[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-llms-work-by-predicting-next-token-zh":3,"article-related-llms-work-by-predicting-next-token-zh":33,"series-industry-c40b20df-d89a-43ae-bb11-11062dcd2cd2":81},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":25,"views":29,"created_at":30,"published_at":31,"topic_cluster_id":32},"c40b20df-d89a-43ae-bb11-11062dcd2cd2","llms-work-by-predicting-next-token-zh","5 個關鍵部件看懂 LLMs","\u003Cp data-speakable=\"summary\">這篇用 5 個部件拆解大型語言\u003Ca href=\"\u002Fnews\u002Fwindsurf-model-router-ide-opinion-june-2026-zh\">模型\u003C\u002Fa>，幫你看懂它怎麼學、怎麼猜下一個 \u003Ca href=\"\u002Ftag\u002Ftoken\">token\u003C\u002Fa>、以及該怎麼選模型。\u003C\u002Fp>\u003Cp>讀完這 5 項，你就能判斷一個 \u003Ca href=\"\u002Ftag\u002Fllm\">LLM\u003C\u002Fa> 的\u003Ca href=\"\u002Fnews\u002Fspacex-shou-gou-cursor-bu-hua-suan-ai-bian-cheng-zh\">能力\u003C\u002Fa>從哪裡來，也能分辨它是靠規模、結構，還是後訓練調整在發揮作用。\u003C\u002Fp>\u003Ctable>\u003Cthead>\u003Ctr>\u003Cth>項目\u003C\u002Fth>\u003Cth>做什麼\u003C\u002Fth>\u003Cth>關鍵細節\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd>訓練資料\u003C\u002Ftd>\u003Ctd>提供語言樣本\u003C\u002Ftd>\u003Ctd>來自書籍、文章、網站與程式碼\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>Tokenization\u003C\u002Ftd>\u003Ctd>把文字切成可運算單位\u003C\u002Ftd>\u003Ctd>可用詞、子詞或字元\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>Transformer attention\u003C\u002Ftd>\u003Ctd>抓住 token 之間的關係\u003C\u002Ftd>\u003Ctd>靠 query、key、value 向量\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>Parameters\u003C\u002Ftd>\u003Ctd>儲存學到的行為\u003C\u002Ftd>\u003Ctd>可達數十億到數兆\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>Fine-tuning \u002F RLHF\u003C\u002Ftd>\u003Ctd>把通用模型調成特定用途\u003C\u002Ftd>\u003Ctd>提升對齊、可用性與一致性\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003Ch2>1. 大量訓練資料先把語感養出來\u003C\u002Fh2>\u003Cp>LLM 不是先懂語言再上線，而是先看過大量文本，從中學出語法、風格、常見搭配與知識分布。資料來源通常包含書籍、文章、網站與程式碼，範圍越廣，模型越能在不同語境下維持表現。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781889466449-7e9g.png\" alt=\"5 個關鍵部件看懂 LLMs\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>在這一步，資料品質比單純的數量更重要。去重、清理錯誤、過濾不想要的內容，都是避免模型學歪的基本功。\u003C\u002Fp>\u003Cul>\u003Cli>來源：書籍、文章、網站、程式碼\u003C\u002Fli>\u003Cli>處理：清理、去重、過濾\u003C\u002Fli>\u003Cli>目的：讓模型學到可泛化的語言模式\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>2. Tokenization 讓文字變成模型看得懂的單位\u003C\u002Fh2>\u003Cp>模型不直接讀「句子」，而是先把文字切成 token。token 可以是詞、子詞，甚至字元，這讓系統能處理生字、罕見詞與變形字，不必依賴完整詞彙表。\u003C\u002Fp>\u003Cp>這也是 LLM 最核心的運作方式之一：它不是像人一樣理解整段文字，而是根據 token 序列去預測下一個最可能出現的 token。\u003C\u002Fp>\u003Cul>\u003Cli>常見切法：詞、子詞、字元\u003C\u002Fli>\u003Cli>好處：能處理新詞與罕見詞\u003C\u002Fli>\u003Cli>結果：文字變成可計算的輸入\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>3. Transformer attention 決定哪些詞最重要\u003C\u002Fh2>\u003Cp>LLM 多半建立在 transformer 架構上，而 self-attention 是它能處理長句與上下文關係的關鍵。這個機制會衡量序列中不同 token 的重要性，即使兩個詞隔得很遠，也能建立關聯。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781889465978-pti1.png\" alt=\"5 個關鍵部件看懂 LLMs\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>實作上，模型會把每個 token 映射成 query、key、value 三種向量，再計算權重分配資訊。這讓模型能在回答問題時，抓住真正\u003Ca href=\"\u002Fnews\u002Fmanus-ai-github-clone-kits-zh\">相關\u003C\u002Fa>的線索，而不是只看最近的字。