Fundamentals of generative AI
Free Google Cloud Generative AI Leader practice — 6 questions on Fundamentals of generative AI, with explanations. No sign-up. Full 12-question mixed test →
A retail executive asks why their generative AI chatbot occasionally claims the company offers services they discontinued two years ago. The model was last trained in early 2023. What concept BEST explains this behavior?
Knowledge cutoff is the specific concept describing a model's inability to know facts or events that occurred after its training data collection ended. A 2023-trained model cannot know about 2024-2026 business changes without fine-tuning or retrieval-augmented generation (RAG).
A product manager wants to understand how a large language model processes the sentence 'AI transforms business' during inference. At the most fundamental level, what does the model convert this sentence into before generating a response?
Tokenization is the FIRST step: text is split into tokens (words, subwords, or characters) and each token is mapped to an integer ID. This happens before embeddings (which convert token IDs to vectors) or probability calculations. Leaders must understand tokens are the atomic units of LLM processing and directly impact costs and context limits.
An enterprise architect is comparing Google's Gemini models to earlier text-only LLMs. A stakeholder asks what 'multimodal' fundamentally means for their use cases. Which explanation is MOST accurate for a business leader?
Multimodal specifically refers to handling multiple MODALITIES of data (text, image, audio, video, code) as native inputs and outputs. Gemini 1.5 Pro, for example, can natively process a video and answer questions about it, or generate images from text—this cross-modality capability is what 'multimodal' means architecturally.
During a board presentation on responsible AI, a director asks how Google's principles specifically address the risk of generative AI producing harmful content. Which principle directly mandates proactive steps to prevent models from generating unsafe outputs?
Google's 'Be built and tested for safety' principle explicitly addresses designing safeguards to prevent harmful content generation (violence, hate speech, dangerous instructions). This includes safety classifiers, content filters, and red-teaming. It's the principle most directly about preventing unsafe outputs.
A CFO reviewing a generative AI initiative asks, 'What actually makes these models so expensive to create compared to traditional machine learning models we've built before?' Which factor is the PRIMARY driver of foundation model training costs?
Pre-training foundation models (GPT, PaLM, Gemini) requires massive compute: months on thousands of specialized accelerators (TPUs/GPUs) performing matrix operations across billions of parameters on trillions of tokens. This is the dominant cost—easily tens to hundreds of millions of dollars. Traditional ML models have far fewer parameters and train much faster.
A marketing leader asks their AI team to explain why a well-crafted prompt can dramatically change model outputs without retraining. What core transformer mechanism enables the model to dynamically adjust its response based on the specific prompt context provided?
The attention mechanism is the core innovation of transformers that allows models to consider the ENTIRE input prompt and conversation context when generating each token. Attention weights determine which parts of the prompt are most relevant to the current generation step—this is why prompt engineering works at inference time with no retraining. Leaders should understand attention as 'selective focus' across the context.
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.
$99.99 $34.99 with code FREETEST33 — valid through September 2.
Get my $34.99 deal →