IVQ 651-700

---

Section 66: Embedded LLMs & Local Inference (10 Questions)

1. What are the tradeoffs of running LLMs in embedded systems vs. in the cloud?

2. How would you deploy a quantized LLM on a mobile device?

3. What’s the role of 4-bit or 8-bit quantization in edge deployment?

4. How do you minimize memory footprint without degrading generation quality?

5. What’s the difference between ONNX and GGUF for local model deployment?

6. How would you cache results effectively for offline GenAI tools?

7. How do local models like Phi-2 or TinyLLaMA compare to GPT-3.5 for constrained apps?

8. How can you do RAG with a vector DB on-device?

9. What architecture supports synchronized updates between edge and cloud models?

10. How do you design fallback mechanisms for low-power GenAI clients?

---

Section 67: Self-Improving & Adaptive Systems (10 Questions)

1. What is meta-prompting and how can it help GenAI self-improve?

2. How do LLMs assess the quality of their own outputs?

3. What is self-refinement in the context of LLMs?

4. How do you evaluate the effectiveness of self-critique loops in agents?

5. How can a model rewrite its own prompt to better serve a user?

6. What techniques allow models to learn from user outcomes without labeled data?

7. How would you implement prompt evolution based on continuous feedback?

8. What are policy-gradient style updates in self-improving agents?

9. How do autonomous agents identify areas where they need external help?

10. What are the safety concerns with fully self-improving GenAI systems?

---

Section 68: Domain-Specific Reasoning (10 Questions)

1. How do you tune an LLM for high-stakes financial advice scenarios?

2. What guardrails are needed for GenAI use in legal document drafting?

3. How would you validate medical LLM output for diagnosis support?

4. How do you design prompts for code generation in embedded systems vs. full-stack apps?

5. How would you evaluate an LLM's ability to summarize legal case law?

6. How can GenAI support pharma research documentation workflows?

7. What evaluation methods are best for academic LLM assistants?

8. How do you incorporate domain-specific taxonomies into GenAI models?

9. What are typical risks of using GenAI in scientific writing?

10. How do you extend GenAI to support patent drafting or technical IP filings?

---

Section 69: Voice, Video & Multimodal GenAI (10 Questions)

1. How does Whisper differ from traditional speech-to-text systems?

2. How do you chain audio transcription with LLM summarization workflows?

3. What are key considerations when adding TTS to a GenAI agent?

4. How do you align visual cues with LLM-generated video scripts?

5. What are good ways to caption real-time streams using GenAI?

6. How do you create GenAI workflows that mix voice, gesture, and text inputs?

7. What’s the role of VLMs (Vision Language Models) like Flamingo or GPT-4V?

8. How do you prompt image-generation models like DALL·E or Midjourney consistently?

9. What are video editing use cases where GenAI can save time or cost?

10. How would you generate a narrated video from a multi-step document?

---

Section 70: Future Architectures & Emerging Ideas (10 Questions)

1. What is the concept of “mixture of experts” and how does it scale LLMs?

2. How does the Recurrent Memory Transformer differ from vanilla Transformers?

3. What are state-space models (SSMs), and why are they promising?

4. How do you compare Mamba, RWKV, and Transformer-based models in real-time use?

5. What is FlashAttention and why does it matter for LLM performance?

6. How do sparsity techniques reduce compute without reducing accuracy?

7. What’s the role of retrieval-as-first-class-citizen in future GenAI stacks?

8. What is the importance of modularity in model and system design?

9. How will agent-based modeling evolve GenAI toward autonomy?

10. What is your prediction on the convergence of LLMs and cognitive architectures?

---