Without wishing to cause a shitfit, can someone explain to a retard like me in simple terms how these modern LLMs aren't actually AI?
>>107066722AI stands for Artificial Intelligence LLMs are not intelligent
>>107066734/thread
>>107066722they're text completion softwares based on probabilities.
>>107066722"AI" is a marketing term that doesn't have a fixed definition. So anything can be AI.
>>107066722You know how your phone tries to autocomplete words? "AI" tries to autocomplete answers to questions based upon reddit replies and wikipedia entries. That'll be 3 trillion dollars please
>>107066722This. >>107066795 Ask an LLM to guess a number between 0 and 9. Nine times out ten it will be either, 7 or 3. And very likely in this sequence too.>but why?Because these two numbers are the most present in the dataset. And the statistical averages is what LLMs spit out to you.
>>107066722LLMs generate text that seems to make sense, which can go a long way in many applications and can simulate many aspects of intelligence.But in the end, they can only generate text based on their training data and can't actually make "new" logical connections, and since every token they produce in their current "memory" becomes more and more computationally expensive, they don't have functional longer term memories and intuitions anyways.There are many hacks and mitigations to solve some of the weaknesses of LLMs, and some of those can definitely produce a very useful product and can continue to do so, but we'll probably need a new architecture or paradigm to really make a new breakthrough.
>>107066722it's not "ai" but it's good enough to be useful and replace coders/artists so now they cope with semantics.
>>107066722technically they're not because real intelligence is more than being able to put words together. BUT a lot of people just like to say that to minimize how good it actually is, like "it's just predicting the next word bro" or "still a nothingburger, just a dumb machine", "ones and zeros" etc. And in reality, for all intents and purposes it's already better than a lot of people at their own job and the population is only getting dumber, so i don't think it actually matters whether we call it intelligence or not.
my question is, are LLMs a step towards general AGI or are they just something completely different? Human intelligence improved by learning so my sense is that they are an important piece towards AGI, but what is the missing piece(s)?
>>107066866Humanity has a pretty tenuous grasp on what "intelligence" even is. Don't count on us being able to synthesize it.
>>107066866they're a step toward fooling you into thinking its AGI
>>107066722
>>107066866>are LLMs a step towards general AGI no>Human intelligence improved by learning so my sense is that they are an important piece towards AGI,LLMs do not learn
>>107066812>anything can be AI
>>107066722It's just infinite goal posting shifting because "AI" is an extremely nonspecific term that can technically be used to describe literally anything a computer does.Anyone who tries to tell you that "AI" has some specific meaning that LLMs don't fall under is a moron who doesn't know what the hell they're talking about.
ai is a poorly defined term. llms are just a math function that predicts the next token based on the previous tokens. at any rate, it would appear they are capable of producing coherent text. if that's how you want to define ai, go ahead. I had an Nintendo 64 game that had AI opponents. my tv has AI color tuning processors or some shit. people seem to use the term to mean anything they want.
>>107066734Define "intelligent"
>>107066722It is AI.AI is just some form of intelligence.Enemies in video games are considered AI even if all they were was a state table with actions like idle_animation, run to player, attack.Things have been labeled as AI since the 70s with LISP and Expert Systems.
>>107067495You masturbate to weird gay shit.
>>107066734AI has been historically used just as an umbrella term for behavior that the computer does without human intervention. You could program a tic-tac-toe AI to play perfectly using only database. It just creates an illusion of intelligence. It's still AI.
>>107066722today I took like 2 code challenges I asked AI to give me, both times I sent my complete version for review to the AI, both times AI told me my version was wrong, and both times the "correct" version that AI gave me was wrong and mine was correct
>>107067611except this term (AGI) has been hijacked by brainlets into meaning artificial conciense
>>107067274>they're totally intelligent, if you simply use a sufficiently weak and retarded definition
you can memorize ieee rfcs like a true autist and talk to your computer in tcp and your computer can talk back to you in tcp, but that does not make you a router.
>>107067746You don't have any definition for intelligence.
>>107068599Nobody has a good one, but some people like to pretend they do to suit their argument.
>>107067274**Intelligent** describes someone or something that shows the ability to learn, understand, and think clearly and effectively.It refers to possessing or demonstrating qualities such as:* **Reasoning ability** – thinking logically and making sound judgments.* **Understanding** – grasping complex ideas or relationships.* **Learning capacity** – acquiring and applying knowledge or skills.* **Adaptability** – adjusting to new circumstances or solving unfamiliar problems.In essence, an *intelligent* person or being shows mental sharpness, insight, and the capacity to use knowledge thoughtfully and effectively.
