回複 'dhyang_wxc' 的評論 : Thanks. I will not trust AI ever, because everything is in a latent black box, and the gradient of descent can take many different path.
dhyang_wxc 發表評論於
回複 '矽穀工匠' 的評論 :
是的,這方麵有可能突破一個層次。
上回我們對話,你介紹了Confusion matrix,我看過之後,就想到了chain of thoughts。這個matrix用來衡量人工智能的靜態結果還可以。然而chain of thoughts在這個matrix裡,在critical points 或 pivot points 必須處在對角線上,但其餘部分卻可能是枉則直的,未必一定是沿著對角線發展。AI如果學會了這個,顯然會進步許多。
矽穀工匠 發表評論於
回複 'dhyang_wxc' 的評論 : 是啊,機器輸出是不受控的。現在看看test time computer, 小模型,chain of thoughts能走多遠吧,估計都是摸著石頭過河。deep search倒是確實可以顛覆穀歌,不小心幹掉了S E O。這個deep search好像沒有用上什麽C N N。
openAI的這夥人,可能正在lie until it is true,但他們大概已經知道沒有truth在前麵。lie就是為了在你所講的這些方麵裡,低於人智的某一個有突破,收回投入。不想出來個DS。DS報價低廉,跨者不行,沒利潤的積累也就沒有前途。但DS使其他AI的回收利潤近乎不可能實現。我不看好AI的前景,繼續觀望。。。
矽穀工匠 發表評論於
回複 'gwangmsn' 的評論 : Yes. Neurons use moledular gate nano channel, but MOSFET use transistor gate. You must transfer a couple of electrons for the potential to reach new equilibrium of zero and one. But overall, I think there are many things that can be done. Interestingly Musk consider Grok a "truth seeking machine".
gwangmsn 發表評論於
回複 '矽穀工匠' 的評論 : algorithms can be modified to match biology brain. It just like current math algorithm doesn't simulate biology brain. It will be different algorithm and just we didn't find the algorithm yet.
矽穀工匠 發表評論於
回複 'gwangmsn' 的評論 : Neurons are based on molecule transduction and very hard to reproduce with Silicon. Carver Mead tried analogy MOSFET in the past and now they are no where to be found.