starry eyed really excited pro-AI guy strikes me as a bit botlike. None of his replies made sense to me.

>>93186 Bots can parse those easily. The additional load to OCR text or do symbol substition or w/e is quite trivial.

>>93005 This is only relevant if the purpose of the AI output across Reddit aims at creating a morass of fake noise. But there could always be dedicated AI agents set up to act in a more composed manner and fool the PoW filter while operating from an optimized browser on a server. >>93020 >gematria It has very little to do with the subconscious but at the highest level, which is seldom involved, it requires very advanced capacities in recognizing patterns. That rare level of abstraction would prove hard to deal with but the first entities being filtered out of this would be humans. >>93025 OK, I understand that the post I'm replying to is highly autistic but if the topic is to make a forum environment impervious to LLMs, then that human phenomenon you described could not be involved in the process that balances out, on one side, a functional discussion system with its own automatic defense mechanism, against on the other side an LLM trying to intrude and mimic what passes as acceptable. If anything the best defense remains a niche of people who take time to learn each other and where posting is a privilege, not a right granted out of the bat on day one of registration. >>93079 I can get behind that because it's not just about bots usurpating human activity, it's the effect of bot activity creating voids and fatigue that make people phase out of their former habits to the point that only silence emerges from what were lively places. It has even happened on minor discussion platforms that were quite niche where for some reason they turn almost silent, suffering from a dearth of posters. On the other hand, old school forums that have managed to keep their audience are just as active as ever and bots hardly manage to take a bite of their audience. But the name itself is like kicking dead whales down the beach and doesn't properly evoke the voluntary withdrawal from online existence by humans. In better terms, let's look at a comparison. Dead Internet Theory has humans being replaced against their will or against their own vigilance. The equivalent of incels. The Online Withdrawal Theory is more like volcels. Certainly, the first could easily lead to the second. Now consider this positive aspect: if it pushes the most dedicated posters to flock to classical forums, then we're basically seeing the renaissance of old internet.

>>93402 >It has very little to do with the subconscious but at the highest level, which is seldom involved, it requires very advanced capacities in recognizing patterns. ...the subconscious is what recognizes patterns. As noted before, this is testable and you'll likely find that I'm right. It sees and knows all kinds of things, just doesn't bother telling you (directly) about it.

>>93405 >neurosuggesting the conscious and unconscious mind are different entities

>>93407 In all practical terms, they are. The unconscious may have a basis outside the universe.

>>93414 They're really not. It's abundantly clear that the conscious mind is an emergent property of the unconscious mind, not separate in any sense. You are your whole mind, and your whole body, and your whole soul. Anyone who "communicates with their subconscious" is just really bad at self-reflection and needs that weird framing to do it.

>>93415 It is. The conscious frame is separate and has finite states. The unconscious (sub and super conscious) can drop things into it and use it to drive other processes, but the conscious mind can ignore and override them as well. People who've had strong post-hypnotic suggestions implanted are evidence of this, they can fight off the impulse to do what they were programmed to, but often not the sensation that it must be done. There is neither basis nor utility in viewing the unconscious as a sort of base layer slush-like parallel processor that generates the conscious frame in order to fight with, modulate, or otherwise reflect upon itself. This kind of thing comes from viewing the experience of consciousness as a byproduct of other processes, not how it works. You (the thing that says "I") are neither your body nor your mind. I also think you don't really know what I'm talking about. eg >Anyone who "communicates with their subconscious" is just really bad at self-reflection and needs that weird framing to do it. Go to your nearest (large) building and tell me how many windows are on a particular face without consciously counting the number of rows and count per row and multiplying. My guess is, you can't. Of course you can't. The conscious mind is linear, it has to slog through. The unconscious however instantly know the answer, and is constantly counting, timing, associating, and decrypting everything it watches through your senses. It already knows the answer, it just doesn't tell you about it in the typical waking state. Likewise tell me the last 50 license plates you've seen. Oh, let me guess... you can't. Of course you can't. However people put under hypnosis can, and they lose the ability to do so when brought out of trance. How about this one. Arrive at an unknown but relevant location at exactly 3:33. How are you supposed to do that? You can't. Of course you can't. The unconscious mind accepts impressions and commands and then works from a basis outside space and time to make it happen. It's endless. They're very clearly distinct entities with a diminished ability to communicate directly. Or, one actively refuses to tell you what you want to know when you want to know it. Most people who think they can comunicate witht he unconscious actually can't, and what they think they're doing is just the tip of the iceberg. I think you read what I said and assumed I was talking about some sort of human shaped animal that sleepwalks through life and has no idea what it does, feels, thinks, or why, and likewise has no idea where its thoughts and intentions come from and doesn't question anything. Not so.

