I wonder about the mechanisms of the engine of the AI math hype. I'm gonna play devil's advocate for a moment. Let's create an alternate version of me that fully believes in it. How would you answer the question that, that version of me poses, specifically "How is AI so good at math"? I'd like to hear the opinions of AI accels on this. I wonder if they have an intriguing answer that's not just "daddy Altman told me so". For the record I am very staunchly anti-AI and I think I have some well founded reasons to believe 99% of the recent hype is founded entirely in bullshit, completely hollow with no actual meat and potatoes. But I'd like to be proven wrong! Even if accels don't exactly make convincing arguments. I guess in that department they just don't have the mental firepower.
>>17078600Once you get beyond the "easy" part of math where you just learn to crunch numbers (this is, ironically, the part consumer grade LLM's notoriously suck at), math really becomes about making connections and finding novel approaches. This is an area where LLM's become a valid tool.The whole point of an LLM is it tries to predict the next token from a probability distribution of previous tokens. The relatedness of seemingly distinct concepts is baked into their architecture. Don't ask an LLM to solve an integral for you. That's what Wolfram Alpha is for. What a sufficiently advanced LLM can do is point you to the class of equations which represent the best-match model for the problem you're facing.
>>17078600You can launch a thousand instances of it and have each one try a thousand different approache, while efficiently communicating with the others, in a matter of hours. Agentic "AI" has superhuman abilities. If you reduce mathematics into the business of solving specific problems within a pre-established framework, it's gonna beat you at math, alright.
>>17078600LLMs are pattern matchers. mathematics beyond arithmetic is pattern matchingthat's it, that's how they are "so good" at it: they match patterns faster and more reliably than humansthe>it's all bullshitdefense is the same defense that everyone who wasn't Hardy put up against Ramanujan
>>1707866370 IQ take.
>>17078663You can probably turn most problems into a n-dimensional representation and use an approach closer to a graphical than argumentative one ... then find an argument that fits the solved representation. Don't ask me btw, I am really bad at math but decent at visualizing shit. ^^
>You can probably turn most problems into a n-dimensional representation and use an approach closer to a graphical than argumentative one
>>17078600They won't ever take the sanctum. The academics might buckle to the pressure, but the backward number fuckers will grow and promulgate and have their day.
>>17078694Hey don't tell me I've accidentally been right! That was not even an educated bellyfeeling.
>>17078769it's obviously "true", it's just not interesting
>>17078671how else could you have understood it amirite xDdas per Hung Le's 40+ hours on the l_1 embedding papers, if there's novel and useful structures buried inside the papers that only reveal themselves after letting the same model blow up ~40 pages to >90, then it's impossible who exactly the idiot is
>>17078962Meds.
