生成式人工智慧的演進:從規則到推理
人工智慧的歷史標誌著一場根本性的轉變:從明確的人工程式設計,轉向基於模式的統計預測。這種演進使現代人工智慧能夠執行複雜的 推理 任務。
1. 什麼:規則為本時代
早期的人工智慧依賴於 專家系統。在這些系統中,每一個可能的回應或動作都是由人類以嚴格的 如果-那麼邏輯。
- 限制: 這些系統極其脆弱。它們無法處理細微差異、俚語、拼字錯誤,或任何超出其特定硬編碼程式的場景。
2. 為何:統計學突破
突破來自於處理大量未標記資料的能力。與手動規則不同, 大型語言模型(LLMs) 學習詞彙之間的統計關係。
- 變換器(Transformer): 2017年推出的革命性模型架構。
- 注意力機制: 變換器的核心組件,讓模型能權衡序列中不同詞語的重要性,以理解深層語境(例如,知道長段落中的「它」指代的是什麼)。
3. 如何:從預測到推理
現代生成式人工智慧根本上是 非決定性的。它計算「下一個詞元」的概率分佈,而非遵循固定的決策樹。
透過反覆根據整個前序內容預測最有可能的下一個單詞,模型產生創意內容,並看似能「推理」自然語言中提供的複雜指令。
概率陷阱
人工智慧不是事實資料庫;它是一台統計引擎。由於它僅預測最有可能出現的下一個單詞,因此可能陷入 「幻覺」——以絕對信心呈現錯誤資訊。
TERMINALbash — 80x24
> Ready. Click "Run" to execute.
>
Question 1
What is the primary difference between rule-based chatbots and modern Generative AI?
Question 2
What does the 'Attention Mechanism' in a Transformer model do?
Challenge: Designing a Tutoring App
Apply your knowledge of AI evolution.
You are designing a tutoring app. You need to choose between a rule-based "if-then" system and an LLM.
Task 1
Identify a scenario where the rule-based system would fail but the LLM would succeed.
Solution:
Handling a student asking the same question in a creative or slang-heavy way (e.g., "Yo, how do I do math?" vs "Please explain the equations."). A rule-based system would likely throw an error if the exact phrasing wasn't programmed.
Handling a student asking the same question in a creative or slang-heavy way (e.g., "Yo, how do I do math?" vs "Please explain the equations."). A rule-based system would likely throw an error if the exact phrasing wasn't programmed.
Task 2
Suggest a "Metaprompt" to ensure the LLM doesn't just give the answer but acts like a tutor.
Solution:
"You are a helpful tutor. Do not provide direct answers. Instead, ask leading questions to help the student find the solution themselves."
"You are a helpful tutor. Do not provide direct answers. Instead, ask leading questions to help the student find the solution themselves."