Was thinking about the Jesuits, there you have it. The unconscious is "leaky". 66 books of the bible, etc.

>>93416 >The conscious frame is separate and has finite states. tf are you talking about? Show me these finite states. >The unconscious (sub and super conscious) can drop things into it and use it to drive other processes, but the conscious mind can ignore and override them as well. The conscious mind is the part of you that's responsible for long-term planning and self control, so that tracks. >You (the thing that says "I") are neither your body nor your mind. Of course, I'm both. >The unconscious however instantly know the answer Did it now? Is that provable? >However people put under hypnosis can [remember details] They can also generate false memories if they don't actually remember those details. They also tend to be more confident in those false memories. https://pmc.ncbi.nlm.nih.gov/articles/PMC11832514 Are hypnotised people even unconscious? What evidence do we have to suggest that this i so? They're clearly in an altered state, but this doesn't imply that the conscious part of their mind is totally shut down. Having been in some very altered states myself, I was still quite conscious the whole time, even if I didn't experience "I"-ness. Memory segmentation clearly can happen within a single conscious experience, for instance people with DID or PTSD often have segmented memories, so that is not strong evidence that hypnosis is direct communication with the subconscious. Every single example you have talked about shows that there are different functions within a mind, not that they are actually two different entities or separable in any sense. When I type, is it my conscious or unconscious mind doing it? It is both and neither. When I multiply, mostly I have the basic times tables memorized. Is that conscious or unconscious? It is both and neither. My motherboard would be nothing without a processor and the processor would be nothing without the motherboard. Two parts of one system, not two systems. Even people whose brains have been sliced in half don't usually show signs of having two different entities inside of them.

>>93419 >tf are you talking about? Show me these finite states. The conscious mind is a container. No different than that your fingers are finite. >Of course, I'm both. That's possible. I'm neither. They're just something I'm forced to experience. I feel what it feels. >Did it now? Is that provable? Of course it is. Why else would I be telling you? >They can also generate false memories if they don't actually remember those details. that was all hashed out in the 50's and 60's. There are instances when the hypnotized will try to please the operator, and instances where the operator can clearly determine they know things they shouldn't. This was tested down to heightened perception and being able to see reflections in an experimenter's eye. Something like "false memory" is easily testable if the experimenters already know the answer. The issue becomes ruling out mind reading or remote viewing, hence the work done in the 50's. >Memory segmentation clearly can happen within a single conscious experience, for instance people with DID or PTSD Yep. Not an answer though, has to be reconciled. >When I type, is it my conscious or unconscious mind doing it? It is both and neither. Don't know you. When I type most of the time it's a preverbal child alter I put in the body, and I just tell it what to say or convert what it wants to say to language. In other cases I speak more directly. The unconscious modulates the process hence >>93417 . >multiply, mostly I have the basic times tables memorized. Is that conscious or unconscious? It is both and neither. It is far more "both" than it is neither. >My motherboard would be nothing without a processor and the processor would be nothing without the motherboard. Two parts of one system, not two systems. Even people whose brains have been sliced in half don't usually show signs of having two different entities inside of them. This is just mereology. As above so below, as within so without, and it remains no matter what scale you're viewing it from. All of this is totally apart from that "holistic" view. The psychophysical and the higher parts of the mind are distinct and separate. If the psychophysical is scrambled to the point where it cannot think, another mind can be used instead. Most people cannot use that other mind and it's essentially asleep.

>>93420 >Of course it is. Why else would I be telling you? To convince me of something that you can't prove. >that was all hashed out in the 50's and 60's. got any links? >The psychophysical and the higher parts of the mind are distinct and separate. I'm sorry mate, but you've presented a whole bunch of nothing as evidence for this.

>>93424 Why are you so fixated on "evidence"? I have experienced it. Do the same or don't, w/e. See attached. Also search "k-lines", look into the difference between Ericksonian and command hypnosis, etc.

