Textbook of AI

Artificial Intelligence (AI) is the broad field of computer science concerned with building machines capable of performing tasks that, if done by a human, would be said to require intelligence. The field encompasses perception, reasoning, learning, planning, language understanding and action in the world.

Origin

The term was coined at the 1956 Dartmouth Workshop, organised by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon. The proposal famously asserted 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." The summer's participants, including Allen Newell, Herbert Simon, Ray Solomonoff, Trenchard More, Arthur Samuel and Oliver Selfridge, would shape the next two decades of the field.

Earlier landmarks include Alan Turing's 1950 paper Computing Machinery and Intelligence, which proposed the imitation game (now the Turing test) as an operational substitute for "can machines think?", and Warren McCulloch and Walter Pitts's 1943 model of neurons as logical threshold units.

Definitional axes

Russell and Norvig's textbook organises AI definitions along two axes, thinking versus acting and human-like versus rational, yielding four schools:

1. Thinking humanly, cognitive modelling, computational psychology.
2. Thinking rationally, logic, automated theorem proving.
3. Acting humanly, the Turing-test tradition.
4. Acting rationally, the rational agent that perceives its environment and chooses actions to maximise expected utility.

The "acting rationally" quadrant dominates contemporary research. An agent is formally a function from percept histories to actions; an environment is a transition function over states. The agent's task is to maximise some performance measure on this environment.

Subfields and approaches

AI is not a single technology but a family of paradigms:

• Symbolic AI / GOFAI, logic, search, expert systems, knowledge graphs.
• Statistical machine learning, regression, support vector machines, random forests, probabilistic graphical models.
• Deep learning, neural networks with many layers, the dominant paradigm since 2012.
• Reinforcement learning, agents that learn from reward signals.
• Robotics, embodied AI in physical systems.
• Natural language processing, speech, translation, language models.
• Computer vision, perception from images and video.

Modern frontier systems, GPT-4, Claude, Gemini, Llama, are foundation models trained on web-scale text and increasingly multimodal data. They blend several paradigms: deep neural architectures (the transformer), self-supervised pre-training, supervised fine-tuning, and reinforcement learning from human feedback (RLHF).

The AI effect

The AI effect describes the tendency of the field's perceived boundary to recede. Once a problem is solved, chess, optical character recognition, spell checking, route planning, it stops being called AI and becomes "just software." Pamela McCorduck observed: "every time somebody figured out how to make a computer do something … the chorus of critics was to say 'that's not thinking.'" The field's history is therefore one of constantly expanding ambitions and constantly reset goalposts.