AI systems are categorized by capability (reactive vs learning), modality (text, vision, speech), and deployment (batch, real-time, on-device). Taxonomy helps you pick the right tool and set expectations.
Capability spectrum
- Reactive — chess engines evaluating positions; no memory of past games unless programmed
- Learning — models updated from new data (recommendations, fraud scores)
- Interactive agents — chatbots combining retrieval, tools, and policies
By modality
| Modality | Example tasks |
|---|---|
| Tabular | Churn prediction, credit risk |
| Text (NLP) | Search, sentiment, summarization |
| Vision | Face detection, defect inspection |
| Speech | Transcription, voice commands |
| Multimodal | Image + caption, document QA |
Deployment shapes
# Deployment modes (conceptual labels)
modes = {
"batch": "nightly scoring job",
"online": "API per user request",
"edge": "on phone or camera",
}
print(modes["online"])Practice: Optional pseudocode only—run locally in Jupyter if helpful. No model training required for this literacy track.
Latency, privacy, and update frequency drive batch vs online vs edge choices.
Important interview questions and answers
- Q: Tabular vs NLP AI?
A: Same ML ideas—features differ (columns vs tokens/embeddings). - Q: Edge AI trade-off?
A: Lower latency and privacy; harder to update models centrally.
Self-check
- Name three modalities and one task each.
- What is reactive vs learning AI?
Tip: Modality (text vs vision) drives tooling more than the word "AI" on a slide.
Interview prep
- Reactive vs learning?
- Reactive evaluates current state; learning systems update from new data.
- Edge deployment trade-off?
- Lower latency and privacy; harder centralized updates.