The first working AI systems of the 1950s took two different routes. The Logic Theorist, written by Allen Newell, Herbert Simon, and Cliff Shaw in 1955 and 1956, proved theorems by searching with rules of thumb. The perceptron, built on Frank Rosenblatt's 1958 design, adjusted numeric weights from examples until it sorted inputs correctly. Both routes are still in use, and the history of each shows which early claims held up.
Key takeaways
- The Logic Theorist was specified in 1956 and first ran on a machine in late 1956 and early 1957, according to a 2026 historical reconstruction.
- A widely cited figure says it proved 38 of the first 52 theorems in Chapter 2 of Principia Mathematica.
- In July 1958, the U.S. Office of Naval Research unveiled a perceptron demonstration in which an IBM 704 learned to tell left-marked punch cards from right-marked ones after 50 trials.
- A single-layer perceptron cannot learn every pattern, and a method for training multiple layers was published in Nature in 1986.
The Logic Theorist: search guided by rules of thumb
Newell and Simon described the program in a 1956 RAND report titled "The Logic Theory Machine." Its task was to prove statements of symbolic logic taken from Whitehead and Russell's Principia Mathematica. Instead of trying every possible sequence of steps, which grows too fast to finish, it used heuristics, meaning rules of thumb that usually point toward a proof. A 2026 reconstruction by Jeff Shrager lists five of its methods: substitution, detachment (modus ponens), forward chaining, backward chaining, and sublevel replacement.
The program was not yet running on a computer when the Dartmouth summer workshop met in 1956. Newell and Simon worked parts of it by hand, with people acting as program components. Machine-executable proofs came on the RAND JOHNNIAC computer in late 1956 and early 1957. The often-quoted result, 38 of the first 52 theorems in Chapter 2, is tied to the 1957 runs, though Shrager notes that the figure is cited widely without a named origin. His own reanimation of the program from its transcribed code, run in a new interpreter, proved 16 of 23 attempted theorems.
The perceptron: weights adjusted by examples
Rosenblatt, a research psychologist at the Cornell Aeronautical Laboratory in Buffalo, New York, published "The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain" in Psychological Review in 1958. The paper proposed a network of simple units whose connection strengths change with experience. According to Cornell's own account, the Office of Naval Research unveiled the device in July 1958. An IBM 704, a machine weighing about five tons, was fed punch cards and, after 50 trials, learned to tell cards marked on the left from cards marked on the right.
Press coverage ran ahead of the evidence. Rosenblatt called the perceptron the first machine capable of having an original idea, a New York Times headline described a Navy device that learns by doing, and The New Yorker raised the prospect of a rival to the human brain. The demonstration itself showed something narrower: a program could change its own parameters to separate two kinds of input.
A perceptron learning a rule, step by step
This is an illustrative example using the standard textbook error-correction rule, not a reconstruction of the 1958 hardware. The unit outputs 1 when the weighted sum of its two inputs plus a bias is above zero, and 0 otherwise. It starts with both weights and the bias at zero and is shown the four input pairs in a fixed order, with the target being logical AND (output 1 only for the pair 1, 1). Whenever its output is wrong, each weight moves by the error times its input, and the bias moves by the error.
| Pass | Corrections made | Weights and bias at the end of the pass |
|---|---|---|
| 1 | 1 | 1 and 1, bias 1 |
| 2 | 3 | 2 and 1, bias 0 |
| 3 | 3 | 2 and 1, bias -1 |
| 4 | 2 | 2 and 2, bias -1 |
| 5 | 1 | 2 and 1, bias -2 |
| 6 | 0 | 2 and 1, bias -2 (no errors, learning stops) |
After 10 corrections over five passes, the unit gets all four cases right. With weights 2 and 1 and bias -2, only the pair 1, 1 gives a positive sum. The same rule can never settle on a correct answer for exclusive-or, where the output is 1 when exactly one input is 1, because no single straight line separates those cases from the others.
Why the two routes diverged
The Logic Theorist stored its knowledge as explicit rules written by people, so its steps could be read and checked. The perceptron stored its knowledge as numbers found by training, so it could handle examples nobody had described in advance but offered no readable account of why it answered as it did. That trade-off persists in current systems.
| Feature | Logic Theorist | Perceptron |
|---|---|---|
| Year | 1955 to 1956 design, 1957 runs | 1958 paper and demonstration |
| Builders | Newell, Simon, Shaw | Rosenblatt at Cornell Aeronautical Laboratory |
| Knowledge held as | Written logic rules and search | Numeric weights learned from examples |
| Demonstrated | Proofs of textbook logic theorems | Sorting punch cards by mark position |
| Main limit | Search explodes as problems grow | One layer cannot represent every pattern |
What happened to the perceptron
Cornell's history notes that the perceptron had only one layer and that Rosenblatt did not know how to train multilayer networks. Marvin Minsky and Seymour Papert's 1969 book Perceptrons criticized the approach, interest faded, and the episode helped start a period of reduced federal funding often called the AI winter. Rosenblatt died in 1971. In 1986, David Rumelhart, Geoffrey Hinton, and Ronald Williams described back-propagation in Nature, a procedure that lets hidden units in a multilayer network learn useful features, which the authors set apart from the earlier perceptron-convergence procedure.
FAQ
Was the Logic Theorist the first AI program?
It is widely described that way, and Shrager's 2026 paper calls it widely regarded as the first. The title depends on how AI is defined, so it is safer to say it was among the first.
Did the perceptron actually learn?
Yes, in a narrow sense. It changed its internal weights from labeled examples until it separated two classes, as the left versus right card demonstration showed. It did not understand the cards, and it could not learn patterns outside its single-layer limit.
Who invented AI?
No one person did. The Logic Theorist came from three people, the perceptron from Rosenblatt's work at Cornell, and the name of the field from the 1956 Dartmouth proposal by McCarthy, Minsky, Rochester, and Shannon.
Are these methods obsolete?
No. Heuristic search and rule-based reasoning remain in planners and checkers of logic, and trained networks descend from the perceptron idea of adjustable weights.
Related reading
- Innovators behind the birth of artificial intelligence
- How AI developed from its roots
- Sophia the robot and the gap between demo and reality
- Telling real AI from hype
More in AI Basics.
Last verified: October 2026
Sources
- The Logic Theory Machine, a complex information processing system, Newell A, Simon HA, RAND Report P-868, 1956
- Executable Archaeology: Reanimating the Logic Theorist from its IPL-V Source, Shrager J, arXiv 2603.13514, March 2026
- The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain, Rosenblatt F, Psychological Review 65(6), 1958
- Professor's perceptron paved the way for AI, 60 years too soon, Lefkowitz M, Cornell Chronicle, September 25, 2019
- Learning representations by back-propagating errors, Rumelhart DE, Hinton GE, Williams RJ, Nature 323, 1986
- Perceptrons: An Introduction to Computational Geometry, Minsky M, Papert S, MIT Press, 1969