History & Philosophy of Artificial Intelligence
From ancient myths of artificial beings to modern transformer networks — trace the ideas, people, and moments that shaped what AI is today, and grapple with the foundational question: can machines truly think?
Where the Idea of Artificial Minds Began
The dream of artificial life predates computing by millennia. Ancient Greek myths described Talos, a bronze automaton protecting Crete, and Hephaestus forging golden maidens with intelligence and speech. Medieval alchemists sought a homunculus — a miniature artificial human. These weren't idle fantasies; they reflected a deep human intuition that intelligence might be substrate-independent, that mind might be separable from biology.
The philosophical groundwork was laid long before the first computer. René Descartes (1637) proposed that animals were merely complex automata, and raised the question of whether a sufficiently complex machine could be indistinguishable from a thinking being. Gottfried Leibniz imagined a calculus ratiocinator — a universal calculating machine that could resolve any dispute by computation. George Boole (1854) gave us the algebra of logic, showing that thought could be formalised in symbols and rules.
By the 19th century, Charles Babbage designed the Analytical Engine — a programmable mechanical computer — and Ada Lovelace wrote what many consider the first algorithm, while also raising a prescient caveat: the machine "has no power of originating anything. It can only do what we know how to order it to perform." This question — whether a machine can originate, create, or understand — remains the central philosophical problem of AI.
Can Machines Think? The Core Philosophical Debate
Turing's 1950 paper "Computing Machinery and Intelligence" reframed the question brilliantly. Rather than asking the hopelessly vague "can machines think?", he proposed a concrete behavioural test: if a machine can fool a human judge in a text-based conversation, we should attribute intelligence to it. This move from introspection to behaviour was revolutionary.
But the test has fierce critics. John Searle's Chinese Room (1980) is the most influential: imagine a person locked in a room, following rules to manipulate Chinese symbols they don't understand, producing correct outputs to Chinese inputs. They pass the test but understand nothing. Syntax, Searle argues, is not sufficient for semantics — rules without understanding are not intelligence.
The Frame Problem (McCarthy & Hayes, 1969) revealed another gap: how does an intelligent system know what doesn't change when an action occurs? Humans handle this effortlessly; classical AI systems collapsed under the combinatorial explosion. Modern LLMs sidestep the frame problem via statistical pattern matching, but whether they genuinely "solve" it remains philosophically contested.
The philosophical debates of the 1960s–80s are not obsolete. Every argument you will encounter about whether LLMs "really" understand language, whether AGI is possible, or whether AI systems can be conscious, traces directly back to Turing, Searle, Dreyfus, and McCarthy. Learn the original arguments — they are more sophisticated than most contemporary commentary.
The Three Major Paradigms of AI
1. Symbolic / Good Old-Fashioned AI (GOFAI)
The dominant paradigm from 1956–1980s. Intelligence is the manipulation of symbols according to explicit rules. Expert systems encoded human knowledge as IF-THEN rules. Strengths: interpretable, reliable within defined domains. Weaknesses: brittle outside training distribution, requires human-authored knowledge, cannot learn from data.
2. Connectionism / Neural Networks
Inspired by the brain. Intelligence emerges from the interaction of many simple units (neurons) with learned connection weights. Strengths: learns from data, generalises to new inputs, handles perceptual tasks. Weaknesses: black-box, requires vast data, computationally expensive, vulnerable to adversarial examples.
3. Statistical / Probabilistic AI
Intelligence as inference under uncertainty. Bayesian networks, Markov models, and probabilistic graphical models. Strengths: principled handling of uncertainty, integrates prior knowledge with data. Weaknesses: computational intractability for large models, requires careful model specification.
Modern AI is a hybrid — transformers are neural networks trained with gradient descent (connectionist) but often used with explicit prompting strategies (symbolic) and evaluated probabilistically. Understanding all three paradigms is essential for any AI practitioner.