\u003C\u002Fp>\u003Ccode>token → embedding → query\u002Fkey\u002Fvalue → attention weights → output\u003C\u002Fcode>\u003C\u002Fh2>\u003Ch2>4. Parameters 把學到的模式存進模型\u003C\u002Fh2>\u003Cp>訓練時，模型會不斷調整內部權重，也就是 parameters。這些參數決定模型怎麼處理輸入、怎麼產生輸出；規模大的模型可能有數十億甚至數兆參數，因此在能力與成本之間會出現明顯取捨。\u003C\u002Fp>\u003Cp>參數越多，模型通常越能吸收語法、寫作風格、推理套路與領域模式，但也更吃算力與記憶體。若部署環境有限，小型語言模型會是更實際的選擇。\u003C\u002Fp>\u003Cul>\u003Cli>Parameters：訓練後形成的內部權重\u003C\u002Fli>\u003Cli>規模：可到數十億或數兆\u003C\u002Fli>\u003Cli>小模型：更適合低資源設備與緊縮預算\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>5. Fine-tuning 和 RLHF 把通用模型調成可用工具\u003C\u002Fh2>\u003Cp>預訓練完成後，模型通常還會進一步 fine-tuning，讓它更適合客服、摘要、寫程式或特定領域問答。這一步不是重學語言，而是把已經會說話的模型，調整成更符合任務需求的版本。\u003C\u002Fp>\u003Cp>常見方法是 RLHF，也就是 \u003Ca href=\"https:\u002F\u002Fwww.ibm.com\u002Fthink\u002Ftopics\u002Freinforcement-learning-from-human-feedback\">reinforcement learning from human feedback\u003C\u002Fa>。人類會比較不同輸出，模型再學會偏好更有幫助、更安全、也更一致的回答。\u003C\u002Fp>\u003Cul>\u003Cli>Fine-tuning：把通用模型改成特定用途\u003C\u002Fli>\u003Cli>RLHF：用人類偏好改善輸出\u003C\u002Fli>\u003Cli>目標：提高對齊、可用性與一致性\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>哪種適合你\u003C\u002Fh2>\u003Cp>如果你只想抓住整體脈絡，記住一句話就夠：LLM 是靠下一個 token 預測運作，先用大量資料學語言，再靠 attention、parameters 與後訓練方法把能力磨出來。\u003C\u002Fp>\u003Cp>如果你在看部署成本，先看 parameters 和模型大小；如果你在看輸出品質，先看 fine-tuning 與 RLHF；如果你想理解原理，先從 tokenization 和 self-attention 下手。\u003C\u002Fp>","5 個關鍵部件帶你看懂 LLMs 如何從資料、token、注意力到對齊，進而判斷訓練與部署該看什麼。","www.ibm.com","https:\u002F\u002Fwww.ibm.com\u002Fthink\u002Ftopics\u002Flarge-language-models",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781889466449-7e9g.png","industry","zh","d0dd8e84-c799-4d99-b7ef-14f9eac9f7dc",[17,18,19,20,21,22,23,24],"LLM","tokenization","transformer","self-attention","fine-tuning","RLHF","parameters","next token prediction",[26,27,28],"LLM 的核心是預測下一個 token，不是逐字理解整段文字。","訓練資料、tokenization 與 self-attention 是理解模型行為的三個基礎。","parameters 決定模型規模與成本，fine-tuning 和 RLHF 決定它是否更符合任務需求。",0,"2026-06-19T17:17:20.910277+00:00","2026-06-19T17:17:20.903+00:00","fe20f6f6-432b-47bf-a410-a5f516d885ed",{"tags":34,"relatedLang":40,"relatedPosts":44},[35,36,37,38],{"name":19,"slug":19},{"name":18,"slug":18},{"name":21,"slug":21},{"name":17,"slug":39},"llm",{"id":15,"slug":41,"title":42,"language":43},"llms-work-by-predicting-next-token-en","LLMs work by predicting the next token","en",[45,51,57,63,69,75],{"id":46,"slug":47,"title":48,"cover_image":49,"image_url":49,"created_at":50,"category":13},"5a410687-834e-4767-8bac-11251400de18","pentagon-should-not-use-grok-wartime-targeting-zh","五角大廈不該用 Grok 做戰時打擊決策","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781903867710-wz9f.png","2026-06-19T21:17:18.215115+00:00",{"id":52,"slug":53,"title":54,"cover_image":55,"image_url":55,"created_at":56,"category":13},"1570bfba-1886-43c4-a0c6-5ef97b5da551","grok-latest-controversies-regulation-story-zh","5 則 Grok 爭議，已變成監管問題","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781902971533-q3fy.png","2026-06-19T21:02:21.155982+00:00",{"id":58,"slug":59,"title":60,"cover_image":61,"image_url":61,"created_at":62,"category":13},"82982d74-02ac-4638-adf7-fc28d119c252","aibox-ax8850-hardware-first-integration-zh","AIBOX 