Foundation models, tokens, prompts, and how generative AI produces output
Domain Weight
Fundamentals of gen AI accounts for 27% of the GCP-GAIL exam.
This domain is the conceptual foundation of the exam. Questions frame generative AI in business terms: what it is, how it differs from earlier AI, and what its limits mean for a decision maker.
From AI to Generative AI
- Artificial intelligence is the broad field of machines performing tasks that normally need human intelligence
- Machine learning learns patterns from data instead of following hand-written rules
- Generative AI is a subset that creates new content such as text, images, audio, and code
- A foundation model is a large model pretrained on broad data and adapted to many tasks
- A large language model is a foundation model specialized for text and code
- Multimodal models accept and produce more than one type of content, such as text, images, and audio
Training
The model learns patterns from large data sets. Expensive, infrequent, and done by the model provider.
Inference
The trained model generates output for a prompt. Happens on every request and drives ongoing cost and latency.
| Model Type | Input to Output | Example Use |
|---|
| Text and multimodal LLM | Text and media to text | Summarize, draft, answer questions with Gemini |
| Image generation | Text to image | Create marketing visuals with Imagen |
| Video generation | Text or image to video | Produce short clips with Veo |
| Speech | Audio to text and text to speech | Transcribe and voice with Chirp |
| Embedding | Text to numeric vector | Semantic search, clustering, and retrieval |
How Models Generate Output
- A prompt is the instruction and context supplied to the model
- The model works in tokens, which are word fragments rather than whole words
- It predicts the next likely token repeatedly to build a response
- The context window is the maximum amount of text, in tokens, the model can consider at once
- Temperature and related settings control how random or focused the output is
- Hallucination is confident but incorrect output, because the model predicts plausible text rather than verified facts
- Embeddings turn text into vectors so that similar meanings sit close together in vector space
Grounding Reduces Hallucination
A model without external data answers from its training alone. Grounding supplies trusted content at request time so the answer can be tied to real sources and returned with citations.
Exam Focus Areas
- Generative AI predicts the next token, so it does not look up facts unless it is grounded in trusted data
- Embeddings convert text into vectors and are what make semantic search and retrieval work
- A larger context window lets you supply more information but does not by itself guarantee a correct answer
- For leaders the message is that value comes from data, use case, and adoption, not from the model alone
Practice This Domain
Test your understanding with free practice questions at /certifications/google-cloud/generative-ai-leader/practice — focus on: What generative AI is and how it differs from traditional ML, Large language models (LLMs): training, inference, tokens, and context windows, Foundation models and their capabilities (text, image, code, multimodal).
