The code before ChatGPT
A guided descent from the first running AI programs to the math and philosophy underneath modern machine learning — written as a linear exhibit, not academic papers.
Educational reference layout · CustomSite spec · Inspired by primary sources & standard histories of AI
Depth I — The modern pioneers
The first programs that behaved like reasoners
Alan Turing framed the question “Can machines think?” in 1950, but the first working AI program on a machine is widely credited to Allen Newell, Herbert A. Simon, and Cliff Shaw: the Logic Theorist (1955), which searched for proofs in Principia Mathematica using heuristics — not just rote steps.
Nearby “firsts” in the same era
- Arthur Samuel (1952) — checkers that improved with play; a practical early example of learning.
- Frank Rosenblatt (1957–58) — the perceptron: weights updated from mistakes — a direct ancestor of image models.
- Marvin Minsky & Dean Edmonds (1951) — SNARC, a physical neural-net maze experiment with vacuum tubes.
Watch: deeper context
Neurons as math
McCulloch & Pitts (1943) modeled a neuron as a threshold logic unit: networks of such units could, in principle, compute logical functions. That paper linked neuroscience vocabulary to computing.
Donald Hebb (1949) gave a learning principle — strengthen co-active connections — that later became the intuition behind weight updates in neural networks.
Watch: deeper context
When logic became voltage
Claude Shannon (1937) showed Boolean algebra maps cleanly onto switching circuits: the conceptual bridge from “truth values” to physical on/off states — the substrate of all digital machines.
Once reasoning could be encoded as rules over bits, software could carry “thought-like” procedures without magic — only disciplined engineering.
Watch: deeper context
Paper, gears, and the idea of a universal procedure
Ramon Llull experimented with combinatoric disks (Ars Magna, 13th c.) — a mechanical metaphor for generating combinations of concepts.
Leibniz pushed binary numeration and dreamed of a calculus of concepts; Ada Lovelace described algorithmic steps for the Analytical Engine and anticipated that machines might manipulate symbols beyond arithmetic.
Watch: deeper context
Formal systems & limits
Aristotle’s syllogisms gave a template for structured deduction; Pāṇini’s precise grammar shows rule systems older than silicon.
Kurt Gödel (1931) showed any sufficiently expressive formal system has true statements unprovable inside the system — a reminder that “perfect closure” is impossible, and why modern AI leans on probability, not only logic rules.
Watch: deeper context
Algorithms before Python
Mesopotamian tablets recorded procedural recipes; the word algorithm traces to Al-Khwarizmi’s systematic methods.
The I Ching’s broken/unbroken lines fascinated Leibniz as an ancient echo of binary representation — a cultural hint that humans long encoded structure into two symbols at a time.
Watch: deeper context
First distinction
At the bedrock, many traditions converge on one move: draw a boundary between this and not-this. Spencer-Brown’s Laws of Form famously calls that primordial act a “mark.” Without distinction, there is no information.
This exhibit is a narrative ladder, not a metaphysical claim. It is here to show how today’s AI stacks on centuries of logic, engineering, and honest limits — so clients see your studio can ship clear structure, not just pretty pages.
Watch: deeper context