跳到主要內容

發表文章

目前顯示的是有「Large Language Model」標籤的文章

目前AI(LLM)的原理概論:LLM 沒有真正理解,而是統計式理解!

緣起: 我在公司負責 導入AI 技術與治理規範,做了一些研究。 邀請大家一起閱讀與討論,以了解目前 AI 目前真實的能力與限制 ! 才不會被賣 AI 的人搞得很焦慮 ! 我試過了Microsoft  Copilot Chat 、 Claude 、 Gemini 、 ChatGPT , 我最喜歡 Microsoft  Copilot 的答案,更可貴的它提到引發 AI 熱潮, Google 無私公開的論文: Attention Is All You Need 我的指令 (Prompt) : 下面句子的 it 代表什麼 ? 你是怎麼知道的,會不會有出錯的時候 ? The animal didn't cross the street because it was too tired. 請說明清楚技術,如過出現專有名詞請說明清楚: 包含但不僅限於 Large Language modle , Natural Language Processing   Microsoft Copilot chat : 這個問題其實碰到了 大型語言模型( Large Language Model, LLM )最核心的能力之一:指代消解( Coreference Resolution ) 。 先說結論: 在句子 "The animal didn't cross the street because it was too tired." 大部分人類以及大部分 LLM ,都會判斷: it = the animal (那隻動物) 也就是: 那隻動物沒有穿越馬路,因為牠太累了。 但重點是: AI 並不是「真的知道」 it 是 animal ,而是根據語言模式、機率、語意關聯去推斷,因此有可能出錯。 一、人類為什麼知道 it 是 animal ? 句子: 1     The animal didn't cross the street because it was too tired. 拆解: 1     主...

標籤

顯示更多