Alibaba Cloud AI and Model Services
What to know
Alibaba Cloud's AI portfolio spans foundation models, managed model-serving platforms, and prebuilt AI APIs. Candidates should understand the Model Studio / DashScope-style model service (invoking foundation models via API), PAI (Platform for AI) for training, notebook, and model deployment, and prebuilt vision, speech, and NLP APIs for common tasks. Know the concepts that recur: tokens as the billing and context unit, prompt design, retrieval-augmented generation (RAG) to ground answers in your own data, fine-tuning to adapt a model, and inference endpoints for serving.
Understand the operational trade-offs: prebuilt API (fast, low effort), RAG (grounded, updatable knowledge, no retraining), fine-tuning (changes model behavior, needs labeled data), and training from scratch (rarely justified). Recognize data governance: your prompts and data handling, plus keeping sensitive data out of unapproved pipelines.
Traps
Questions conflate RAG with fine-tuning. RAG adds external knowledge at query time; fine-tuning changes weights. Tokens are not characters or words. A larger model is not automatically the right answer - latency and cost matter. "The model is always factually correct" is false; grounding and evaluation are required.
Rule of thumb
Need current or private facts? Use RAG. Need a consistent style or task behavior? Consider fine-tuning. Need a quick capability? Use a prebuilt API.
Worked example
A support bot must answer from an internal knowledge base that changes weekly. Answer: RAG over the knowledge base with a foundation model - updates require re-indexing, not retraining.