Large Language Model Fundamentals

⏷ 6. What problem does self-attention solve?

⏷ 7. How do queries, keys, and values work in attention?

⏷ 8. Why do transformers use multiple attention heads?

⏷ 9. Why do transformers need positional information?

⏷ 10. How do encoder-only, encoder-decoder, and decoder-only transformers differ?

⏷ 11. What is causal masking?

⏷ 12. What is a context window, and what counts against it?

⏷ 13. How do input and output tokens affect LLM latency?

⏷ 14. How do greedy decoding, temperature, and top-p sampling differ?

⏷ 15. Why can the same prompt produce different answers?

⏷ 16. How do output limits, stop sequences, and end tokens work?

⏷ 17. Why can models struggle even when the input fits inside the context window?

⏷ 18. Why is context not the same as memory or model knowledge?

⏷ 19. What happens during language-model pretraining?

⏷ 20. How do base, instruct, chat, and reasoning-oriented models differ?

⏷ 21. How do instruction tuning and preference optimization differ?

⏷ 22. When should you use prompting, retrieval, or fine-tuning?

⏷ 23. How do full fine-tuning and parameter-efficient fine-tuning differ?

⏷ 24. What is model distillation, and when is it useful?

⏷ 25. What is quantization, and what does it trade?

⏷ 26. How do hosted and open-weight models differ?

⏷ 27. How do parameter count and model size relate to quality, latency, and cost?

⏷ 28. Why do language models hallucinate?

⏷ 29. How would you select an LLM for a production feature?

⏷ 30. How do KV caching and batching improve LLM serving?

⏷ 31. How do the prefill and decode phases of LLM inference differ?

⏷ 32. How do multi-head, multi-query, and grouped-query attention differ during inference?

⏷ 33. How does speculative decoding reduce generation latency without changing the target distribution?

⏷ 34. How do mixture-of-experts models work, and what do they trade?

⏷ 35. Which metrics describe LLM serving performance, and why is one latency number insufficient?