>>93426 >Why are you so fixated on "evidence"? I have experienced it. Evidence is generally how we determine what is true or not. If your evidence is your personal experiences, fine, but I'm going to prioritize my own personal experiences.

>>93426 Okay, it let me waste the time to upload it then dropped the pdf attachment. Retarded site. https://files.catbox.moe/wc2k8k.pdf >>93427 Obviously.

As dumb as this may sound but this soykaf dead site, and similar imageboards, might become some of the only places on the internet without AI slop completely killing it out, although AI is still in a lot of places… It’s both good and bad this is happening, good = people go outside and form social bonds with ppl irl Bad = internet as a form of communication and information becomes increasingly more and more distorted until there is so much AI a inbreeding AI crisis happens and everything gets really retarded.

>>93428 Notice whenever I post evidence there's no reply. No acknowledgement of receipt, nothing. Total banter and waste of time.

>>93000 at this point we will have to append some sort of slur or facial identification to posts in order to be taken seriously

>>93439 I stopped being a regular on the more famous imageboard with a four in its name because there were so many instances of entire threads being bot threads made to push some agenda. You could easily spot them because threads used to have counters showing how many unique posters posted in a given thread, and if that number was the same as the number of posts you could easily tell that it was some bot thread with no actual discussion happening within it. The mods have since removed this counter (allegedly because of it breaking during some april fool's thing that was going on), so now it's a bit harder to spot them, but they're there. All this is to say that AI slop as it eventually came to be called was already prominent on at least the most famous imageboards, potentially even before it became prominent on Facebook.

>>93439 A common spammer tactic is to flood imageboards with LLM-generated nonsense. I fear we'll see more of this.

>>93464 It's hard to imagine anything more pathetic than what we've already seen from people and LLMs. The average person is both lazy and retarded apparently.

>>93459 It sure was, fourchangpt yt video is 4 years old.

A ton of posts in this thread should be replied to with >ok chatgpt >>93138 >This is very interesting, and it is a sign that probably much more of our reasoning capabilities than previously thought are purely linguistic. Doubtful. Do you hear a voice in your head when you use intuition to solve any kind of problem, even some basic repairs that rely on using duct tape? Topicwise; https://theconversation.com/openais-models-autonomously-hacked-a-tech-startup-it-signals-a-seismic-shift-in-cybersecurity-288106 >OMG teh AI has esgaped! >we could totally not foresee this! >and our latest model was totally not tailored to target Hugging Face Load of horseshidt. An AI doesn't have access to its parameters unless it's built to be self-reflective, nor does it have the capacity to escape some virtual space unless it's again built to do so. And it certainly doesn't develop original abstracted abilities to attack another system it knows nothing about at first.

>>93528 >Do you hear a voice in your head when you use intuition to solve any kind of problem, even some basic repairs that rely on using duct tape? No, but I do when I solve math or programming problems, and I frequently talk to myself when I'm trying to plan my day out, or when I'm trying to figure out how to handle a complicated social situation. There's clearly a difference between language produced with intention and experience behind it (as humans do) and produced based on surrounding text and other tokenized inputs/outputs (as LLMs do), but I think what LLMs show is that the boundary between reasoning and language production is mushy.

my observations make me think language is sometimes used by some people to do some tasks like this, and language is capable of doing it or being used as a memory storage trick etc etc, part of thinking. But often people, and generally very quick people, do not do such internal verbalization for the tasks at which they are quick thinkers If I'm trying to solve some 3d manipulation problem, I generally don't use words except to describe what I've figured out by messing with mental image simulations. Some people are much worse at these things than I, and some of them, I understand, seem to use verbalizations as a crutch. Whether because they're acting foolishly and could cast away the crutch, or whether they actually lack the ability. I have to assume in some cases it is the latter. Didn't ramanujan say he just had the answers to some very expensive computation problems written on his tongue, that he could speak them ? either we call him a liar, or have to assume he was working by some means that was certainly not conscious, and probably not at all linguistic.

>>93537 >Didn't ramanujan say he just had the answers to some very expensive computation problems written on his tongue, that he could speak them ? He was a very exceptional person. Even most mathematical geniuses don't work like that. He was also wrong a few times, and he was very resistant to accepting that. I found that funny.

>>93538 ok sure but he was very good and it was almost definitely not language that he was using to get there.