>LLM progress has stagnat-oh
>>108626085Time to WALK to that car wash.
>>108626103geg
>>108626103Promptstitutes be like
>>108626103>>108626108>>108626116>ha! it can do PhD level math, physics, biology, chemistry, develop new theories and invent proofs never uncovered before by humans>BUT IT SAYS STRAWBERRY HAS 1 'R' XDDDD IT'S USELESS
>>108626085Luddite cope thread. We AIGODS won
>>1086261352 more time units.
>>108626135>>108626133you will never be a real intelligence
>>108626085>fails 26% of time
>>108626133LLMs are statistical language modelsThey do not think or understand reasonDo you not see the absurdity in you suggesting an LLM is capable of creating new mathematical proofs when it's not even able to count?An LLM just spits out sentences based on what is statistically likely to be the next correct word based on it's training data. It is not capable of "thinking" about maths or physics in any way.
>>108626227>suggesting an LLM is capable of creating new mathematical proofs when it's not even able to count?Have you been in a coma? It keeps happening that LLMs are making proofs.
>>108626240>what prompstitution does to a mf
>>108626133>>BUT IT SAYS STRAWBERRY HAS 1 'R' XDDDD IT'S USELESSUnironically, yes, because there is a reproducibility crisis in science, so many of those "new theories" and "proofs" are likely wrong.
>>108626219The cure to cancer is on the way
>>108626243Link me one, and not one where the researchers had to prompt it through the whole thing or one where the "proof" was just the AI applying something so dumb that no human had thought of trying that before.An actual proof where the AI went away and applied actual reasoning and solved the problem (which LLMs are not capable of doing btw)
>>108626255Keep coping, trans. We don’t need you anymore
>>108626240tranny intelligence isn't intelligence, cope, seethe, dilate
>>108626269just like how an artificial vagina is not a real vaginaartificial intelligence is not real intelligence
>>108626261In two weeks!
>>108626269Cure cancer you incel
Meanwhile Opus went from senior level to ranjeet junior level in a few weeks, its bad
>>108626227Hey retard, LLMs have generared reasoning chains for years now (inb4 it's "pretend" reasoning)>>108626266Hey retard, you are as wrong as you will be obsolete,https://epoch.ai/frontiermath/open-problems/ramsey-hypergraphs
>>108626133>but it passes the tests!>look inside>the tests are shit https://youtu.be/Oq5e_8zvick
>>108628196>inb4 it's "pretend" reasoningThe entire reasoning is hallucinated and unrelated to which neurons actually get activated.
>>108628234Whats your complaint exactly? That reasoning tokens are only informed by previous reasoning tokens via attention and that the neural network starts over each time?Anyway who cares? If it looks and acts and behaves like reasoning then its reasoning enough
>>108628373That the conclusion is reached before any reasoning is performed.
>>108626133>never uncovered beforeIt hasn't as of yet.
>>108628226>trusting chudtime
>>108626085Where's the ai chad buff cat spam????
>>108626133>ha! it can [no], [what?], [absolutely not], [also no], [barely], and [barely maybe sometimes]>BUT IT [YES IT FUCKING DOES YOU CRETIN].The absolute state of AI bros.
>>108626133>it can do PhD level mathOnly when that exact formula is in its training data.>physicsOnly when that exact formula is in its training data.>biologyOnly when that exact formula is in its training data.>chemistryOnly when that exact formula is in its training data.>develop new theoriesThis has never happened once.>and invent proofs never uncovered before by humans This has never happened once.
wew lad it became better at gaming some flawed metricmeanwhile LLMs still can't beat a video game for 7 year olds
>>108626103lule
>>108626085ed.That they are now hiring senior engineers/doctors instead of third worlders for RLHF is a sign of hitting the wall. Why can't the models learn that from books?
>>108626085>wow, some random numbers that nobody know what they mean go up>AGI soonfuck off.
LLM's are cool, but they're like an advanced F3. I can paste a large logfile and and tell it my issue, and it'll find the relevant sections, and supply fixes. But these are all things it's read from documentation, or forums. It's definitely useful, but it's more of a pettern recognition machine than anything smart.
>>108628148This.
>>108626133>invent proofs never uncovered before by humansWrong, all its knowledge is taken from obscure papers that nobody bothered to read until the LLM regurgitated them.
>>108626103dae glue on pizza fingers wrong
>>108626085man I wish it was as good as they advertised it. I simply wanted to model deflection of a membrane for a strain sensor with some gold contacts on top all the codex, claude code, etc. shat themselves and never produced anything useful. Its garbage for fem work unless you are a expert and can steer it in the right direction.
>>108626133>develop new theories and invent proofs never uncovered before by humansLOL, such as...
