{
  "lesson": "12-edge-inference",
  "title": "Edge Inference — Apple Neural Engine, Qualcomm Hexagon, WebGPU/WebLLM, Jetson",
  "questions": [
    {
      "stage": "pre",
      "question": "What is the core constraint that makes mobile LLM inference slower than datacenter, per the lesson?",
      "options": [
        "Storage capacity",
        "Wi-Fi latency",
        "Compute throughput",
        "Memory bandwidth (mobile DRAM at 50-90 GB/s vs HBM3 at 2-3 TB/s)"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Why does Apple's Neural Engine avoid CPU-NPU copy overhead?",
      "options": [
        "Apple Silicon ships unified memory — CPU and ANE share the same pool",
        "It uses PCIe 5.0",
        "Core ML disables KV cache",
        "It transcodes weights to FP4 before copy"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Which quantization format does the lesson recommend for WebGPU + WebLLM in the browser?",
      "options": [
        "GGUF Q4_K_M",
        "NVFP4",
        "FP8",
        "Q4 MLC (q4f16_1) compiled via mlc_llm convert_weight"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Roughly what WebGPU mobile coverage does the lesson report for 2026?",
      "options": [
        "Under 10%",
        "Only iOS Safari",
        "About 70-75%, with Firefox Android still catching up",
        "100% across all browsers"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "Why is keeping 128K context impractical on a typical phone?",
      "options": [
        "Tokenizers fail above 8K",
        "Model weights plus KV cache for 32K tokens plus OS overhead easily exceed the 8 GB RAM budget",
        "WebGPU caps context at 4K",
        "iOS forbids long context"
      ],
      "correct": 1,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "Why is voice highlighted as the killer app for edge inference?",
      "options": [
        "Voice models do not need KV cache",
        "Voice models always fit in 50 MB",
        "Voice agents are latency-sensitive (first token < 500 ms) and local inference eliminates network latency entirely",
        "Voice runs in WebAssembly only"
      ],
      "correct": 2,
      "explanation": ""
    }
  ]
}