Today “AI” usually means machine learning, and especially deep learning and large language models. But the classical ideas in these notes — search, constraints, logic, probability, decision theory — are still inside modern systems, often combined with learned components.
How the pieces fit
| Paradigm | Knowledge comes from | Core techniques | Where to read more |
|---|---|---|---|
| Search & planning | A problem formulation | BFS/A*, alpha-beta, MCTS, STRIPS planners, CSP solvers | This subject |
| Logic & knowledge | Hand-written facts and rules | Resolution, chaining, ontologies, knowledge graphs | This subject |
| Probabilistic reasoning | Expert-built or learned models | Bayesian networks, HMMs, MDPs | This subject |
| Machine learning | Labelled/unlabelled data | Regression, trees, SVMs, clustering, evaluation | Machine Learning, Data Science |
| Deep learning | Large datasets + gradient descent | CNNs, RNNs, Transformers, generative models | Deep Learning |
| Generative AI / LLMs | Web-scale pretraining + fine-tuning + RLHF | Prompting, RAG, agents, evaluation | Generative AI |
| Reinforcement learning | Interaction and reward | Q-learning, policy gradients, self-play | This subject (intro) + Deep Learning |
Classical AI inside modern systems
| Classical idea | Still used in |
|---|---|
| A* and graph search | Maps and ride-hailing routing, game path-finding, robot motion planning |
| Minimax / MCTS | Chess engines (Stockfish: alpha-beta + a neural evaluation), AlphaZero, game AI |
| Constraint solving & SAT | Airline crew and timetable scheduling, chip verification, package managers' dependency resolution |
| Planning (PDDL/HTN) | Logistics, space missions, NPC behaviour in games, LLM agents that decompose tasks |
| Logic & knowledge graphs | Search-engine knowledge panels, business rules engines, grounding LLM answers (RAG over graphs) |
| Bayesian reasoning & HMMs | Spam filters, medical decision support, Kalman/particle filters in robots and GPS |
| MDPs & RL | Recommendation, ad bidding, robotics, and RLHF for aligning LLMs |
| Beam search | Decoding text in translation and speech models |
Natural language processing
| Era | Approach | Example |
|---|---|---|
| 1960s–80s | Hand-written rules and grammars | ELIZA, SHRDLU |
| 1990s–2000s | Statistical: n-gram language models, HMMs, CRFs | Statistical machine translation, POS tagging |
| 2013–2017 | Word embeddings + RNN/LSTM sequence models | word2vec, seq2seq translation |
| 2018–now | Pretrained Transformers, then LLMs | BERT, GPT family, instruction-tuned chat models |
The idea behind every language model is the same — predict the next word from the context. A bigram model is the smallest version:
from collections import Counter, defaultdict
corpus = "the cat sat on the mat . the cat ate the fish . the dog sat on the log .".split()
bigrams = defaultdict(Counter)
for w1, w2 in zip(corpus, corpus[1:]):
bigrams[w1][w2] += 1
def next_word_probs(word):
counts = bigrams[word]
total = sum(counts.values())
return {w: round(c / total, 2) for w, c in counts.most_common()}
def generate(word, n=5):
out = [word]
for _ in range(n):
word = bigrams[word].most_common(1)[0][0] # greedy decoding
out.append(word)
return " ".join(out)
print(next_word_probs("the")) # {'cat': 0.33, 'mat': 0.17, 'fish': 0.17, 'dog': 0.17, 'log': 0.17}
print(generate("dog")) # dog sat on the cat sat ← greedy decoding soon loopsAn LLM replaces the one-word context with thousands of tokens and the count table with a Transformer of billions of parameters — see the Deep Learning (Transformers) and Generative AI notes.
| NLP task | Input → output |
|---|---|
| Classification (sentiment, spam, intent) | Text → label |
| Named-entity recognition, POS tagging | Text → label per token |
| Machine translation, summarisation | Text → text |
| Question answering, retrieval | Question (+ documents) → answer |
| Speech recognition / synthesis | Audio ↔ text |
Computer vision
| Task | Output | Typical models |
|---|---|---|
| Image classification | One label per image | CNNs (ResNet), Vision Transformers |
| Object detection | Boxes + labels | YOLO, DETR |
| Segmentation | A label per pixel (semantic / instance) | U-Net, Mask R-CNN, SAM |
| Tracking, pose, depth | Trajectories, keypoints, 3-D structure | Specialised CNN/Transformer heads |
| Generation & editing | New images | Diffusion models |
| Vision–language | Captions, answers about images | CLIP-style encoders, multimodal LLMs |
Vision went through the same shift as NLP: hand-designed features (edges, SIFT, HOG) + a classifier, replaced after 2012 by features learned end-to-end by CNNs, and more recently by Transformers and multimodal models.
robotics · speech · recommender systems · autonomous vehicles · multi-agent systems · AI for science (protein folding) · neuro-symbolic AI
Is classical (symbolic) AI obsolete now that we have deep learning?
What is the difference between narrow AI systems today and AGI?