Case Study · Docu-Series · In Production

Robot That
Can Think.

A four-part documented reconstruction of Frank Rosenblatt's perceptron — the 1958 machine that modern AI goes back to. Every scene is sourced. No photograph of the event it recreates is known to exist.

// EPISODE 1 · 62 SECONDS
// Every scene is a documented reconstruction — the source trail is below.

The morning machine learning made the papers.

July 7, 1958. In a government computer room in Suitland, Maryland, a 29-year-old psychologist borrows the Weather Bureau's IBM 704 and shows reporters a program that learns from its mistakes. The Navy tells the press the machine's descendants will walk, talk, see, and know they exist.

The next morning it runs on page 25 of the New York Times — and on the front of the Boston Globe as "Navy's Going to Build Robot That Can Think."

Read the Story ↓ Episode Guide ↓
Project · Robot That Can Think, four episodes Subject · Frank Rosenblatt's perceptron, 1957–1960 Method · Documented reconstruction, sources public Status · Episode 1 out now · Episodes 2–4 in development
the story

A machine learns a child's distinction — left from right.

Episode 1 reconstructs a single day: the July 1958 press demonstration where a borrowed IBM 704, running Rosenblatt's perceptron program, taught itself to tell left-marked punch cards from right-marked ones in about fifty tries.

The proof lived in a stack of punch cards and a simple program. Guess left or right. When the guess was wrong, the machine rewrote its own numbers; when it was right, it left them alone. After enough cards, the guesses stopped being wrong.

That error-correction rule is the seed of machine learning — published the same year in Rosenblatt's Psychological Review paper, and still recognizable inside every neural network running today.

The Navy's interest was practical: the Office of Naval Research funded the work hoping for a machine that could read aerial photographs. The press release promised more — a machine that would "walk, talk, see, write, reproduce itself and be conscious of its existence."

The claim outran the cardboard. That gap — between what the machine did and what the headline said — is the story the series tells across four episodes.

the record

The machine that learned left from right.

// Headline verified verbatim — Daily Boston Globe, July 8, 1958. Typesetting recreated; all other page text greeked by rule.

1957
Rosenblatt proposes the "photoperceptron" at Cornell Aeronautical Laboratory in Buffalo. The Office of Naval Research funds it as Project PARA.
July 7, 1958
A borrowed IBM 704 at the Suitland federal campus learns to tell left-marked cards from right in about fifty tries, in front of the press.
July 8, 1958
The New York Times, page 25: a machine expected to "walk, talk, see, write, reproduce itself and be conscious of its existence." The Boston Globe: "Navy's Going to Build Robot That Can Think."
1958
Psychological Review publishes the perceptron model — the error-correction rule, in print.
1958–1960
The Mark I Perceptron is built in Buffalo: 400 photocells in a 20×20 array — a machine with an eye.
1960
Cornell Aeronautical Laboratory shows the Mark I to the press. Its descendants are reading this page.
the method

Accuracy is the product.

Anyone can generate a period-looking room. The discipline is in what the series refuses to invent. Four rules govern every frame, and every episode ships with its public source trail.

RULE 01

Every scene has a source basis.

Each beat in the cut is licensed by a document — a newspaper report, a published paper, an archival photograph, a federal records collection. The shot book logs the source for every setup.

RULE 02

Only verified text is legible.

If period type appears on screen — a headline, a cabinet label, a diagram — it is either verified verbatim against the record or rendered deliberately illegible. No AI pseudo-text survives review.

RULE 03

Contested numbers stay off the voice track.

Where sources conflict — one paper says fifty training attempts, another says thirty to forty — the narration doesn't pick a winner. The conflict is logged, and the claim stays out of the script.

RULE 04

Reconstruction is disclosed, on screen.

Where no photograph of an event exists, the episode says so. The disclosure is part of the film, not a footnote on a website — because for a channel whose claim is accuracy, that card is the handshake.

// DISCLOSURE

Every scene is a documented reconstruction. No photograph of the July 1958 demonstration is known to exist. The rooms, the machines, and the moments are rebuilt from the written record — and the written record is linked below.

episode guide

Four episodes. One machine.

Episode 1

The Demonstration

July 1958, Suitland, Maryland. A borrowed IBM 704, a stack of punch cards, and the press conference that put "robot that can think" on the front page. The machine learns left from right — and explains, in its own voice, how.

▶ Watch on YouTube ↗
Episode 2

The Eye

Television said next week; the build took two years. The Mark I Perceptron rises in Buffalo — 400 photocells in a 20×20 array, weights turned by motors. The Navy builds a brain, and gives it an eye.

In development
Episode 3

Untitled

The series continues through the perceptron's rise, its critics, and the long winter that followed. Episode details are announced as each script clears fact-check.

In development
Episode 4

Untitled

The machine's descendants. Where the error-correction rule of 1958 ends up — and why the headline the Globe ran was less wrong than it looked, just sixty years early.

In development
sources & deeper reading

The public trail.

New York Times, July 8, 1958 — "New Navy Device Learns by Doing," p. 25. The demonstration, as reported the next morning. nytimes.com ↗
Rosenblatt, "The Perceptron" — Psychological Review, 1958. The model itself, including the error-correction rule the series dramatizes. psycnet.apa.org ↗
The Mark I Perceptron — the 1958–1960 build at Cornell Aeronautical Laboratory: 400 photocells, a 20×20 electric eye. wikipedia.org ↗
CAL Mark I press-conference records — University of Minnesota archives. The 1960 unveiling, in the primary record. archives.lib.umn.edu ↗
Cornell on Rosenblatt's legacy — "Professor's perceptron paved the way for AI — 60 years too soon." news.cornell.edu ↗
Suitland weather-computing history — NOAA heritage record of the federal campus where the borrowed IBM 704 lived. vlab.noaa.gov ↗

// Each episode's full beat-by-beat source log — including conflicts between contemporary reports and how the cut resolves them — is maintained in the project shot book and summarized in every episode description.

work with the studio

Have a story that needs this kind of rigor?

Larrydojo builds character-consistent, fact-checked AI video — from public-domain adaptation to documented historical reconstruction. See how the pipeline works in the Dick Tracy Origins case study, or start a project.

Dick Tracy Origins → Send a brief →