1958: The Perceptron makes headlines

On 8 July 1958, the New York Times ran a story that would become one of the most famous (and most hyped) announcements in the early history of artificial intelligence. Under the headline "New Navy Device Learns By Doing," the paper described the Perceptron, a pattern-recognition machine developed by psychologist Frank Rosenblatt at the Cornell Aeronautical Laboratory with funding from the U.S. Office of Naval Research.

The Perceptron was inspired by simplified models of neurons. Fed with input from a grid of photocells, it could adjust the weighted connections between simulated neurons until it learned to distinguish simple shapes, such as telling a square from a triangle. Rosenblatt built it first as a program running on an IBM 704, and later as dedicated hardware called the Mark I Perceptron, complete with a wall of wires and potentiometers standing in for adjustable synapses.

What made the press coverage so notable was its tone. The Navy suggested the device might one day walk, talk, see, write, reproduce itself, and be conscious of its own existence. Such claims were wildly premature, and the backlash came within a decade: in 1969, Marvin Minsky and Seymour Papert published "Perceptrons," a rigorous critique showing the mathematical limits of single-layer networks. Funding and enthusiasm for neural networks cooled sharply, contributing to what historians later called an AI winter.

Yet the core idea Rosenblatt championed did not disappear. It resurfaced, layered and reworked, as the backpropagation-trained multilayer networks of the 1980s, and again, decades later, as the deep learning systems now embedded in everyday life.

• The 1958 Perceptron: a single layer, hand-tuned weights, simple shape recognition
• Today's neural networks: billions of parameters, trained on vast datasets, powering translation, image generation, and conversation

Looking back from today's AI landscape, the story of 8 July 1958 is a useful reminder that breakthroughs and overclaiming often travel together. The Perceptron's basic architecture, adjustable weighted connections that learn from examples, remains at the heart of modern machine learning, even as the field continues to relearn the lesson that genuine progress rarely matches the boldest headlines.