The AI Toy

AI you could put your hands on

In the first three months of 1990, Kevin E. Martin gave readers of COMPUTE!'s Gazette a practical introduction to neural networks. His Future Computing: Neural Networks series paired explanations with programs for the Commodore 64. A home computer associated with games could become a small laboratory for learning from examples.

That is the significance of The AI Toy. On a machine that could be regarded as a “toy computer” by 1990, readers could train a network, test it, inspect its changing connections, and save what it had learned. Commercial magazine publishing put those experiences within reach of people outside university and corporate research labs.

The series gave its readers a concrete baseline for what AI could do: learn a relationship between inputs and answers, develop a compact internal representation, and discover groups in data. Those are substantial ideas to encounter through code running on your own computer.

January: how connections become knowledge

The first installment builds a vocabulary for neural networks. Artificial neurons receive signals through connections with adjustable strengths, called weights. Some connections encourage a response; others inhibit it. A pattern of activity represents an input, and the network's connections help determine its output.

Martin introduces association: presenting one pattern can bring back another. Think of learning to connect a particular shape with a particular answer. He distinguishes learning, when connections change, from retrieval, when the network uses its existing connections. He explicitly credits Donald Hebb's 1949 account of learning through changes in connections and introduces several families of neural networks.

This installment supplies the mental model for the experiments that follow. The reader can begin to understand where a network's learned behavior lives: in the strengths of its connections.

February: teach it an answer

The second installment makes that model runnable. A machine-language program, BP.ML, performs the calculations, while short BASIC programs let readers configure a network and supply its examples. Backpropagation works backward from an incorrect answer to calculate how the connection strengths should change.

In XOR, two input units feed two hidden units and one output unit. The task is easy to state: answer 1 when the two inputs differ and 0 when they match. The program trains on the four possible cases. The hidden layer gives the network a way to learn this relationship, a classic example of why the structure of a network matters.

ENCODE uses four input units, only two hidden units, and four output units. Information must pass through that narrow middle layer, encouraging a compact learned code. Its output targets rearrange the input patterns, so this particular exercise is best understood as learning an internal representation. It is a small, inspectable example of a central idea in later machine learning.

The reader also meets practical training choices: how large each adjustment should be, how previous adjustments affect the next one, and when the remaining error is small enough to stop. Learning becomes something to experiment with, observe, and improve.

March: let it discover groups

The third installment turns to competitive learning. The network receives examples without a supplied correct category for each one. Output units compete to respond to an input, and the winning unit adjusts its connections. With repetition, different units can specialize in different patterns.

DIPOLE uses the CL.ML engine. Its world is a four-by-four grid with two neighboring cells switched on at a time. The program presents 24 such patterns to a network with 16 inputs and two competing outputs. The article shows how training can produce groups associated with different regions of the grid.

Readers can inspect the connection weights to see how those groups take shape. They also encounter the limits of the method: the categories that emerge depend on the data and training, and useful organization is not guaranteed. This makes the experiment a lesson in both the possibilities and the practical uncertainty of learning systems.

Why a commercial magazine series belongs here

A history of AI includes the moments when people gained access to its ideas and tools. Martin's series was a significant piece of commercial AI education: three connected installments moved from an explanation of learning to working supervised and unsupervised examples, with BASIC interfaces and machine-language engines that readers could study and use.

The scale is part of the story. Small networks made the relationship between examples, weights, training, and results visible. A reader could follow an entire experiment rather than treat the computer's answer as a mystery. The C64's limitations helped make the mechanics concrete.

For a reader in 1990, that experience could change the meaning of “artificial intelligence.” It offered a grounded set of expectations: a machine could learn from examples, represent information internally, and organize patterns. Time, training choices, and the design of the problem still mattered. This Atlas includes the series as a milestone in making neural-network research accessible through a commercial publication and usable software.

Earlier foundations, continuing ideas

The series connects to a longer research history. McCulloch and Pitts's mathematical neurons, Hebb's learning hypothesis, Hopfield's associative networks, and the influential 1986 backpropagation paper provide useful points of reference. Martin's March reading list directs readers to research collections and the Parallel Distributed Processing books.

Later systems developed these broader ideas at much greater scale. Learning-based document recognition and AlexNet show the growth of networks trained from examples. Latent diffusion uses a learned compressed image representation as part of an image-generation system. These are connections between ideas; the evidence examined for this Atlas does not establish that those later works directly drew on Martin's series.

Explore the articles and code today

Payton Byrd's NeuralNetworks repository provides a place to explore the series and its programs. Start with the combined edition for the collected articles, C64 programs, annotated source, and documented errata. Read the repository's current instructions and validation notes before running an experiment.

The original magazine scans and disk archive remain linked alongside this account. Together they let a new reader follow the same progression: understand the model, try the examples, examine what the network learned, and ask what to change next.