不是拼軟體，關鍵在把 AX8850 的硬體吃滿","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781900274678-ladh.png","2026-06-19T20:17:23.586922+00:00",{"id":64,"slug":65,"title":66,"cover_image":67,"image_url":67,"created_at":68,"category":13},"d8a73bff-aaa0-45c4-a8e3-a1764a5c01ce","ai-coding-assistant-roi-measured-zh","AI 寫碼助手有 ROI，但前提是你真的去量","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781893067880-82y2.png","2026-06-19T18:17:19.809941+00:00",{"id":70,"slug":71,"title":72,"cover_image":73,"image_url":73,"created_at":74,"category":13},"2c14071a-9780-4d9d-9ed0-c12fa1b40501","red-hat-ai-mavenir-telco-ai-stack-zh","Red Hat AI 把電信 AI 變成堆疊","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781885880665-z3qj.png","2026-06-19T16:17:38.29401+00:00",{"id":76,"slug":77,"title":78,"cover_image":79,"image_url":79,"created_at":80,"category":13},"eda7a48a-4eda-4111-9103-38a24eaaec6a","manus-ai-github-clone-kits-zh","Manus AI 相關 GitHub 倉庫多是克隆套件","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781882272627-ya10.png","2026-06-19T15:17:26.427636+00:00",[82,87,92,97,102,107,112,117,122,127],{"id":83,"slug":84,"title":85,"created_at":86},"ee073da7-28b3-4752-a319-5a501459fb87","ai-in-2026-what-actually-matters-now-zh","2026 AI 真正重要的事","2026-03-26T07:09:12.008134+00:00",{"id":88,"slug":89,"title":90,"created_at":91},"83bd1795-8548-44c9-9a7e-de50a0923f71","trump-ai-framework-power-speech-state-preemption-zh","川普 AI 框架瞄準電力、言論與州權","2026-03-26T07:12:18.695466+00:00",{"id":93,"slug":94,"title":95,"created_at":96},"ea6be18b-c903-4e54-97b7-5f7447a612e0","nvidia-gtc-2026-big-ai-announcements-zh","NVIDIA GTC 2026 重點拆解","2026-03-26T07:14:26.62638+00:00",{"id":98,"slug":99,"title":100,"created_at":101},"4bcec76f-4c36-4daa-909f-54cd702f7c93","claude-users-spreading-out-and-getting-better-zh","Claude 用戶更分散，也更會用","2026-03-26T07:22:52.325888+00:00",{"id":103,"slug":104,"title":105,"created_at":106},"bd903b15-2473-4178-9789-b7557816e535","openclaw-raises-hard-question-for-ai-models-zh","OpenClaw 逼問 AI 模型價值","2026-03-26T07:24:54.707486+00:00",{"id":108,"slug":109,"title":110,"created_at":111},"eeac6b9e-ad9d-4831-8eec-8bba3f9bca6a","gap-google-gemini-checkout-fashion-search-zh","Gap 把結帳搬進 Gemini","2026-03-26T07:28:23.937768+00:00",{"id":113,"slug":114,"title":115,"created_at":116},"0740e53f-605d-4d57-8601-c10beb126f3c","google-pushes-gemini-transition-to-march-2026-zh","Google 把 Gemini 轉換延到 2026 年 3…","2026-03-26T07:30:12.825269+00:00",{"id":118,"slug":119,"title":120,"created_at":121},"e660d801-2421-4529-8fa9-86b82b066990","metas-llama-4-benchmark-scandal-gets-worse-zh","Meta Llama 4 分數風波又擴大","2026-03-26T07:34:21.156421+00:00",{"id":123,"slug":124,"title":125,"created_at":126},"183f9e7c-e143-40bb-a6d5-67ba84a3a8bc","accenture-mistral-ai-sovereign-enterprise-deal-zh","Accenture 攜手 Mistral AI 賣主權 AI","2026-03-26T07:38:14.818906+00:00",{"id":128,"slug":129,"title":130,"created_at":131},"191d9b1b-768a-478c-978c-dd7431a38149","mistral-ai-faces-its-hardest-year-yet-zh","Mistral AI 迎來最硬的一年","2026-03-26T07:40:23.716374+00:00"]