AI - progress, punctuated

Stories of speaking statues, oracles, golems, and mechanical beings placed knowledge or agency outside the individual human mind. They were not AI, but they reveal the enduring desire to consult an intelligence other than our own.

Long traditions of logic sought reliable forms of reasoning. George Boole’s nineteenth-century algebra of logic helped make propositions amenable to mechanical manipulation—a lineage that would later become fundamental to digital computation.

Babbage designed the Analytical Engine as a programmable general-purpose machine. Lovelace recognized that such machinery might manipulate symbols according to rules, not merely calculate quantities. The machine was not completed, but the conceptual separation of machine and program endured.

A universal machine could simulate any process expressible as a finite sequence of formal operations. It did not provide intelligence, but it established the general computational substrate on which later attempts could be built.

McCulloch and Pitts described simplified neurons using mathematical logic. Their model was radically incomplete as biology, but consequential as an idea: networks of simple units might produce complex computation.

Turing shifted attention away from an unknowable inner state toward observable behavior in conversation. The proposal did not define intelligence, but it created a durable interface through which people would later experience and debate machine capability.

Naming AI created an institution as much as a technical program. A shared banner helped organize researchers, funding, conferences, and expectations around the possibility of machine intelligence.

The perceptron adjusted its weights from examples. Its capabilities were narrow and its limitations substantial, but the demonstration established a recurring hope: intelligence might be learned rather than fully programmed.

Search, games, theorem proving, and early language systems produced striking results in carefully bounded environments. The difficulty was not that the systems did nothing; it was that their success did not transfer easily to the open world.

Systems such as MYCIN demonstrated that carefully encoded rules could perform impressively within narrow domains. The approach succeeded enough to reveal its own constraint: the labor and brittleness involved in keeping a large rule base correct.

Lisp machines and specialized vendors grew around expert systems. When cheaper general-purpose computers improved and expected returns failed to appear broadly, the surrounding commercial ecosystem contracted.

Backpropagation had important precursors. The influential 1986 work by Rumelhart, Hinton, and Williams demonstrated and popularized its use for learning internal representations in multilayer networks. Its later consequence depended on faster hardware, larger data, and additional methods.

Convolutional neural networks encoded useful assumptions about spatial structure and achieved practical results in handwritten-character recognition well before 2012. They waited for larger datasets and substantially more computation to transform computer vision.

Long short-term memory networks improved the ability of recurrent networks to learn dependencies across longer sequences. They later became central to speech, translation, and language systems before Transformers displaced recurrence in many large-scale applications.

The web supplied vast digital material. Programmable GPUs supplied efficient parallel arithmetic. ImageNet supplied a large labeled benchmark for visual recognition. Each was insufficient alone; together with older learning methods, they changed the scale of feasible experiments.

AlexNet paired a deep convolutional network with the ImageNet dataset and an efficient GPU implementation, winning the 2012 competition by a striking margin. It did not invent every ingredient; it made their combined consequence undeniable.

AlphaGo combined deep neural networks, tree search, and reinforcement learning to defeat Lee Sedol. Its importance was partly technical and partly cultural: it revised public intuition about which forms of judgment remained distant.

The Transformer dispensed with recurrence and convolution for sequence transduction, using attention mechanisms that were highly compatible with parallel hardware. Its eventual importance came from both algorithmic performance and its ability to exploit scale.

Large models trained over broad corpora displayed increasingly general task performance through prompting and adaptation. The improvement was real, but so were increasing concentrations of compute, data, capital, and environmental cost.

A conversational interface and post-trained model allowed millions of people to direct a general-purpose language system without specialized knowledge. The punctuation occurred in accessibility and cultural perception as much as in underlying architecture.