2.4. Reflection Checkpoint: Fundamentals of Generative AI

Key Takeaways

• Foundation Models are the umbrella category; LLMs specialize in text and Diffusion Models specialize in images — all are pre-trained on massive datasets, then optionally fine-tuned for a specific task.
• Models process language as tokens (the billing unit) and represent meaning as embeddings (vectors); prompt engineering is how you steer output quality without retraining anything.
• Generative AI's core advantage — a single model adaptable to many tasks via natural language — comes paired with its core risk: outputs aren't grounded in verified fact, so hallucinations, nondeterminism, and bias are inherent, not edge cases.
• Choosing a model means balancing performance, modality, latency/size, compliance, cost, and fine-tuning support against the specific task — a real-time chatbot and an offline legal-summarization tool have opposite priorities.
• AWS's generative AI stack has a clear division of purpose: Bedrock for consuming a choice of managed Foundation Models via API, SageMaker JumpStart for deploying a specific open-source model into your own environment, Amazon Q as a ready-made assistant application, PartyRock for no-account experimentation, and Agents for Bedrock when the model needs to take multi-step actions rather than just generate text.

Connecting Forward

You now understand what generative AI is, its trade-offs, and which AWS service fits which need. Phase 3 goes deeper into actually building with foundation models — the design considerations, prompt engineering techniques, and training/fine-tuning choices that turn a general-purpose model into a working application.

Self-Check Questions

• Your customer support chatbot confidently tells a user about a return policy that doesn't exist. Which generative AI risk is this, and which technique (covered in the next phase) exists specifically to ground responses in verified facts?
• One team wants to experiment with a generative AI idea before committing to an AWS account, and another wants to fine-tune a specific open-source model they found online. Which AWS service fits each, and why wouldn't Bedrock alone cover both?