Fields Medalist Villani dismissed LLMs as unintelligent as recently as June 2026. But Navier Stokes result has him in shock."Cataclysm" for human mathematicsIYKYK
>>109857303Was that the unconfirmed proof they stole from an actual maths prof?
>>109857303AI can never replace mathematicians like other professions, math guys are Gods, 160 average IQOther STEM subjects are also largely protected
did anyone bother reading the paper or is this like lk 99 where everyone claims its totally real?
some fucking faggot is "in shock" after talking shit, who careshope he jumps off a bridge
>>109857334LK99 is real, and its coming back soon
>literally who twitter screencapall fields
>>109857312The "proof they stole" was also AI generated, by the way.
>>109857303The first post has absolutely nothing to do with the second. He does not contradict himself in any way.Do you guys really not understand LLMs or do you not want to understand?
>mathematician shocked that his life's work is a midwit autistic endeavor that fundamentally requires no real intelligence or understandingI thought they had made peace with this in the mid-20th century but I guess some still needed to learn the lesson
He was probably shaken by the level of plagiarism
>>109857392>>109857621>>109857434>technology board hates technology and doesn't understand it Like pottery
>>109857392You mean the one that was generated after the prof shared all of his work over a year? I got a better question: How will fluid sims improve with the discovery of this formula?
>>109857303Good, math is absolutely miserable.
>>109857691A text prediction machine cannot get smarter. It can just get better as in more precise. It's just not smart.Yes it can help to find patterns in a huge chunk of data.No it cannot come up with some new brilliant idea. That's just impossible. The idea has to be fed in by a human.
>>109857919A pattern recognition machine can find patterns that humans struggle to see, whether it's because of the abundance or complexity of the data. So yes it can find new ideas. Calling it a text prediction machine is pointlessly reductive and reveals that you're too emotional about the subject to think clearly about it.
>>109857919You realize current frontier models can solve every logic puzzle you throw at them, right?
Kek French fag btfo by AI gods. Sacrebleu.
>>109857303>get the car wash example wronganother case of using the free tier lobotomite models and extrapolating from there
>>109858030>>109857992You clearly did not understand my post
>>109857312>proof done with chatbotsSo AI stealing from itself nice one techbroPay artists
>>109858086I think you don't understand your own post, and I suspect you're going to move the goalposts on what constitutes an original idea, as if combining information from other people in a new way to do or explain something previously unknown doesn't count.
>>109858086it's Artificial General Intelligence, not Artificial General Inventivity.
>>109857303i think they said it cost 10 million in tokens? and the prize is like 1 million. so yeah good idea>but as technology advances cost will come down!no actually cost will go up as complexity increases.>i'm in shockhes just a mathematician, so i forgive his naivete at being fooled by this circus.
>>109858199The complexity of what? The tasks we give them to do? Yeah, maybe. Otherwise you can have 27B or even smaller models running on laptops that can beat the best models from 2 years ago.
>>109857303LLMs ARE unintelligent, though, so what is he crying about?
>>109858030>current frontier models can solve every logic puzzle you throw at themProof of this delusion? I've been hearing this from your cult for 2 years straight now and you've been wrong. What's changed? Are they not using transformers anymore?
>>109858229Probably this:>>109857570
>>109858171i'm moving you're mom off my bed now, because i have to sleep
>>109857992>So yes it can find new ideasNo, it can't. A LLM models the distribution of human data it was trained, which gives you an objective metric for novelty: novelty is what doesn't fit in with that distribution.If a LLM could produce novelty, you could train it on its own slop and expect progress. Instead, you get model collapse. If humanity couldn't produce novelty, a LLM trained on all human texts produced before the 20th century would be roughly the same as a model trained on all human texts produced so far, which would be roughly the same as one trained on all that plus every human text produced in the next millennium etc.Simple as that.
>>109857727There is no formula. What was "discovered" (stolen) is that there is no closed form solution because NS sucks. Everyone already knew that since day 1 and algorithms to work around it have been standard since the first days of NS use in practice.
>>109858240Psyho (who you may remember was the human who nearly died coding tirelessly to beat GPT at the AtCoder World programming contest last year, and an expert puzzle solver of course) extensively tested it:https://x.com/FakePsyho/status/2096588951251243219https://x.com/FakePsyho/status/2097318305753174284Or if you prefer something more rigorous, see pic rel from the AtCoder guys.>Are they not using transformers anymore?If you understood what transformers are capable of you wouldn't be asking such a dumb question.This was already predictable 2 years ago:https://arxiv.org/abs/2410.14706v2By the way, it appears ARC-AGI 4 will be all about testing inventivity, so I look forward to seeing what fresh copes you can come up with next year:https://x.com/arcprize/status/2098849962754978152
Agi is very close and will finally reveal all the secrets of the universe
>>109858272>No, it can't.What are hallucinations
>>109857303He was correct, though.
>>109858469AI psychosis status: terminal
>>109858520Not new ideas, mish-mash of unrelated ideas into unrelated contexts. For example, 1+1=3 is not a new idea, it is the old idea "3" with the old idea "1+1" and the old idea "=" all mashed together.
