Artificial intelligence is the field of computer science dedicated to creating systems that can perform tasks normally requiring human intelligence — recognizing speech, understanding language, making decisions, and learning from experience. Unlike traditional software that follows explicit if/then rules, AI systems learn patterns from data and generalize to new situations they haven’t seen before.

In the summer of 1956, John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon organized a workshop at Dartmouth College in Hanover, New Hampshire. They proposed that “every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.” McCarthy coined the term “Artificial Intelligence” specifically for this proposal, choosing it for its neutrality over existing terms like “cybernetics.”

All AI systems that exist today are narrow AI — they excel at one specific task but cannot transfer that ability to other domains. A chess engine that beats grandmasters cannot hold a conversation. A language model that writes essays cannot drive a car. Narrow AI is powerful within its domain but brittle outside it.

Image recognition: Classifying photos (Google Photos)
Language models: ChatGPT, Claude, Gemini
Recommendation engines: Netflix, Spotify, YouTube
Game playing: AlphaGo, Stockfish
Self-driving: Tesla Autopilot, Waymo

Artificial General Intelligence would match human-level intelligence across all cognitive tasks — reasoning, learning, creativity, social understanding — without being limited to a specific domain. AGI could learn any new task the way humans do, transferring knowledge between domains. AGI does not yet exist and remains an active research goal with no consensus on when (or if) it will be achieved.

A hypothetical level beyond AGI where machine intelligence surpasses all human cognitive abilities. Purely theoretical — no serious researcher claims we are close to ASI.

No memory, no learning from past experience. The same input always produces the same output. IBM’s Deep Blue (1997) evaluated 200 million chess positions per second but couldn’t remember previous games or improve over time.

Can use past data and recent experience to inform decisions. This is where most modern AI lives. Self-driving cars observe other vehicles over time. LLMs use context windows. These systems learn during training but have limited ability to update in real-time.

Would understand emotions, beliefs, and intentions of other agents. Could predict behavior based on mental states. Does not yet exist — though some researchers argue LLMs show early, limited signs of modeling others’ perspectives.

Would possess consciousness and self-awareness — an understanding of its own existence and internal states. Purely theoretical. No AI system has demonstrated anything approaching genuine self-awareness.

Represents knowledge using explicit symbols and rules. Intelligence emerges from logical reasoning over structured representations. Dominated AI from the 1950s through the 1980s. Produced expert systems like MYCIN (medical diagnosis, 1976) and DENDRAL (chemical analysis, 1965).

Strengths: Transparent, interpretable, works with small datasets, excellent at logical reasoning
Limits: Requires hand-coded rules, brittle in ambiguous situations, doesn’t scale to complex real-world problems

Represents knowledge through weighted connections between artificial neurons, inspired by the brain. Intelligence emerges from learning patterns in data. Struggled in the 1970s–80s but exploded after 2012 with deep learning. Powers all modern AI breakthroughs: image recognition, language models, game playing.

Strengths: Learns from raw data, scales to massive problems, handles ambiguity and noise
Limits: Requires huge datasets, computationally expensive, “black box” — hard to interpret why it makes specific decisions

In his 1950 paper “Computing Machinery and Intelligence,” Alan Turing proposed the “Imitation Game”: a human interrogator communicates via text with two hidden entities — one human, one machine. If the interrogator cannot reliably tell which is which, the machine demonstrates intelligence. Turing argued this behavioral test sidesteps the “meaningless” question of whether machines truly “think.”

John Searle challenged the Turing Test with a thought experiment: imagine a person in a room who follows rulebooks to manipulate Chinese symbols, producing correct responses without understanding Chinese. Searle argued that passing the Turing Test doesn’t prove understanding — a system can manipulate symbols syntactically without any semantic comprehension.

We are in an era of narrow AI that is extraordinarily capable within specific domains. Large language models can write code, translate languages, and pass professional exams. Diffusion models generate photorealistic images. RL agents master complex games. Yet every one of these systems is narrow — brilliant at its task, unable to generalize beyond it.

1. AI is the science of making machines that learn from data
2. All current AI is narrow (task-specific), not general
3. Two paradigms: symbolic (rules) vs. connectionist (learning)
4. Connectionism (neural networks) dominates modern AI
5. The Turing Test measures behavior, not understanding

Ch 2: History of AI — The full timeline from Turing to transformers

Ch 3: ML Paradigms — Supervised, unsupervised, and reinforcement learning

Ch 5: Perceptrons — How neural networks actually work

Ch 9: Transformers — The architecture behind modern AI

Ch 10: LLMs — How ChatGPT and Claude actually work