Memory hook: Rules repeat instructions; ML learns patterns; agents take actions.
Must remember
Traditional rules execute explicit logic and may be best when decisions are stable and explainable rules are available. Supervised ML learns from labeled examples; unsupervised methods discover structure; reinforcement learning optimizes behavior through rewards. Generative AI produces content from learned patterns. An agent can choose tools/actions toward a goal, increasing both usefulness and control requirements.
Choose the approach from the problem: prediction/classification, anomaly detection, recommendations, extraction, generation or workflow assistance. A deterministic calculation does not need a language model. A high-stakes decision may need a human review boundary even when automation is technically possible.
Foundation models support many tasks but have limitations: incorrect answers, bias, sensitivity to instructions and bounded context. Tokens are processing units and influence capacity/cost; embeddings represent similarity for retrieval, not proof of truth. Prompting supplies instructions, retrieval adds external context and tuning adapts supported behavior. Training a model from scratch usually needs substantially more data, expertise and investment.
Evaluate build, buy and adapt choices across differentiation, time, data control, integration, cost and skills. A prebuilt solution may be appropriate for a common workflow; proprietary requirements may justify customization. Avoid confusing a convincing demonstration with reliable behavior on representative unseen work.
Choose under exam pressure
| Requirement | Choice and reason |
|---|---|
| Stable arithmetic/business rule | Conventional software or rules. |
| Predict a labeled business outcome | Suitable supervised ML. |
| Draft an answer using current private knowledge | Authorized retrieval plus generation and evaluation. |
Traps
- Generative AI is not automatically the best tool for every automation.
- A larger model is not guaranteed to meet the business requirement better.
Active recall
1. What distinguishes an agent?
It can select and execute actions/tools toward a goal.
2. What do embeddings represent?
A learned numerical representation useful for similarity and other tasks.
3. Prompting versus retrieval?
Instructions versus supplying relevant external context.
4. Why use a non-AI baseline?
To establish whether AI adds enough value to justify complexity and risk.
5. Why test unseen examples?
To estimate performance beyond the demonstration or training cases.