(untitled)

This post is just a reminder to you, lurking friend. The same people who fearmonger about glorified markov chains taking over the world are ex-cryptobros, ex-NFT advocates, ex-web3 shi/lls, and zoomies. They are literal dumbasses who don't know how either of this works. Brought to you by TikTok, Twitter and Reddit (trademark). They fear magic because it is something they don't understand, and because it's magic to them it's an equivalent of a god. Internet is indeed dead, but only because people like OP are allowed to bring there piles of garbage from 4chan.

>people like OP So, yourself?

>>68020 pundit here, op is likely a virgo, based on their description of magic, one could speculate they would use ai primarily to activate the left hemisphere of their brain and explore like paleolithic man 300,000 or so years ago. not saying that's a bad thing, but it could account for OP referring to OP in the OP itself. Keep in mind though, blah blah blibli, many contexts or bullshit and many opinions and we still don't really know and I'm just a Human Intelligence so don't ask me. Overall, ye

Ye it's amusing to see those threads (2 or even 3 in the same day) which seem to be written by a 16 year old, and to see the thread kept alive by a bunch of anons who seem to entertain the idea, or in any case reply instead of just ignoring the old and tired premise.

>>68022 >written by a 16 year old Let's be honest, most people on this site have never watched lain, have never coded anything, and are 16 years old.

>>68022 But actually the even scarier thought is that these people are not 16 but instead they are fully grown 25+ year olds who have already wasted all their learning years. Pretty dreadful, right? And the thing is that you don't even need a language model to fool this kind of person. They've already been fooled by the flood of normal chatbots on social media like twitter, facebook, 4chan, etc for years now. The dystopia they imagine where everyone online is an undetectable bot already exists for them, they just don't know it. It's enough to make me wonder sometimes if things wouldn't be better if they really did get replaced with language models.

>>68023 >have never coded anything Not required to be in this chan >are 16 years old The fate of every community that is not gatekeeping >have never watched lain And this is the real problem. There should have some kind of requirement and should be tested. Gatekeeping is the only way to keep any kind of community healthy.

>>68026 >There should have some kind of requirement and should be tested. It's fine. the newbies here don't know about our super duper secret hiding spot on here :)

>>68026 That's a great idea. The first test should be a coding test.

>>68037 >The first test should be a coding test. ChatGPT can easily solve it, though. At least any coding test can easily be done with an AI.

>>68038 >ChatGPT can easily solve it, though This is fine, if you're dumber than chatgpt you should be kept out.

30+ ex-web3 dev SEL enjoyer who thinks LLMs are impressive here. Fight me.

>>68044 Studied NLP and machine learning in college, here to fight you. It's amazing from the perspective of an NLP. ..but it's really dumb and not an "AI". When we say "AI" we presume it's "intelligent" and that's just not possible kek. It also can't answer basic questions. Tried using it to help me write my thesis, just asked it to simply explain something, like you would read in a dictionary. And well, it was wrong over 90% of the time. It's not even as good as a dictionary. But it can talk pretty!

>>68048 I didn't take any machine learning classes when I was in college, but I did do some NLP, and later spent several years trying to build practical applications on top of that theory. I failed. It might be arrogant to say so but I believe my failures there are the failures of the field; grammar simply cannot model human language on a deeper level. The idea that our thoughts can be accurately described in terms of structured logic is wrong. You generate sentences based on any such structure and don't impose a narrow format and you get stilted garbage that any native speaker will rightly recognize as gibberish, even if it makes sense to someone who is versed in the logic underlying it. LLMs do actually everything I was trying to do and more that I had concluded was impossible. You can pooh pooh it all you want but it delivers where NLP never could and anyone can see it with five minutes of objective observation. I feel super bad for Noam Chomsky and everybody else who put their lives into that soykaf, and I don't begrudge all the cope, but it's fuarrrking over, yall got steamrolled.

>>68050 Also, since I neglected to fully read this sentence >It's amazing from the perspective of an NLP. And I haven't really contradicted you I figure I should also go into potential for general intelligence, which I do think it has even though my main reason for finding it impressive is that it has handily solved age old CS problems. Basically, despite observable limitations, those limitations keep being surpassed. There is no good theoretical reason further progress is impossible, because the nature of intelligence is not confined to our ideas about human minds, it's just the mechanical capacity for doing arbitrary things. This tech keeps gaining more of that capacity.