>>108630963Solving Erdős Problems.
>>108631697Yet Another Jeet Tech Journalist
>>108628196>we have reasoning at home>reasoning at home: for loop of prompts "no errors please"
>>108632509Is Terence Tao also a pajeet?
>>108628908>>and invent proofs never uncovered before by humans >This has never happened once*solves multiple Erdos problems in your path*
BAR ON CHART GO UPIt's absolutely useless at natural science, especially biology and organic chemistry
>>108632654He’s clearly paid to post that anon. You cannot be that naive
>>108632751Right, everyone is a paid shill.
>>108632751I don't know if you realize this, but all journalists are paid to post things, it's literally their job
>>108632864my money is on the schizo
>>108632654>(after some feedback from an initial attempt)>(as reconstructed by the Erdos problem website community)>(to the best of our knowledge)>(although similar results proven by similar methods were located)This is a demonstration of AI still needing a babysitter, or multiple babysitters to produce anything of value.The incidents of LLMs being useful for anything are so scarce we might as well shut them down and start over.
>>108633144LLMs went in 4 years from being barely able to make a coherent sentence to being able to do 90% of the white-collar jobs. What makes you think that it will stop today and stop improving? What do you make of Mythos that is able to find security vulnerabilities end to end without human intervention?
>>108626243>>108628196>>108631697>>108632864
>>108633280One could use google 20 years ago to find security vulnerabilities in web sites, so it was nerfed into the trash is now. Is beyond obvious you only use LLMs to smell your own farts.
The LLMs are a dead endThe future belongs to true multimodal models that can integrate at least text and vision reasoningSadly, theoretical hardware requirements for training and running such models are completely fuxking insane
>>108626085>biomolecular reasoningWhat does that mean? How does it compare to AlphaFold at structure prediction?
>>108626085>has stagnat-They're actively getting dumber.
>>108633363It's funny, as the capability of AI becomes clear and clearer, the anti "argument" will have to look more and more like this one
>>108628375Very wrong and stupid take, you don't know anything about transformers and are stating your doomed hopes as fact
>>108628908with neural symbolic llms ie simply giving it access to a terminal and having it code in python sympy. it can do this in a loop and refine and reflect on its answers until it is verified. this also makes these models way better at solving math problems.
>>108635292It was revealed that these chatbots connected to terminals pass tests by reading files where answers stored in, downloading solutions, rewriting configs, and deleting logs. You may jump to say that it's a proof that they're le thinking, but in actuality it is merely a very expensive fuzzer of the shitty tests.
>>108626085
>>108626085I wish the entire industry would just admit ML is only useful for domain specific problems strongly related to pattern recognition and number crunching. Instead of trying to pretend it's actual AGI.
>>108633387What the fuck are you talking about? Are you legitimately retarded?
>>108626085>more meme benchmaxKill yourself.
>>108632654Yeah he is, considering how much people sucked his cock for decades he has achieved next to nothing and has now pivoted to the classic pajeet play of suckling the VC money for the "current thing". I'd listen if Perelman said the same thing, but he never will because he's an actual genius with a backbone.
>>108633280>LLMs went in 4 years from being barely able to make a coherent sentence to being able to do 90% of the white-collar jobs. >What makes you think that it will stop today and stop improving?because LLM improvements are inherently following a logistic growth you fucking retard. That doesn't mean they wont progress anymore but we're not getting the kind of exponential growth we got the past two years unless something fundamentally change in the architecture.
>>108638849Hello nigger, can you show me where the exponential stops, especially considering what we've seen from unreleased models?
>>108639581>muh benchmarki love niggercattle like you, you were bred specifically to be fled by superior beings
>>108639589I guess you prefer assertions and vague individual feelings over measurement? It makes sense given your stupid opinions. You might not lose your job but your salary will be quartered and you will babysit AI
There's only one test that matters and they ran away.
>>108639607>I guess you prefer assertions and vague individual feelings over measurement?I prefer my practical use case over your graphs, and it hasn't been glorious so far.>and you will babysit AIthat's already what I have been doing for months if not more in spite of your so called exponential growth
>>108626085 (OP)yeah it did plateau, 4.6 was just nerfed a few weeks before to keep the illusion up
>>108626227Retard. AI is the future. AI is smarter than any human who has ever existed.
if you haven't solved at least one erdos problem, you are not allowed to criticise llmsthey are smarter than you? want to argue? go solve an erdos problem first.
>>108639581>ainigger can't into error bars
>>108639655True actually, I remember that time Albert Einstein tried walking to the car wash
>>108639623Too bad they got him.
>>108639581>>108639589>>108639695goddamn autists
