Next-world prediction machine is the moon-landing moment for AI.

SAAAAAAAAAAAAAAAAAR!!!!!!

>>106197903 (OP)
>moon landing moment
They going to fake the entire thing then spend the next 5 decades pretending they didn't film it in a studio?

>>106197903 (OP)
>pre-college prodigies
so it’s about as good at writing proofs as a high schooler? Am I reading that right?

>>106197903 (OP)
>a token guesser when given the answers guessed the answers
wow

>>106197903 (OP)
You work this upon yourself by following these witless cheerleaders on twitter. You deserve all the grief you get.

>>106197903 (OP)
LLMs can't solve unsolved math problems they are dumb

AI is an industry of frauds hyping up calculators. it's so fucking dumb.

>>106197903 (OP)
careful working is proof of deliberate fuckery, what a faggot

>>106197903 (OP)
>It's hard to overstate the significance of this
But you make it look so easy lying fag

Presented without further comment.

>>106198095
>calculators
They wish they could function as the world's least efficient calculator. Doesn't matter if it's 2+2 or quadratics, LLMs get the answer by a combination of luck and repetition. We trust calculators because they always get the same answer to the same input. LLMs do not and even when it's right 90% of the time, having to check it in case it's the 10% takes so fucking long that it defeats the point.
Kinda the same reason quantum computing has gone nowhere in years. It returns an answer, but there's no way to verify the answer without doing it on normal computers anyway. Why? Because even when you think you've just done a variant of something that's worked every other time, well sometimes it doesn't work and how are you going to know if this was that time or not? Makes the whole thing worthless.

>>106197903 (OP)
Let me guess, the proofs written by those pre-college (see: high school) "prodigies" have been scraped and fed into the LLMs prior to this challenge?

>>106197903 (OP)
>moon landing moment
>did it once, cant ever do it again
yep thats AI

>>106198754
What I don't understand is why they always talk about "agentic" AI but they can't somehow just make their model, which already uses python, to just run the 2+2 in the instance. Or maybe its doing that sometimes, but clearly not always. Or these letter counting tasks. This is all stuff you can easily calculate with a programming language, yet this "agentic" AI is literally just sifting through its training set to find the answer to how many b's are in blueberry and then somehow givng a probabiblistic answer on something that has a definitive answer

>>106197903 (OP)
A sophisticated enough next word predictor can do anything.

>>106198068
>LLMs can't solve unsolved math problems they are dumb

The future looks bleak

>>106198754
Right. If you have to check why didn't you just compute the right answer in the first place. LLMs are worthless.

>>106197998
Most graduates in mathematics couldn't solve these problems. The fact that the competition is for pre-college kids doesn't mean the problems are easy. These kids basically spend their life on them.

>>106198754
You're being closeminded. Yes, we expected it and were sold on the idea of it making calculators redundant. Sure, that's so far not happened and for the sake of argument may never happen. However, calculators don't do proofs. If the AI can illuminate even parts of the journey toward novel discoveries, while human intellect is the one holding the flashlight as it were, then that would still be valuable. The problem is human intelligence has sunken so low that most people want these AI to do their thinking for them, and aren't looking for an assistant per se which requires skillful and active wielding.

>>106198741
Retards still don't understand what tokenisation is.

>>106198741
>>106198910
qwen 2.5 answers this correctly tho

>moon-landing moment
Doing something relatively uselessL

>>106198917
Claude will also tell you the date, because every day they update the prompt with the current date. But knowing the date and reading individual characters in a token isn't what LLMs are good at. Trying to judge an AI by asking dumb questions like how many b's in blueberry is retarded.

Shit that never happened and boomer kool aid

>>106198890
No one is investing trillions of dollars in me on the premise that surely in two more months I will be able to.

>>106198979
Just because you're poor, doesn't mean that everyone else has to be.

>>106198904
Unless youre retarded and just don't have the mental capacity to understand, anyone can "solve" these math problems GIVEN the answers.
The point is the kids that have to solve these problems don't get trained on the exact answers.
Thats like claiming I know 5000 words in japanese because I have a dictionary in front of me right now

The only thing ai is good for is shitting out large quantities of text and code I can't be bothered to type myself

>>106198992
I was replying to the person who implied that winning the olympiads makes you as good as a high schooler.
The assumption is that the AI was not given the answers obviously.

>>106199031
Okay, point taken. But I don't think that assumption is valid. You can see the LLMs fail at basic math unless the exact answer is in their training set. Kind of obvous because LLMs aren't really made for math, which is why I think its wierd that they are always pushing this "math olympiad" shit so hard when its obviously untrue

>>106199209
Maybe they are super good at abstract math rather then something that you can do with a calculator

>>106198917
Reasoning models can do it, gpt 5 can do it too but openAI is cheapskating and wants to give the worst model for free now despite rivals providing a much better model in free tier

>>106198938
>don't judge the artificial INTELLIGENCE for not knowing how to count!
do retards really?

>>106200148
Skill issue. If you had written b-l-u-e-b-e-r-r-y it would have counted correctly.

>>106198910
tokenization is their problem to solve, its not an excuse

>>106200269
It's not a problem because only a retard would use LLM to ask to count letters. It's like saying calculators are bad at arithmetic because they can't write poetry.

>>106200298
their average costumer could not give a rats ass whether its an llm or not, for them its "AI that gives magical answers", and if it fails at that then thats openais problem
if "AGI", aka their explicit goal cant solve a simple problem like this, then thats not much of an intelligence and will fail in other non obvious cases as well because of it

>>106200330
Again, the AI could easily answer the question if the developers changed the definition of a from a whole word to a single letter, but it would be a ridiculous decision to feed letters instead of words just to appease some autists who want to count letters.

>>106200360
while they may be retarded to care about it that much, it is a limitation of the architecture and therefore a valid criticism that will have to be solved eventually

>>106200398
It's not the architecture, it's a compromise between performance and cost.

>>106198904
has been trained on right answers, bound to give right answers

>>106200410
every ai architecture is filled with compromises, and when they find a way for it to not take too much performance, im sure they will implement it and then it will no longer be a limitation

>>106198907
what they can do is mix and match already existing info, they're good at that. what they cannot do is come up with novel ideas for problems. they may be able to luck out by hallucinating something which makes sense for them based on their training data, and no human figured to do so far, and it would be luck even so. but coming up with some e=mc^2 out of the blue is not going to happen, because the whole thing is not working like a human brain does. it's unable to do basic shit that human brains can, but it can do some shit that humans can, fast, way faster, that's true. but llms are not the whole package