Terence Tao: AI's math proofs pass the eye test, fail miserably on the smell test

[lostsoul]

[lostsoul]

FateOfMuffins
The thing with the "smell" is that we don't have the relevant data to feed into these models. It's something that mathematicians develop over countless hours of trial and error, but the problem is in literature, only the final proofs are written down. How we actually came up with the solutions and the entire trial and error process is generally... not written down. Suppose it took a mathematician say 7 years to prove this big theorem, but all you're going to see is the final proof. You don't get to see the hundreds of pages of scrap work that they did over those years, if they're even written down at all. Some stroke of genius may be written down only as messy notes on a napkin from a random cafe one morning. Etc.

Edit: It's somewhat of a big bottleneck for humans learning math as well. You can have solutions for contest problems but it's not entirely useful for students. They'll read it, maybe even understand it, but are unable to replicate it themselves because the solutions never really go into how they thought of them in the first place. Why did they do this step instead of that step, how would one even conceive this solution in the first place, etc etc. You'll have students who upon listening to the solution think that it's so easy and understand once I explain it, but they can't do it themselves at home. Sometimes I try to show my students how to approach some of these problems by doing it completely from scratch and showing just how much of the kitchen sink I'm throwing at it - how I've probably tried 3 different ways that didn't stick and that they never see in the solution key, before finding something promising.

[lostsoul]

FateOfMuffins 的这个观察非常关键。AI 有数据，就能模仿。数学中， 发现 和 发表 几乎毫不相关。
AI 只有 发表 的数据，完全没有 发现 这方面的数据。没有这数据，它就永远过不了 smell 这一关。

[枫林晓1]

奇点说的是这个超越那个。指的是整体

不是说ai超越每一个人。人类里个别的高能力者。ai可能在某一方向上永远无法超越。

与奇点无关。奇点是ai超于人类。

目前ai只不过是缸中之脑。自然无法做很多人能做的事情。不过。你敢放出来吗？

让他们自我复制试试。