Artificial Intelligence

Artificial intelligence builds systems that perceive, reason, learn, and act. In practice, modern AI is dominated by machine learning — models trained on data rather than rules written by hand.

Overview

AI ranges from classical symbolic systems (rule engines, search, planning) through statistical ML to today’s deep neural networks and transformer-based foundation models.

Sub-Fields

• Machine learning (supervised, unsupervised, reinforcement).
• Deep learning (CNN, RNN, Transformer, diffusion).
• Natural language processing (NLP).
• Computer vision.
• Robotics & autonomous systems.
• Knowledge representation, planning, search.
• Reinforcement learning (PPO, SAC, DQN).

ML vs Deep Learning

• Classical ML — feature-engineered; linear/logistic regression, trees, SVM, gradient boosting (XGBoost, LightGBM).
• Deep learning — learns features; needs more data & compute (GPU/TPU).
• For tabular data, gradient-boosted trees still often beat deep nets.
• For images, text, audio — deep learning wins.

Large Language Models

• Transformer architecture (attention is all you need, 2017).
• Pre-train on huge corpora; fine-tune or use few-shot prompting.
• RAG — retrieve-then-generate for up-to-date / proprietary data.
• Function/tool calling, agentic workflows.
• Evaluation — benchmarks, golden sets, human review, LLM-as-judge.

ML Workflow

1. Define the problem & success metric.
2. Collect, label, and clean data.
3. Split train / validation / test.
4. Feature engineering / model selection.
5. Train; track experiments.
6. Evaluate; check fairness & drift.
7. Deploy; monitor for performance decay.

Tools & Standards

• Frameworks: PyTorch, TensorFlow, JAX, scikit-learn.
• LLMs: OpenAI, Anthropic Claude, Google Gemini, Meta Llama, Mistral.
• MLOps: MLflow, Weights & Biases, Kubeflow, SageMaker, Vertex AI.
• Data: pandas, Polars, DuckDB, Spark.
• Governance: NIST AI RMF, ISO/IEC 42001, EU AI Act.