>>107067793but le Chinese room...
>>107066722They cant do math because they cant do logic theyre just copying text from their training data. Just give a complex many decimal math problem it hasn't seen b4 and watch it fail.If you train a computer to judge images as red or blue based on each pixel hue andit could do it to every input would it be intelligent? Obv not, llm does this just with words (and other stuff) and a really complex algo. But its output is just based on data input and rote response. Its pretty easy to reach its data limitations, talk about a complex system and ask it what it lacks in capabilities, it cant handle the negation (my ex i was trying to.get it to help analyze some malware)
>>107068688Also if we did have an agi it would use llm as memory (as in brain, long term) and really only make it a metacognitive llm which would be only a marginal improvement and probably not worth the extra $$
The red pill is that they are actually a fraction of collective human intelligence.
>>107066828jesus christ
>>107068688>They cant do math because they cant do logic theyre just copying text from their training data.As is typical with dismissals of LLMs, this description is reductionist to the point of being incorrect.>Just give a complex many decimal math problem it hasn't seen b4 and watch it fail.And my car is shit because it sucks at toasting bread. I don't get this dichotomy where unless these systems are perfect and can solve the Riemann Hypothesis while simultaneously giving you the best blowjob of your life, they're useless.
It's not AI yet, but if you give Sam Altman 1 trillion dollars he will make it happen for sure.
>>107070754LLMs are just text completion turned up to 11, so advanced it's hard to tell it's text completion, but it still is when you look a little closer
What makes the human brain so different from these text predictors anyway?Whenver you "think" of something it's essentially the same and looking through a vector database in your brain to find the topic and use related words in order to explain it.
>>107070965The difference is that you can’t fill a human mind, the sheer number of connections allow for nearly unlimited data storage potential and I suspect one, if not several, properties of intelligence emerge from that.
>>107070965What makes the human mind so different is that it is a receiver of consciousness that translates the environment around us into perception.
coal thread
>>107067274We can't even define human intelligence and we barely understand the brain, do you think we're going to reach AGI by 2030? Spoilers we won't and all the SV tech bros are full of shit
>>107070952I feel like this bot didn't like the game.
>>107066722technically, many technologies colloquially can be called "ai", including chess bots from the 80s. but LLMs are just autocomplete 2.0 + magic 8-ball on steroids
>>107071632>do you think we're going to reach AGI by 2030?
>>107066722They absolutely do constitute intelligence by any reasonable definition. Five years ago, people would have considered any degree of autonomous problem solving or beating half a pokemon game at a rate far above random noise to be intelligence. It's funny how words like "intelligence" and "art" have been going through extremely specific redefinitions lately designed explicitly to exclude AI, especially when previously these same people would have told you that there can be no objective fine-line criteria.
>>107066795But isn't the point of the "Artificial" part of Artificial Intelligence to be able to seemingly replicate human behavior and solve problems without actually doing so like a human or having actual intelligence? Wasn't the recurring theme in every single science fiction story that robots are actually super-advanced calculators that give the impression of thinking but aren't?
>>107073356>Wasn't the recurring theme in every single science fiction story that robots are actually super-advanced calculators that give the impression of thinking but aren't?Such as? Data from Star Trek has agency, the Architect from the Matrix redesigned the simulation from the ground up in the presense of new data about human psychology. Soma has multiple models of AI behaviour in one game.
>>107073356https://www.youtube.com/watch?v=ol2WP0hc0NY
>>107073416I was thinking of I Robot. But even in the cases you mentioned, I'm not talking about AI capabilities being advanced enough, but rather the difference between real intelligence and artificial intelligence. Generally, in fiction they focus on when we can consider an AI to be a “person” or not, which is another matter entirely, but before being considered a person, a model advanced enough to communicate with humans (regardless of it being able to understand the words or not) and act dynamically on its own without needing to follow commands for each step of its function (regardless of it having any initial "function" or not) would already be considered AI.
It will be AI when it starts analyzing its own answers for correctness instead of telling you to drink bleach then saying "good catch - you're right, drinking bleach is fatal" after you drink a gallon and tell it you're unwell.When I first started using Linux it made me chown my entire root directory to some other user and broke the entire system so I had to drive 3 hours to where the server was to fix it then drive back.
