Memory hook: Task, data, model, measure.
Must remember
- AI is the umbrella; ML learns patterns; generative AI produces new content. Supervised learning uses labelled examples, unsupervised learning discovers structure, and reinforcement learning learns from rewards.
- An LLM handles language tokens; multimodal models also consume or produce images, audio or video. Diffusion models progressively denoise a representation to generate content.
- Choose a model by modality, quality on representative tasks, context capacity, latency, cost, availability, deployment constraints and customization needs. A larger model is not automatically the best business choice.
- Gemini is Google's multimodal model family; Gemma provides open models; Imagen targets images; Veo targets video. Open model weights still have licensing and operational obligations.
- Structured tables, semi-structured JSON and unstructured documents need different preparation. Check completeness, accuracy, consent, relevance, freshness and accessibility before training or retrieval.
- Remember the stack: infrastructure supplies compute; models supply learned capability; platforms manage development; agents coordinate actions; applications deliver a user outcome. Ingest, prepare, train, deploy, monitor is a lifecycle, not a one-time launch.
Choose under exam pressure
| Requirement | Choice and reason |
|---|---|
| Classify labelled support tickets | Supervised learning; assess errors by class. |
| Produce a short product video | A video-capable model such as Veo, with rights and safety review. |
Traps
- A context window is capacity, not a guarantee that every fact will be used correctly.
- An open model does not eliminate hosting costs or make training data unrestricted.
Active recall
1. A report needs fresh sales figures. Retrain first?
No. Retrieve the current authorized figures and ground the response.
2. Why split evaluation data from training?
To measure generalization rather than memorization.
3. Which model family targets images?
Imagen; compare quality and editing requirements before choosing.
4. What is a token?
A unit of model input or output, often a word fragment, that affects context and cost.
5. A small model meets every acceptance test. Why use it?
It can meet the need with lower latency and cost.