>>109858030Yesterday's frontier model could get the number of Rs in strawberry reliably correct.This is a problem that appeared years ago and keeps coming back.
>>109858553That is what human do too.
>>109858560Nope, still can't get it right, they had to literally hard-code the answer to this. Change the berry name and it keeps failing just as before.
>>109857303who gives a fuck if it's intelligent or not, a tool is either useful or not, that's all that matter
>>109858574Lol no. Go get your AI psychosis checked professionally.
>>109858272>words words wordsAI bharat won, benchod
>>109857392Lol what are you talking about?They had been working on a novel approach for over a year, SOME of that work involved using codex, they did not simply ask an AI to do the work for them and suddenly they were handed an answer.The work they inputed to codex (their own work, not AI done work) was then scraped by Open AI and used to specifically train a highly advanced experimental model with their novel approach to the navier stokes problem. If openAI had not scraped their previous work it would've taken their model 10000x or 1000000x longer since it would've started from the "normal" navier stokes approaches and then slowly branched out into more experimental ones over time, instead it started off using a VERY specific non-standard approach from day 1 that other mathematicians had already been working on specifically for navier stokes.Theu cheated and stole their work.
>>109858546Good, I wanted to be sure that you saw my post and that you knew everyone witnessed you get completely BTFO yet again, I can sleep peacefully now.>>109858582Here's an extra
>>109858560This is as of this very moment with the latest frontier ai model from openai.
>>109858520Hallucinations are where the word predictor gets the context wrong you Bosnian ape lmfao
>>109858607>gpt-6-astra gets it wrong>has to retry by deliberately misspelling the word using the same context to make it pretend to be rightlol, lmao even.
>>109858582who fucking cares dumbass? do you even understand how they are tokenized? it can NEVER do that because it doesnt get as input a string of character, its input is actually a token that gets turned into a high dimensional vector and after that is passed through the network. when the tokenizer sees a word like 'strawberry' it just uses the whole token instead of breaking it down letter by letter. i wondder if it could count numbers correctly if you sent an llm separate tokens for 's' 't' 'r' 'a' 'w' 'b' 'e' 'r' 'r' 'y'. and geniunely who cares? can't you see actual use cases for frontier models outside of counting letters?
>>109858614Hallucinations can be leveraged to find new ideas https://github.com/DivergentAI/dreamGPT
>>109858624lmao at your life dude
>>109858633The reason modern models no longer have trouble with this btw is that they can reliably associate word tokens with the tokens of the letters they're composed of. They still get tripped up sometimes by word sets like months or days of the weeks, since they associate those more strongly with numbers for internal calculations and for programming, pretty interesting stuff really.Also models with vision sidestep the problem entirely when parsing a screenshot of the word since they do see individual characters.
>>109858272>No, it can't. A LLM models the distribution of human data it was trained, which gives you an objective metric for novelty: novelty is what doesn't fit in with that distribution.https://en.wikipedia.org/wiki/AlphaFoldIt beats any human at protein structure prediction. The authors won the Nobel prize.
>>109858199>no actually cost will go up as complexity increases.Gtfo of here with your stupid ass outdated opinion.
>>109858609I've never seen Astra use an emoji unprompted.
>>109858602>The reason modern models no longer have trouble with this btw is that they can reliably associate word tokens with the tokens of the letters they're composed ofThat's because OpenAI trained their models on tasks like counting the number of letters in a word lol. A smarter way to overcome an LLMs inability to read letter would be to make it run a program and use the output of the program to discern the number of letters. But the best and least computationally expensive way to overcome this would be to have a small model before the tokenizer that only exists to sometimes give the model a string of characters instead of a token.
>>109858633So it's stupid.
>>109858757Yeah, in fact agentic models usually do a quick py script for stuff like that since they're well aware of the common pitfalls (they almost never copy long strings of text "by hand" for example.) And I'm reasonably certain modern models have various subprocesses to handle specific task types, it's why they're finally good at "mental" arithmetics for example.
>>109858769yes its stupid in the domain of counting letters but gpt6 astra is likely a better programmer and mathematician than you, if not better it's still faster than you
>>109858769>Hah, that Nobel prize winner sucks at facial recognition and can't even speak English without an accent, what a moron
>>109858712Not an LLM.Trained on a databank of proteins we already knew how they folded.Protein folding is something we already could simulate (folding@home).
>>109858791>And I'm reasonably certain modern models have various subprocesses to handle specific task types, it's why they're finally good at "mental" arithmetics for example.I wonder if frontier models have an MoE expert that was not trained by gradient descent but by regular algorithms to do math or maybe only arithmetic?
>>109858810Unironically yes.
it's tragic. Math was one of the last remnants of 19th century romanticism in our age. I hate maths myself but I find it comforting to know that absolute spergs like Villaini or Perelman could have fun in their little worlds, and feel part of some centuries old lineage.Now they're more useless than fucking chess players ffs
>>109858812Nobody is talking about LLMs dipshit
