Topic overview
Model mechanics: tokenization, embeddings, decoding, context windows, fine-tuning, and RLHF.
Applied prompting: instruction and context design, zero-shot vs. few-shot, system vs. user prompts, structured output formats, and prompt injection as a security concern.
Core concepts
Temperature and sampling, grounding vs. hallucination, instruction following, retrieval as a grounding mechanism, model selection and cost/latency tradeoffs — plus, on the prompting side, output schemas (JSON mode, function-calling constraints), prompt versioning, and evaluating prompt changes against a labeled test set rather than eyeballing outputs.
Why it matters
The same model produces wildly different quality output depending on how a task is framed and constrained. LLMs are useful in production when paired with clear product constraints, strong evaluations, and prompts treated as versioned, tested artifacts — not one-off strings.
Interview relevance
Expect questions on model behavior, limitations, and failure modes (hallucination, context limits), plus what makes a prompt reliable at scale, how prompt injection differs from a normal user error, and how you'd evaluate a prompt or model change before shipping it.