Google Cloud Certified Generative AI Leader

This is where you build the vocabulary you later use to assess offerings and risks.

• Core terms from AI and machine learning to foundation model, LLM and prompt engineering
• Supervised, unsupervised and reinforcement learning, the ML lifecycle from data ingestion to model management
• Data quality and data types: structured or unstructured, labelled or unlabelled
• Choosing foundation models, the layers of the gen AI landscape and Google's models Gemini, Gemma, Imagen and Veo

In the largest section you sort Google Cloud's portfolio by purpose and audience.

• Google Cloud's strengths: TPUs (Google's own AI chips) and GPUs, an open approach, control over your data
• For everyday work: the Gemini app, Gemini Enterprise, Gemini for Google Workspace
• For customer experience: Agent Search and the Customer Engagement Suite with Conversational Agents and Agent Assist
• For development teams: Agent Platform with Model Garden, RAG offerings, custom agents with their tools and prebuilt AI APIs

This part shows where models go wrong and which lever corrects them.

• Limits of foundation models: knowledge cutoff, bias, hallucinations (plausible but false answers), edge cases
• Countermeasures: grounding in enterprise data, third-party data or Google Search, RAG (answers drawn from documents retrieved first), fine-tuning, human in the loop
• Prompt techniques: zero-, one- and few-shot, role prompting, prompt chaining, chain-of-thought, ReAct
• Parameters such as temperature, top-p and output length, ongoing monitoring and evaluation of models

Here you carry what you have learned into decisions in your organisation.

• Choosing the right gen AI solution for a specific need
• Steps to introduce generative AI in an organisation and ways to measure its impact
• Secure AI: security across the ML lifecycle, Google's Secure AI Framework (SAIF), Identity and Access Management (IAM), Security Command Center
• Responsible AI: transparency, privacy, bias, fairness, accountability and explainability