02. How LLMs Understand Prompts

flowchart LR
    subgraph HUMAN["Human Reading"]
        A["Sees words"] --> B["Understands meaning"]
        B --> C["Thinks about response"]
        C --> D["Writes response"]
    end

    subgraph LLM["LLM Processing"]
        E["Sees tokens"] --> F["Converts to vectors"]
        F --> G["Pattern matches\nagainst training"]
        G --> H["Predicts next token\nx 1000 times"]
    end

    style HUMAN fill:#3b82f6,color:#fff
    style LLM fill:#f59e0b,color:#fff

"Let's meet for → lunch | coffee | dinner"

flowchart TD
    PROMPT["'Write a quicksort in TypeScript'"] --> TOKENIZE["Tokenizer"]
    TOKENIZE --> TOKENS["Tokens:\n'Write' | ' a' | ' quick' | 'sort' | ' in' | ' Type' | 'Script'"]
    TOKENS --> EMBED["Embedding Layer"]
    EMBED --> VECTORS["Vector Representations\n(numbers the model can process)"]

    style PROMPT fill:#3b82f6,color:#fff
    style TOKENS fill:#f59e0b,color:#fff
    style VECTORS fill:#22c55e,color:#fff

sequenceDiagram
    participant User as User
    participant Token as Tokenizer
    participant Embed as Embedding
    participant Layers as Transformer Layers
    participant Output as Output Layer

    User->>Token: Sends prompt text
    Token->>Token: Split into tokens
    Token->>Embed: Token IDs
    Embed->>Embed: Convert to vectors
    Embed->>Layers: Vector sequences
    Layers->>Layers: Process through attention
    Layers->>Layers: Self-attention, feed-forward
    Layers->>Output: Final hidden states
    Output->>User: Predicted tokens (response)

"cat" → [0.23, -0.45, 0.67, ..., 0.12]  (a vector in 4096-dimensional space)
"dog" → [0.21, -0.42, 0.65, ..., 0.15]  (similar vector — similar meaning)
"car" → [-0.31, 0.52, -0.71, ..., 0.08]  (different vector — different meaning)

flowchart LR
    subgraph INPUT["Input Tokens"]
        T1["The"] --> P1["Pos 0"]
        T2["cat"] --> P2["Pos 1"]
        T3["sat"] --> P3["Pos 2"]
        T4["on"] --> P4["Pos 3"]
        T5["the"] --> P5["Pos 4"]
        T6["mat"] --> P6["Pos 5"]
    end

    subgraph ATTENTION["Self-Attention"]
        P1 --> A1["'The' attends to: cat, sat, on, the, mat"]
        P2 --> A2["'cat' attends to: The, sat, on, the, mat"]
        P3 --> A3["'sat' attends to: The, cat, on, the, mat"]
    end

    style INPUT fill:#3b82f6,color:#fff
    style ATTENTION fill:#f59e0b,color:#fff

System: You are a helpful coding assistant.
User: Write a binary search function in Python.

❌ "Sort this array" → Model might: explain sorting, ask questions, show multiple algorithms
✅ "Sort this array using quicksort in Python with O(n log n) and return the result" → Model does exactly that

flowchart TD
    subgraph VAGUE["Vague Prompt\n'H Write code'"]
        V1["Which language?"] --> V2["High probability\nfor many options"]
        V2 --> VR["Random/ variable\noutput"]
    end

    subgraph SPECIFIC["Specific Prompt\n'Write Python quicksort'"]
        S1["Language: Python"] --> S2["Algorithm: quicksort"]
        S2 --> S3["Narrow probability\ndistribution"]
        S3 --> SR["Predictable\noutput"]
    end

    style VAGUE fill:#ef4444,color:#fff
    style SPECIFIC fill:#22c55e,color:#fff

❌ "Explain promises"
   → Model might explain: JavaScript promises, Promise theory, political promises

✅ "Explain JavaScript Promises to a beginner who knows callbacks
   → Model knows: topic (JavaScript), audience (beginner), prerequisite (callbacks)

❌ "Write a function to find duplicates in an array"
   → Returns: could be O(n²) with nested loops

✅ "Write a function to find duplicates in an array. Time: O(n), Space: O(n).
   Input: [1,2,3,2,4,1], Output: [1,2]"
   → Returns: optimized hash map solution

❌ "List 5 sorting algorithms"
   → Could return: paragraph, bullet list, numbered list, or table

✅ "List 5 sorting algorithms in a markdown table with columns:
   | Algorithm | Time Complexity | Space Complexity | Stable |"
   → Returns: exactly the table you wanted