Generative AI Fundamentals and Qwen Models
What to know
Generative AI here means large language models plus multimodal models (text, image, audio, video) accessed through APIs. Candidates should understand the core vocabulary: tokens (subword units that drive cost and context limits), context window (prompt + completion budget), temperature/top-p (randomness controls), and system vs. user vs. assistant message roles. Know that Qwen is Alibaba's model family, spanning different sizes and modalities, and that larger models generally improve reasoning while smaller/quantized models improve latency and cost.
Traps to watch
- Confusing training (fixed weights) with inference (serving). Parameters like temperature affect inference only.
- Assuming a bigger context window is always better: cost and latency scale with tokens.
- Mixing up fine-tuning (changing weights with labeled data) and RAG (retrieval at inference; no weight change).
- Believing temperature=0 guarantees identical outputs; it reduces but does not eliminate nondeterminism.
- Forgetting that multimodal models need modality-specific inputs (e.g., image URLs or base64).
- Treating "Qwen" as a single model rather than a family with size and modality variants.
Rule of thumb
Match model choice to the task: use a small model for classification/extraction at scale, a larger model for multi-step reasoning, and a multimodal model only when non-text input is required. Estimate cost as (input tokens + output tokens) × per-token price; if a prompt is long and reused, consider caching or trimming.
Worked example
A support bot needs to summarize 20-page documents. A 32k-context model fits a trimmed document; if not, chunk the text and summarize iteratively (map-reduce) rather than switching to a giant model. Set temperature low (0–0.3) for factual summaries.