Will Diaital IntelligenceReplace Bioloqical Intelligence数字智能是否会取代生物智能?GeoffreyHinton杰弗里·辛顿sitvof TorontaWaIC
Two paradiams for intelligence,两种智能范式Thelogic-inspired approach;、逻辑启发范式:The essenceofintligencelsreasoning智能的本质在于推理。This isdone by using symbolic rules to manipulate·通过符号规则对符号表达式进行操作以实现推理。symbolic expressions.Learningcan wait.学习可以暂缓,首先要理解知识如何以符号表达的形式Understandinghowknowledre isrepresentedin进行表示。symbolic expressionsmust comefirst.WAIC
What happenedin thenext30 years?接下来的三十年发生了什么?10years:YoshuaBengio shows this approachworks十年后:YoshuaBengio展示了这种方式可以来建楼for real language.真实的自然语言。20 vears:Computationalinquistsfinally start using二十年后:计算语言学家终于开始接受embeddingvectors."特征向量(嵌入)"30 years: Google invents translormers.三十年后:谷歌发明了Transformer,OpenAl向世OpenAlshowswhat theycando界展示了它的强大能力。WAIC
LargeLanguageModels大语言模型LMsunderstand languageinthe sameway aspeople.,大语言模型理解语言的方式与人类非常相似Theyconvert words intoleature vectors大语言模型把词转化为能和其他词"配合得很好"的特that fit togethernicely,征向量、,LlMs really do understand what they are saying大语言模型确实“理解"它们所说的话。WAIC
ALego analogy for how words work用乐高类比词语是如何运作的.UsingLego blocks, we can model any large3-D,我们可以用乐高积木非常好地建构出各种大型三维结构Words are like high-dimensionalLego blocks which,,词语就像是高维的乐高积木,可以用来建构几乎任何事物can beused formodellinganyhingat allThesemodelscanbecommunicated toothe,这些建构可以被传达给其他人。people.WAIC
ALegoanalogyforhowwordswork用乐高类比词语是如何运作的Therearethousandsaofdifferent words虽然有成千上万个形状各异的词语，但每个“形状"that havedifferent shapes都有一定的灵活性，能够根据上下文进行变形,与其but each shane has some flexibility他词语契合。Itcan deformtofit in withotherwordsin the contexEach word hasmany odlyshaped hands每个词都有许多形状奇特的“手",需要与其他词语Itshakes hands with other words"握手"才能组合在一起,Understandinga sentence ismuchmorelike理解一句话更像是折叠一个蛋白质分子，而不是将其tolding a protein molecule than translating翻详成一种明确无歧义的逻辑表达式。WAIC
Summary sofar目前小结Understandinga sentence consistsof associating理解一个句子,就是为句中的词分配彼此兼容的特征mutuallycompatible feature vectors with the words in向量。the sentence.-LMsunderstandlanguageinmuch thesamewayas·大语言模型理解语言的方式与人类非常相似。它们在people. They are very like us and very unlike normal很多方面像我们,却又与传统计算机软件截然不同。But there is one way in which digital LMs are fan但有一点,数字化的大语言模型远远优于我们类比信号驱动的大脑WAIC
Digitalcomputation数字计算Aundamntalpropertyacaretompoesthat·当代计算机的一个基本特性是:我们可以在不同的物we can run the same proarams (or the sameneura理硬件上运行相同的程序(或相同的神经网络nets)on different phvsical pleces of hardware.This means the knowledgecontained in theprogram (or in the weights) is immortal:Itis,·这意味着程序中的知识(或神经网络的权重)是永生independent of any particular piece of hardware.的:它不依赖于任何特定的硬件。
...
