AI說書 - 從0開始 - 46

.article-container { width: 100%; font-family: Microsoft JhengHei,Helvetica Neue,Helvetica,Arial,sans-serif; } ul, ol { margin: 12px auto; max-width: 740px; color: #535150; line-height: 1.8; padding-left: 0px; } .graf--img { display: table; justify-content: center; align-items: center; text-align: center; color: gray; font-size: 14px; letter-spacing: 0px; margin: 10px auto 50px; width: 100%; position: relative; clear: both; } .graf--img.center img { width: 100%; max-width: 740px; margin: 10px auto 0px; display: block; margin: 0 auto; } .graf--img.full img { width: 100%; } .captionTheme__wrapper { width: 100%; font-style: normal; line-height: 22px; font-size: 16px; max-width: 600px; margin-top: 8px; display: inline-block; } .graf--img.full { max-width: 100%; margin: 40px 0px; display: block; margin: 0 auto; align-items: center; } .graf--figure { text-align: center; color: gray; font-style: italic; font-size: 15px; margin: 28px auto; box-sizing: border-box; } .graf--figure iframe { width: 100%; max-width: 740px; margin: 0 auto; } .graf--p { font-size: 16px; line-height: 1.8; font-family: Microsoft JhengHei fixed, Helvetica Neue ,Microsoft JhengHei, Helvetica, Segoe UI, Tahoma, Arial, sans-serif; letter-spacing: 1px; font-weight: 400; max-width: 740px; color: #535150; text-align: left; } .graf--p > a { color: #00B3C6 !important; text-decoration: none !important; } .graf--li > a { color: #00B3C6 !important; text-decoration: none !important; } .graf--quotesSpecial > a { color: #00B3C6 !important; text-decoration: none !important; } .graf--blockquote > a { color: #00B3C6 !important; text-decoration: none !important; } .graf--h1 > a { color: #00B3C6 !important; text-decoration: none !important; } .graf--h2 > a { color: #00B3C6 !important; text-decoration: none !important; } .graf--h3 > a { color: #00B3C6 !important; text-decoration: none !important; } a.graf--mention { color: #535150 !important; text-decoration: underline !important; font-weight: 700; } .graf--h2 { font-size: 24px; padding: 0; max-width: 740px; text-align: left; letter-spacing: 1px; font-weight: 700; margin-top: 34px; line-height: 1.5; } .graf--h3 { font-size: 18px; padding: 0; max-width: 740px; text-align: left; letter-spacing: 1px; font-weight: 700; margin-top: 28px; line-height: 1.5; } .graf--li { font-size: 16px; padding: 0px 0px 0px 4px; font-weight: 400; letter-spacing: 0px; list-style-position: outside; text-align:left; margin-left: 24px; } .graf--hr { width: 100%; margin: 0px auto; transform: translateY(-50%); position: relative; padding: 0px; text-align: left; max-width: 740px; margin: 0 auto; } .graf--hr hr { height: 0; } .graf--blockquote { padding: 10px 0px 10px 16px; font-size: 16px; color: #7A7574; letter-spacing: 1px; margin: 28px 0px; border-left: 4px solid #DDD9D8; width: 100%; max-width: 740px; text-align: left; } .graf--quotesSpecial { display: table; color: #7A7574; position: relative; padding: 31.5px 40px; text-align: center; letter-spacing: 0px; position: relative; margin: 29px auto; font-family: Microsoft JhengHei fixed, Helvetica Neue, Microsoft JhengHei, Helvetica, Segoe UI, Tahoma, Arial, sans-serif; font-size: 16px; -webkit-box-ordinal-group: 1; -webkit-box-flex: 0; } .embed-wrapper { max-width: 740px; border: 1px solid #DDD9D8; display: block; padding: 12px; border-radius: 8px; margin: 12px 0px; text-decoration: none !important; } .embed-title { font-size: 16px; font-weight: 700; color: #535150; margin-bottom: 8px; text-align: left; line-height: 1.5; text-decoration: none !important; } .embed-description { width: 100%; font-size: 14px; color: #7A7574; line-height: 1.5; max-height: 150px; text-align: left; overflow: hidden; padding: 12px 0px; } .embed-url > a { width: 100%; font-size: 14px; color: #141413 !important; text-decoration: none !important; line-height: 1.5; text-align: left; } .embed-thumbnail-wrapper { padding-left: 12px; } .embed-thumbnail { width:100px; border-radius: 8px; } pre { background: #F6F6F6; border-radius: 8px; padding: 16px; font-size: 16px; color: #535150; line-height: 180%; text-align: left; } .lexical__textBold { font-weight: bold; } .lexical__textItalic { font-style: italic; } .lexical__textUnderline { text-decoration: underline; } .lexical__textStrikethrough { text-decoration: line-through; } .lexical__textUnderlineStrikethrough { text-decoration: underline line-through; } .lexical__textSubscript { font-size: 0.8em; vertical-align: sub; } .lexical__textSuperscript { font-size: 0.8em; vertical-align: super; } .lexical__textCode { background-color: rgb(240, 242, 245); padding: 1px 0.25rem; font-family: Menlo, Consolas, Monaco, monospace; font-size: 94%; } 我想要一天分享一點「LLM從底層堆疊的技術」，並且每篇文章長度控制在三分鐘以內，讓大家不會壓力太大，但是又能夠每天成長一點。延續AI說書 - 從0開始 - 45，我們介紹了 Google 於2017 年提出的 Transformer 架構的 Positional Encoding (PE) 方法為 :圖片出自：Transformers for Natural Language Processing and Computer Vision - Third Edition, Denis Rothman, 2024可以看到 Sin 函數對應偶數位置，而 Cos 函數對應奇數位置，我們可能好奇這函數到底長得怎麼樣，有鑑於此，我們可以在 Google 瀏覽器中鍵入 : 「plot y=sin(2/10000^(2*x/512))」來追蹤位置 2 號 (pos = 2) 的圖形長怎麼樣。如下所示 :

AI說書 - 從0開始 - 46

Learn AI 不 BI 的其他內容

發表回應