On-Device AI: Running Models on Your Phone and Laptop

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-ai-foundations-on-device-final1-creators

1. What is a primary advantage of running AI models locally on personal devices instead of using cloud-based AI services?

   1. The local model will always produce more accurate responses than cloud models
   2. Local models automatically update themselves whenever improvements are released
   3. Consumer devices can run the largest available AI models with billions of parameters
   4. Sensitive data processed by the model stays on the device and is never sent to external servers

2. What does quantization refer to in the context of on-device AI?

   1. A security measure that prevents unauthorized access to local models
   2. A technique to reduce model file size by using lower-precision numbers for calculations
   3. The method AI uses to convert human language into numerical data tokens
   4. The process of downloading a model from the internet to local storage

3. A user installs a local AI app and runs a model entirely offline. Why might their data still NOT be private?

   1. Local models require an internet connection during initial installation, exposing all data
   2. Government regulations force all AI apps to report user data to authorities
   3. The local model automatically shares conversation history with the AI research community
   4. The app wrapper itself could send prompts to a server for analytics or as a fallback to cloud AI

4. Which type of task is MOST likely to achieve similar quality between a local 3B-8B parameter model and a frontier cloud model like ChatGPT?

   1. Writing novel computer code for an emerging programming language
   2. Answering questions about very recent world events that happened this week
   3. Complex multi-step mathematical proofs requiring advanced reasoning
   4. Summarizing a short email or document into key points

5. What does the lesson identify as 'the model-update problem' with on-device AI?

   1. Models become corrupted if not updated every 30 days
   2. Users must manually download new versions of models to get improvements—local models do not auto-improve
   3. The process of updating a local model causes all previous conversations to be deleted
   4. Local models gradually become less accurate as they age without updates

6. What hardware limitation most directly prevents consumer laptops from running the largest AI models available today?

   1. Lack of specialized AI chips designed for machine learning
   2. Insufficient processing speed compared to cloud data centers
   3. Inadequate RAM or VRAM to hold the massive model parameters in memory
   4. Poor internet connectivity that limits model performance

7. Why might someone choose on-device AI over cloud AI for transcribing a lecture while on an airplane?

   1. Because local models are guaranteed to produce higher-quality transcriptions
   2. Because it works offline without requiring internet connectivity
   3. Because cloud AI services are more expensive for transcription tasks
   4. Because local models can understand accented speech better than cloud models

8. What is the primary benefit of reducing the per-call cost of AI to near zero after downloading a local model?

   1. The AI model will produce more accurate responses for free
   2. Users no longer need any internet connection whatsoever
   3. The model will run faster because it's not connecting to external servers
   4. Users can interact with the AI as frequently as they want without accumulating charges

9. What is the main purpose of tools like Ollama or LM Studio mentioned in the lesson?

   1. To provide access to the most advanced cloud-based AI models like GPT-4
   2. To enable running local AI models on personal computers
   3. To optimize web browsers for faster AI-powered search results
   4. To act as app stores where developers can sell AI-powered applications

10. A student wants to process sensitive medical notes through AI while ensuring the data never leaves their laptop. What does the lesson suggest?

    1. Run a local model, but verify the surrounding app doesn't send data to servers
    2. Use any popular AI app since they are all guaranteed to be private
    3. Only use AI services that charge a premium for privacy features
    4. Avoid AI entirely for sensitive data since no solution is truly private

11. What distinguishes 'frontier' AI models from those typically run on consumer hardware?

    1. Frontier models have far more parameters and require significantly more computational resources
    2. Frontier models can only run on smartphones, not laptops
    3. Frontier models are open-source and free to use by anyone
    4. Frontier models were all trained on outdated data

12. Why does latency matter for on-device AI applications?

    1. Higher latency actually improves the quality of AI responses
    2. Lower latency means faster response times, which improves user experience
    3. Latency only affects AI models that generate images, not text
    4. Latency measures how accurate an AI model's responses are

13. The lesson suggests comparing local model outputs to ChatGPT or Claude on the same prompts. What is the purpose of this exercise?

    1. To generate content that can be shared on social media
    2. To prove that local AI is always superior to cloud AI
    3. To find prompts that will cause the local model to crash
    4. To identify which tasks work similarly and which tasks show quality drops locally

14. What parameter range of models does the lesson say can currently run on consumer hardware?

    1. Under 1 billion parameters
    2. Over 1 trillion parameters
    3. 3 billion to 8 billion parameters
    4. 50 billion to 100 billion parameters

15. A user downloads a local AI model and runs it while their phone is in airplane mode. Later, they notice the AI app has a feature that sends usage statistics to the developers. What does the lesson say about this scenario?

    1. The privacy benefit is partially lost because the app sends data to external servers
    2. The user still has full privacy because the model runs locally
    3. The model will run slower because it's sending data in the background
    4. Airplane mode automatically blocks all data transmission from AI apps