Fundamentals of Generative AI
What this domain tests
How foundation models are built and how their controls affect cost, latency, and quality - plus when Bedrock, SageMaker JumpStart, or a prebuilt generative feature is the right vehicle.
Core rules
- Transformers use self-attention to weigh relationships among all tokens at once. Tokens are subword units; the context window caps combined input and output length.
- Pre-training learns general patterns from massive text; fine-tuning adapts weights to a domain; instruction tuning and RLHF align responses with human preference.
- Embeddings are dense vectors where semantic similarity maps to geometric closeness - the basis of semantic search and RAG.
- Sampling controls: higher temperature and top-p increase randomness and creativity; lower them for deterministic extraction. Max tokens caps output length and therefore cost.
- Hallucination is fluent but unsupported output; mitigate with grounding (RAG), guardrails, and low temperature rather than longer prompts alone.
- Diffusion models generate images by iterative denoising; latent diffusion works in compressed space for speed.
Worked example
A team needs short, consistent JSON extracted from invoices at low cost. Reasoning: determinism matters, so temperature near zero; the task is narrow, so prompt engineering comes before fine-tuning; volume is high, so a smaller mid-tier model beats a frontier model on cost per token; grounding is unnecessary because the source text is supplied. Choose Bedrock with a small model, near-zero temperature, and a strict JSON schema in the prompt.
Traps
- Believing a larger model is always better.
- Confusing fine-tuning (changes weights) with RAG (changes context).
- Assuming Bedrock trains base models on your prompts - it does not.
- Forgetting that tokens drive both billing and context limits.