Fundamentals of Generative AI
Free AWS Certified AI Practitioner (AIF-C01) practice — 6 questions on Fundamentals of Generative AI, with explanations. No sign-up. Full 12-question mixed test →
A financial services company is building a chatbot that must provide accurate regulatory compliance information. During testing, the model occasionally generates plausible-sounding but factually incorrect regulatory citations. The team has already implemented prompt engineering with specific instructions to only cite verified sources. Which combination of techniques would BEST address this hallucination issue while maintaining response quality?
RAG grounds the model's responses in verified external data sources, directly addressing hallucination by retrieving factual information before generation. Lower temperature (0.2) reduces randomness and creative fabrication, making outputs more deterministic and factual. This combination is the industry standard approach for accuracy-critical applications.
A media company needs to generate unique product descriptions for 500,000 items daily. Current costs using a foundation model API are $12,000/month. The descriptions follow predictable patterns but must maintain brand voice and avoid repetition. Which approach would MOST cost-effectively meet these requirements?
At 500K daily items with predictable patterns, fine-tuning a smaller open-source model (like Llama or Mistral) and self-hosting eliminates per-token API costs. Fine-tuning on brand examples maintains voice consistency. The high volume makes the upfront fine-tuning investment worthwhile, and ongoing costs become compute-only, drastically reducing the $12K/month API spend.
An AI system generates medical appointment summaries from doctor-patient conversations. During auditing, the team discovers the model occasionally includes demographic assumptions not present in the conversation (e.g., assuming a patient's occupation based on their name). Which technique specifically addresses this bias manifestation?
Constitutional AI allows defining explicit rules that the model must follow, such as 'never infer demographic information not explicitly stated.' This directly addresses the specific bias behavior of making unsupported assumptions. The technique uses the model itself to critique and revise outputs against defined principles, making it effective for this exact use case.
A research team is evaluating embeddings for a semantic search system across legal documents. They compare two models: Model A (384 dimensions, trained on general web data) and Model B (1024 dimensions, trained on legal corpus). Model A shows 72% retrieval accuracy while Model B shows 89% accuracy, but Model B's inference time is 4x slower. The system must handle 10,000 queries/hour with sub-200ms latency. What is the BEST optimization strategy?
This two-stage approach (fast Model A for initial retrieval, slower Model B for re-ranking only 100 results) balances accuracy and latency. Model A handles the 10K queries/hour volume quickly, then Model B's superior accuracy improves only the most promising results. Re-ranking 100 items is fast even with Model B. This hybrid approach is standard for production semantic search under latency constraints.
A company's generative AI application produces training materials. A user inputs: 'Create a Python tutorial using examples from the book Clean Code.' The model generates content that closely paraphrases passages from the copyrighted book. Which statement BEST describes the liability and mitigation approach?
The company deploying the AI application is responsible for its outputs, regardless of user prompts. Courts have held service providers liable for copyright infringement in generated content. Implementing content filtering, plagiarism detection, and citations is the appropriate mitigation. Companies like GitHub Copilot have added duplication detection for this reason.
A generative AI system for customer service must handle sensitive PII (personally identifiable information). The architecture team is debating data handling strategies. The system receives customer queries containing names, addresses, and account numbers, then generates personalized responses. Which approach BEST balances functionality and privacy?
PII tokenization (detecting PII via NER, replacing with tokens like [NAME_1], [ACCOUNT_1], processing the sanitized text, then substituting real values back) prevents PII from being sent to external model APIs or logged. This is the privacy-preserving approach recommended by NIST and used by companies like AWS Comprehend for PII protection. The model never sees actual PII.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$54.99 $17.99 with code FREETEST33 — valid through August 22.
Get my $17.99 deal →