What Is AI?
Core concepts, types of AI, and essential terminology for developers
Overview
Artificial intelligence (AI) is software that performs tasks typically associated with human intelligence—recognizing patterns, making predictions, generating text, or deciding among options. Most production systems today use narrow AI: models trained for specific tasks (spam detection, code completion, image classification). General AI—human-level reasoning across any domain—does not exist yet.
As a developer, you rarely build models from scratch. You integrate pretrained models via APIs or libraries, understand their inputs and outputs, and design applications around their strengths and limits.
Syntax / Usage
Key terms you will encounter in docs, job posts, and architecture discussions:
| Term | Meaning |
|---|---|
| Model | A learned function that maps inputs to outputs (e.g., text → label) |
| Training | Fitting model parameters on data |
| Inference | Running a trained model on new inputs |
| Dataset | Labeled or unlabeled examples used for training or evaluation |
| Feature | An input signal the model uses (pixels, words, numbers) |
| Hyperparameter | Settings chosen before training (learning rate, layer count) |
| Fine-tuning | Adapting a pretrained model to a narrower task with new data |
AI systems fall into broad categories:
- Rule-based / symbolic: explicit logic (if/else, expert systems)—predictable but brittle
- Machine learning: learns patterns from data—flexible but needs data and monitoring
- Deep learning: ML with neural networks—excels at images, text, audio at scale
Examples
Choosing an approach for a product feature:
Task: Classify support tickets by urgency
Rule-based: keyword lists ("urgent", "down") — fast to ship, misses nuance
Classical ML: train on past tickets with labels — good with structured features
LLM API: send ticket text, ask for category — flexible, higher cost/latency
A realistic stack for a SaaS app might combine all three: rules for obvious cases, a small classifier for volume, and an LLM for ambiguous tickets.
Common Mistakes
- Treating "AI" as one technology—APIs, local models, and classical ML behave very differently
- Expecting deterministic outputs from probabilistic models
- Skipping problem definition and jumping straight to "use GPT"
- Confusing automation (scripts) with intelligence (generalization from examples)
- Underestimating data quality, evaluation, and ongoing monitoring
See Also
machine-learning-basics large-language-models responsible-ai