I'mHonestly curious does /g/ agree with Silicon Valley that thinks if we keep scaling LLMs more emergent properties will pop up or with the academia that thinks machines need a world model
>>109095465SV is smoking hopium because it's the path of least resistance and can funnel a shitload of money to pump up stocks that way. that's it. that's the whole thing. you can pack it up and go home now. did you really expect anything different?
>>109095476What about fable? Plus the fact LLMs have started to show more complex behaviors
>>109095486do you have a single clue what you're talking about or are you just fishing? this is a little bit of a nuanced topic. basically there's an infinite number of emergent properties that can and fewer that will show up. this will perpetuate until people are eking out the last frontiers of novel behaviors for LLMs. that's not a substitute for AGI is why I'm asking you this.
>>109095499And you know this how? Are you some expert?
>>109095509do you have any other source to spoonfeed you esoteric information? anyways, I digress. the currently useful emergent properties that exist are coalescent networks using a cascading number of agents, like Gas Town.
>>109095465at least you should keep scaling until it hit the wall (not yet)researching the alternatives right now is just gambling
>>109095515Okay u know stuff but I don't get why LLMs can't be AgI? If we scale enough can't they be good enough to built models
>>109095520okay if you stack a shitload of copper into a tesla coil you can make big ass giant fucking arcs that scare the piss out of people except it's not actual lightning because it's not direct current. the alternative is stacking a shitload of car batteries except nobody is going to pay you to do that unless you're styropyro and he's a chemist, not an electrical engineer, for example.a closer analogue is that we thought the same thing about markov chains. I've been meaning to finish reading this but this is how LLMs are different from markov chains in essence. https://arxiv.org/pdf/1706.03762the point is that there is little to no evidence outside of the regular speculation that the logical conclusion of LLMs (for this go-round) is going to result in AGI, because we thought the same thing about markov chains before that, and other applications before that, like symbolic systems. there appears now to be a way towards AGI and that LLMs will likely either/or be a part of that and be an essential contributor to its development.
>>109095465I was under the impression agents were like specific narrower band of AIs?
>>109095542So what's missing part is it world model lee cun discuses? And how would you get multiple too coordinate
>>109095555imagine AGI exists. it would predict the output of an AI agent consistently without knowing the implementation of the agent, sort of like how with enough encrypted data a key can be derived except that it takes like petabytes or some shit of a good encryption algorithm to do that. most agents are just a pre-trained LLM and everyone calls that AI as a colloquialism.
>>109095557Gas Town. otherwise, as human intelligence understands the nuances of one's own intellect, such as the ability to introspect or retrospect, that you and I for instance can learn a new job just by reading about it. if I gave you a manual for a thing and you had no idea what it was, you would be able to go and ask someone else when you get stuck. AI as we know it (i.e. LLMs) can do that, except there's no concrete model of the world in the agents, and their intelligence is basically by proxy, by facsimile, or is a facade entirely because there's no conscious perception involved.
>>109095567But AGI doesn't exist.
>>109095587But consciousness as I understand is the the brain being able to meld different parts into one unified world. I mean we aren't constantly seeing world but hallucinating it in a controlled manner while being updated when we get it wrong. If you didn't catch the double the then you get point im making. But that's still simplified model the purpose of brain isn't to study math but to keep organism alive in a dynamic world long enough to reproduce. If brain was truly intelligent machine we wouldn't have sex or become obese intelligence is mere side circus
>>109095567is human AGI?
>>109095616there are humans who don't have sex or become obese. either mode of being an organism or living an ascetic life are anecdotal. basically every intelligent organism hallucinates reality, and that is a separate thing from a world model, because reality is almost entirely subjective. there is no objective reality to human intelligence largely outside of consensus. you might've heard the quote that history is written by the victor. training data for an LLM is not so different because most of the mundane reality of day-to-day life is implicit and changes between generations. you don't think that LLMs require the prerequisite resources to be managed not unlike subconsciousness? typically an LLM is not going to be considering its host operating system nor the underlying electrical system powering itself. a couple of great examples are that animal intelligence engages in play, sleep, and dreaming.>>109095631human intelligence is not artificial. AGI can be simulated by human intelligence. >>109095596
>>109095656is human GI?why are you so annoying nigger?
>>109095656>>109095616as a corollary, consider someone who has a hard day having several drinks in the evening. that person erodes the nervous system (because alcohol is a neurotoxin) and wakes up with a fuzzy recollection of a memory. AI/AGI aren't going to do that because neither are capable of becoming stressed and the emotions that go with that. animals can do this too and even though they might not have a baby bottle full of everclear on the nightstand, that PTSD can even be diagnosed in non-human animals.right now, LLMs "hallucinate" like a drunk engages in babbling like >>109095676yes. do you even have an inkling of an idea what the fuck you're reading?
>>109095465I don't understand the planning on the pic. Can't you just implement that now with some adversarial LLM, so that instead of blindly doing next token action, it tells the adversarty what it plans to do, and the adversary RP says "if you do this, that should happen"? Done, predicted outcome.
>>109095679>imagine AGI exists. it would predict the output of an AI agent consistently without knowing the implementation of the agent, sort of like how with enough encrypted data a key can be derived except that it takes like petabytes or some shit of a good encryption algorithm to do that. most agents are just a pre-trained LLM and everyone calls that AI as a colloquialism.can human do this?
>>109095695not enough time in the day.
>>109095465>>109095555Answer me this: when a superintelligence knowledge increases exponentially how does it plan to index/store/retrieve that information.
>>109095465>that thinks if we keep scaling LLMs more emergent properties will pop upThis is not serious science. It's just a justification for MORE ELECTRICITY and MORE DATA CENTERS and MORE CAPEX so they can inflate the financial bubble more in the short term.
We're so behind on studying minds we don't even properly know what current AIs are missing that makes fucking worms mog it with advanced abilities like adaptive forgetting or generalization. LeCunn comparing LLMs to cats proves he doesn't know or care what he's fucking talking about, their intelligence is not comparable to anything natural, maybe slime molds.On the other hand we don't need AI to be human level to do human level tasks. It'd be more than enough to have something like a swarm of tiny AIs adapting to the task at hand like a virtual anthill and they could replace proverbial 99% of time of 99% jobs. Pretending that LLM agents are anywhere near close to that functionality is a self-inflicted modern delusion.Most of talk about AI is midwits hanging onto consciousness as if it meant something and mentioning Turing test without understanding that it was meant to shut the fruitless "conciousness" conversations down - it doesn't matter if the computer conscious, if it's hard to tell if it is, that means it's doing it's job.Sagan was wrong and sci-fi was a mistake.
>>109095465My opinion on this matter requires an excessive payment, chinks must be given the bill before interaction with them.
>>109095465You have stolen my well formulated ideas multiple times. I demand immense payment.
>>109095465I think we need to send more billions to Mr Altman
The real world can supply unlimited tasks unseen by any single pre-trained model, so online learning is a basic requirement of intelligence. You can't scale your way to that.>>109096408>maybe slime molds
