Memory hook: Predict a pattern; generate content; constrain an action.
Must remember
AI is the broad field of machine-based intelligent behavior. ML learns patterns from examples; generative AI produces new content; an agent can select tools/actions toward a goal. Analytics describes and investigates data; business intelligence presents decision-oriented metrics. These capabilities overlap but solve different problems.
Use an existing API when a standard task such as translation, image analysis or speech recognition fits. A foundation model such as Gemini supports broader language/multimodal tasks. Gemini Enterprise Agent Platform brings model and agent development capabilities together; Agent Studio and supported AutoML workflows reduce different kinds of implementation work. BigQuery ML supports selected model workflows through SQL.
Training from scratch offers control at high cost and expertise requirements. Prompting adjusts instructions; retrieval supplies relevant context; tuning adapts supported model behavior with examples. Select the least complex approach that meets quality, privacy, latency and differentiation needs.
AI quality depends on accurate, complete, current, valid, unique and consistent data. Evaluate performance on representative unseen examples, including harmful or biased outcomes. Explainability helps people understand decisions; responsibility includes consent, security, fairness, human oversight and accountability. AI Hypercomputer combines compute such as GPUs/TPUs with software and consumption options, but hardware alone cannot fix poor data or evaluation.
Choose under exam pressure
| Requirement | Choice and reason |
|---|---|
| Standard language translation | A pre-trained Translation API. |
| Answer from private current documents | Authorized retrieval plus a suitable generative model. |
| Analyst wants supported ML using SQL | BigQuery ML. |
Traps
- Fluent output can be incorrect.
- An autonomous agent should not inherit unlimited user or administrator privileges.
Active recall
1. What distinguishes an agent?
It can select and execute tools/actions toward a goal, under controls.
2. Why try a pre-trained API first?
It may meet the need faster with less data, cost and engineering.
3. What does retrieval add?
Relevant external context at inference time.
4. Why evaluate unseen examples?
To estimate generalization rather than memorization of training data.
5. Why is human oversight important?
To govern consequential decisions and catch failures beyond automated controls.