>>68050 >thoughts cannot be accurately described in terms of structured logic I'm not sure I agree with this. A very rigid and grammatical style isn't neccisarily limiting and ultimately (good) ideas have to be logically expressible. Part of the problem with non deeplearning NLP systems is that they are bad at impersonation, not that they are bad at ideas. Both are bad at ideas but these modern systems are very good at impersonation, which makes them seem more human (and therefore maybe plausibly better at ideas). A lot of communication with humans is trying to riddle out what someone means from his words and so its possible these will be useful (some say they already might be) as creativity aids, listen to enough nearly right supermonkeyonakeyboard and maybe you get the inspiration for a sonnet, even if the monkey's no bard.

>>68056 >ultimately (good) ideas have to be logically expressible IMO this isn't true. One may contort their mind to conform to a defined standard, and there are useful reasons to do so (like being a programmer), but this is a distillation and not the space where thinking happens, the essence of which I believe is not discreet logical concepts. A logically inexpressible idea can still be a powerful one. The failure of non deeplearning NLP isn't just impersonation, it's a failure to be able to engage with the layer of intuitive thought underneath rationality, critically important to thought and communication. I believe these new systems can somewhat engage with that layer, because at least they don't immediately and obviously fall apart when encountering its expression, and that's massive potential progress towards thinking machines.

>>68057 >because at least they don't immediately and obviously fall apart when encountering its expression, and that's massive potential progress towards thinking machines. Not that anon, but are you really sure that's not just because they have been fed so much data that they appear to do that while there's not "much" going on behind the scenes? Just for you I tried chatGPT and thought of something really simple but unlikely to be ever said by anyone. Anyone should get this by intution or by logical thinking. Yet it obviously failed this task. My question is, if these models really are thinking in any meaningful way, why do they need so much data? GPT3 apparently was trained on 570GB of text. Now I suggest you count up all the books or text files you have in on your computer and multiply the sum of that to be equal to what you roughly would have read in your life. I don't think I would go to 1gb just by text alone. I am guessing the argument will be that it's easier for humans because we already are born for language and what not. But no human is born with the ability to make images outside of symbols. Most people, unless they take drawing seriously, draw these weird flat symbols. Yet you still don't need to take anywhere near as much data to learn draw well. No matter that it was kinda impossible since phyiscal art materials in the past were unreliable or took long time to dry, so you had to limit yourself in what you could make. People also draw things that can obviously be wildly different than what was they learned. Something that these models don't do, like my attached pic shows. Tell me how the apparent intelligent behaviour is not just caused by the fact that they have whole libraries and big chunks of internet scanned into them?

>>68058 I wouldn't consider "thinking in any meaningful way" to be dependent on what is going on behind the scenes. If a system can objectively do something, that's what it can do, regardless of how. >trained on 570GB of text. Now I suggest you count up all the books or text files If you count all our sensory data it's way more than 570GB, and I think we need it. People deprived of normal stimulation from birth end up with permanent, severe brain damage. Though clearly there are differences in how we learn and how a machine learning program learns; we evolved in an environment of limited access to information, and these models were built in an environment where the easiest way to progress was throwing more hardware and the entire internet at them and seeing what happens. So what? >Anyone should get this by intution or by logical thinking. Yet it obviously failed this task. ChatGPT is not a human. Its capabilities are not the same as one. That doesn't prove LLMs are a dead end. There are many such tests people applied to earlier models that now pass. I don't think there's good reason to think this particular test will never be passed. What you seem to be suggesting is that it is a dead end because what it can do is no more than the sum of its data. That is at least a coherent idea, but I don't see a reason to think it's necessarily true, and most of the time I see it the backing seems to be mostly a bias towards human exceptionalism and against accepting that the world is about to be dramatically transformed. What it does with the data is provably more than flat regurgitation. It's not just copy pasting, there is dynamic engagement and processing of what we give it. To me that suggests it may be able to go further despite the finite nature of the largest possible dataset. Why not? How can you know the limits?

