History and Evolution of AI

AI's prehistory begins with an attempt to formalize thought itself. In 1943, Warren McCulloch and Walter Pitts proposed that networks of simplified, all-or-none 'formal neurons' could, in principle, compute any expression a human could state in propositional logic, giving the first mathematical bridge between neural activity and symbolic reasoning (McCulloch & Pitts, 1943). A decade later, the field acquired both its name and its founding event: the 1955 proposal by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon for a 1956 summer workshop at Dartmouth College, which stated the conjecture 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,' and which coined the term 'artificial intelligence' for the new research area (Dartmouth proposal, 1955). The workshop's organizers went on to establish the dominant early paradigm, later called symbolic AI or GOFAI ('good old-fashioned AI'): the idea that intelligence could be produced by manipulating symbolic representations according to formal rules, an approach that Russell and Norvig trace through the founding decades of the field (Russell & Norvig, AIMA, ch. 1). Several of the Dartmouth workshop's participants and their students spent the following two decades building programs that embodied this symbolic bet directly, from early theorem-proving and search-based problem-solving programs to the rule-based systems described below, treating intelligence as fundamentally a matter of search over symbolic states guided by heuristics rather than brute computation alone.

Symbolic methods produced the first genuinely impressive demonstrations of machine reasoning within narrow, well-specified domains. The clearest commercial success was the expert system: a program that encodes a human specialist's domain rules and applies them to new cases. DENDRAL, begun in 1965 at Stanford by Edward Feigenbaum and geneticist Joshua Lederberg, inferred the molecular structure of organic compounds from mass-spectrometry data with expert-level accuracy, and is widely credited as the first working expert system (Britannica, 'Expert system'). Its success helped launch a wave of rule-based systems through the late 1970s and 1980s applied to medical diagnosis, mineral prospecting, and financial advising, each built on the same premise: that expertise could be captured as an explicit if-then rule base and executed by a general inference engine rather than hand-coded procedurally, with the rule base itself treated as the valuable, reusable asset separable from the underlying inference machinery. Vendors sold the inference engines themselves as reusable 'expert system shells' into which a company could plug its own domain rules, briefly making rule-based AI a genuine commercial product category rather than only a research demonstration.

That optimism collided with a harder reality by the early 1970s. In 1973, mathematician Sir James Lighthill delivered a report to Britain's Science Research Council concluding that AI research had largely failed to deliver on its most ambitious promises or to meaningfully advance understanding of the brain, a judgment that John McCarthy publicly disputed but that nonetheless led British funders to sharply cut AI research support (McCarthy's review of the Lighthill report, 1973). Comparable disillusionment followed in the United States as machine-translation and general-problem-solving projects underdelivered relative to their funding pitches. The resulting contraction in grants and industry investment became known as the first 'AI winter' — a pattern of boom followed by retrenchment that would recur roughly fifteen years later as the expert-systems market itself collapsed, and that remains a standing cautionary reference point whenever AI hype outruns delivered capability. Later historians have debated how much of the funding contraction to attribute to Lighthill specifically versus broader economic pressures of the period, but the report's role as a symbolic turning point in the UK's AI funding story is not seriously disputed.

While symbolic AI dominated funding and headlines, a parallel connectionist tradition pursued intelligence through learned, distributed representations rather than hand-coded rules. Frank Rosenblatt's 1958 perceptron proposed a probabilistic, trainable network model of how the brain might store and organize information, and it became the first widely studied learning machine capable of adjusting its own weights from examples (Rosenblatt, 1958). Interest in such networks cooled through the 1970s as their practical limits became apparent, but the approach was revived decisively in 1986 when David Rumelhart, Geoffrey Hinton, and Ronald Williams showed that the backpropagation algorithm could efficiently train multi-layer networks by propagating error gradients backward through hidden layers, giving connectionist models a general, scalable training method for the first time (Rumelhart, Hinton & Williams, Nature, 1986). This single algorithmic result underlies essentially every deep learning system built since, from the small multi-layer networks of the late 1980s to the billion-parameter architectures described later in this lesson, because every one of them is still, at bottom, adjusting its weights by propagating error gradients backward through its layers.

The following decades produced a sequence of highly publicized milestones that tracked the field's shifting balance between symbolic search, statistical learning, and raw computation. In 1997, IBM's Deep Blue defeated reigning world chess champion Garry Kasparov under standard tournament conditions, a result IBM itself frames as a turning point demonstrating that brute-force search paired with domain evaluation functions could outperform elite human intuition in a bounded, fully observable game (IBM, 'Deep Blue'). In 2011, IBM's Watson system defeated top human champions at the open-domain quiz show Jeopardy!, a feat IBM describes as a major advance for natural language processing because it required parsing ambiguous, punning natural-language clues rather than structured queries (IBM, 'Watson, Jeopardy! Champion'). The 2012 ImageNet competition marked deep learning's decisive return to prominence when a convolutional network dramatically cut the error rate on large-scale image classification (Krizhevsky, Sutskever & Hinton, 'AlexNet', 2012), and reinforcement-learning systems soon combined deep networks with trial-and-error search, as in the Deep Q-Network's human-level play across dozens of Atari games (Mnih et al., Nature, 2015) and AlphaGo's 2016 defeat of a professional Go player, achieved by combining deep neural networks with tree search in a game long considered too combinatorially vast for brute-force methods (Silver et al., Nature, 2016).

The most recent turning point was architectural rather than a single game victory. The 2017 Transformer architecture replaced recurrence and convolution with a purely attention-based mechanism for relating every position in a sequence to every other, and proved dramatically more efficient to train at scale (Vaswani et al., 'Attention Is All You Need', 2017). Transformers trained on broad, largely self-supervised data at massive scale gave rise to what a 2021 Stanford report terms 'foundation models' — models whose capabilities emerge from scale and broad pretraining and that can then be adapted to many downstream tasks, a shift the report characterizes as consolidating diverse AI applications onto a shared technical base (Bommasani et al., 2021). Most recently, these models have been wrapped in agent loops that interleave reasoning traces with external actions and observations, such as the ReAct framework, which lets a language model plan, call tools, and revise its plan based on what it observes — a concrete step from static prediction toward the autonomous, goal-directed agents this curriculum returns to throughout the AI major (Yao et al., 'ReAct').