>>68063 >If you count all our sensory data it's way more than 570GB, and I think we need it. People deprived of normal stimulation from birth end up with permanent, severe brain damage Well because it's abuse. People who get kicked every day too end up with certain issues but it's not because they lack stimulation. When it comes to text, other data besides reading a page has no bearing on it. You might make the case with visual recognition as seeing things from different angles obviously helps, but as text is just symbols on 2d surface the extra data of the paper or whatever object the text is written on is meaningless. Actually like I said before, if the whole "unconcious thought" or "sensory data" explanation would make sense, people just wouldn't draw stick figures when they pick up drawing. And that's a very easy to understand example, but it applies to most cases. From what I read if you don't focus on an activity and just try to do it with "autopilot", your head doesn't learn it that well. You learn it slightly or do it at the same level as before, just more automated. >we evolved in an environment of limited access to information, and these models were built in an environment where the easiest way to progress was throwing more hardware and the entire internet at them and seeing what happens. So what? My guess is because it's kinda a scam to call it AI. Since a more honest approach would what it is would be saying it's just a pattern finder of text. That it appears human-like is not because it has understanding of anything, as once you put some question that's difficult to solve by just finding a pattern in text it fails like you see on what I attached before. It seems human because you literally are communciating with people through something which could be considered a "filter". >There are many such tests people applied to earlier models that now pass. Well these things can do pretty much anything provided you have enough data to feed them. Which isn't surprising as it's a statistical machine. You can correlate anything with anything. My point is that seeing this as a machine figuring something out is misleading. >mostly a bias towards human exceptionalism and against accepting that the world is about to be dramatically transformed. I am kinda tired hearing this. As if the points I'm making are based on feeling and shouldn't be considred. Can't I just as well say there's a billion dolar companies which have a very big interest into making sure people believe these things actually think? Either to get more investors on board, or to create bot if you are lonely to have a "companion", or for people to believe the "AI has won" so they don't need to think about job security and be fine with giving away for free. After all, where data is free the person who wins is the one with the most powerful computer. >Why not? How can you know the limits? One reason would be there will be a point once you run out of useful data to feed it. These datasets are already crazy huge for what they try to do, and at some point you will simply just make it have stronger bias to what's common to find than actually learning anything. Second reason would be that you'll see more and more AI stuff, which will be fed back to these datasets and used for training an AI. So it's kinda eating it's own tail too.

>>68069 >My guess is because it's kinda a scam to call it AI. Since a more honest approach would what it is would be saying it's just a pattern finder of text. This is a very succinct way of putting it and I completely agree with you. Especially after my experience with it and it being right 1 out of 60 prompts for the most basic questions. The media fantasies of this thing "aiding research" are literally fantasies. It wasted more of my time than it saved. I eventually completely gave up on it and just went to a search engine.

>>68069 >Well because it's abuse. People who get kicked every day too end up with certain issues but it's not because they lack stimulation. It's abuse because we need data to exist as people. Also it's not the same thing as violence, the people I'm talking about end up close to brain dead. Similar idea with solitary confinement of prisoners; severe and permanent psychological damage, though obviously not as severe since it's not in the developmental phase. Large volumes of sensory information is a basic human need and a requirement to function, this should not be controversial. That doesn't prove that our cognition is somehow unreal. >From what I read if you don't focus on an activity and just try to do it with "autopilot", your head doesn't learn it that well What 'focus' means here isn't all that quantifiable. I would point to the need to actually practice a thing to be any good at it. You don't get good at writing by studying grammar, what's important is reading books and writing yourself. You don't get good at art by just studying theory, you have to actually spend time drawing and observing. Tactile feedback matters, the process matters. >I am kinda tired hearing this. As if the points I'm making are based on feeling and shouldn't be considred. Consider the part of my statement right before the one you quoted. I'm not saying bias makes you wrong, but for there to be a good argument there should be elaboration beyond the assumption of "needs data -> dead end for instrumental capacity", and all the rhetoric I see about semantics like "scam to call it AI", gatekeeping "thinking" in terms of how humans do it, and ad-hominem stuff is irrelevant noise that contributes nothing to that idea. The actually important question here is how powerful this tech will become, and that deserves focus and questioning assumptions. >One reason would be there will be a point once you run out of useful data to feed it I don't see a strong argument or evidence that it works this way, that capability ceiling is a linear function of dataset size. I think this is just a guess. My guess is that it doesn't work this way; LLMs have shown some ability to extrapolate their data into ability to process what we give it logically and create somewhat novel outputs, which I think is basically the same as what humans do, and we can produce tools that exceed our natural born limits.