OBJECT ROOM / FOUNDATIONAL RECORDS
Follow the records that changed how intelligence was built.
A museum object is not another definition. It is a record made at a particular moment: a machine, paper, proposal, research book, or release page that lets us inspect what people actually built and claimed then.
Records and commentary remain distinct from originals. Any reproduced photograph carries its own creator, source, rights basis, and change notice.
GUIDED PATH / 12 TURNS
From written rules to learned representations
This is not a march toward one inevitable machine. Twelve selected records separate formal rules, physical learning machinery, bounded perception–planning–action, encoded expertise, learned representations, specialised computation, memory, scale, shared control, search, architecture, and public access.
- 01 01 / FORMALISE 1943
A logical calculus of the ideas immanent in nervous activity
Idealised neural events become logical operations.
- 02 02 / BUILD A LEARNER 1958
Electronic Neural Network, Mark I Perceptron
A room-scale electronic assembly gives the perceptron a physical form.
- 03 03 / CONNECT PERCEPTION TO ACTION 1969
Shakey
A wheeled body, radio link, and remote computer close a bounded planning loop.
- 04 04 / ENCODE EXPERTISE 1984
Rule-Based Expert Systems: The MYCIN Experiments
A research programme writes domain knowledge as inspectable rules.
- 05 05 / LEARN REPRESENTATIONS 1986
Learning representations by back-propagating errors
Weights change as output error travels backward through the network.
- 06 06 / SPECIALISE COMPUTATION 1997
Deep Blue custom chess chip, version 2
A custom chip turns bounded chess search into dedicated hardware.
- 07 07 / HOLD CONTEXT 1997
Long Short-Term Memory
A gated recurrent method is designed to keep error signals available across long intervals.
- 08 08 / SCALE AN EXPERIMENT 2012
ImageNet Classification with Deep Convolutional Neural Networks
A deep convolutional network meets ImageNet and a GPU implementation.
- 09 09 / SHARE CONTROL 2014
SAFFiR aboard ex-USS Shadwell
A humanoid research platform tests perception and action while an operator remains able to intervene.
- 10 10 / LEARN AND SEARCH 2016
Mastering the game of Go with deep neural networks and tree search
Policy and value networks guide tree search inside the game of Go.
- 11 11 / CHANGE ARCHITECTURE 2017
Attention Is All You Need
The paper’s encoder–decoder uses attention without recurrence or convolution.
- 12 12 / REACH THE PUBLIC 2022
Introducing ChatGPT
A research preview turns model interaction into a public conversational interface.
-
01 PAPER / AIM-DOC-1943-001
A logical calculus of the ideas immanent in nervous activity
A formal model that treated idealised neural events as logical propositions.
Why it is here
Under explicit “all-or-none” assumptions, McCulloch and Pitts described neural events and their relations with propositional logic, then related qualifying logical expressions to possible nets.
What to notice
Start with the assumptions. This is neither a literal map of the brain nor a trainable modern neural network; it is a deliberately simplified calculus.
- Creator
- Warren S. McCulloch · Walter Pitts
- Date
- 1943-12
- Publisher / holder
- The Bulletin of Mathematical Biophysics, Springer Nature
- Display note
- Metadata and commentary only
-
02 PAPER / AIM-DOC-1950-001
Computing Machinery and Intelligence
A paper that replaced an untestable question with an observable game.
Why it is here
Turing did not settle whether machines think. He proposed a practical way to examine machine behaviour through language, and then worked through objections to that proposal.
What to notice
Notice the move from defining “thinking” to describing a testable encounter. That change of method is the object’s lasting importance.
- Creator
- Alan M. Turing
- Date
- 1950-10
- Publisher / holder
- Mind, Oxford University Press
- Display note
- Metadata and commentary only
-
03 PHYSICAL INSTRUMENT / AIM-ROB-1950-001
Grey Walter cybernetic tortoise
A compact 1950–1952 cybernetic machine connected light and touch responses to motion without a general-purpose digital computer.
Why it is here
The Science Museum Group records a photoelectric cell, movement around floors, obstacle and slope avoidance, and different responses to moderate and bright light. The object makes an early cybernetic argument tangible: a few feedback relationships could produce behaviour that looked richer than any single response, without demonstrating learning, deliberation, or general intelligence.
What to notice
Do not identify this catalogue entry as an original Elmer or Elsie. A 2008 University of Bristol account distinguishes a surviving original at UWE from a replica in the Science Museum, while the current catalogue dates L2000-4441 to 1950–1952 and credits UWE Bristol; these records do not fully resolve the object's identity or current display status. The catalogue image is not reproduced because its CC BY-NC-SA licence restricts commercial use.
- Creator
- William Grey Walter · Burden Neurological Institute
- Date
- 1950
- Publisher / holder
- Science Museum Group Collection
- Collection
- Science Museum Group Collection; credit UWE Bristol · L2000-4441
- Physical size
- Overall: 260 mm × 265 mm × 360 mm × 2.524 kg
- Display note
- Metadata and commentary only
-
04 PROPOSAL / AIM-DOC-1955-001
Proposal for the Dartmouth Summer Research Project on Artificial Intelligence
The project proposal that put “artificial intelligence” in its title.
Why it is here
The organisers proposed a two-month study for the summer of 1956. Its agenda joined language, abstraction, neural networks, creativity, and self-improvement under one research programme.
What to notice
Read it as a proposal, not a victory report. The confidence, open problems, participant plans, and budget all reveal a field being organised before its outcomes were known.
- Creator
- John McCarthy · Marvin Minsky · Nathaniel Rochester · Claude Shannon
- Date
- 1955-08-31
- Publisher / holder
- John McCarthy archive, Stanford University
- Display note
- Metadata and commentary only
-
05 PHYSICAL INSTRUMENT / AIM-EQP-1958-001
Electronic Neural Network, Mark I Perceptron
An electronic assembly built at the Cornell Aeronautical Laboratory gave Rosenblatt’s perceptron a physical form.
Why it is here
The Smithsonian catalogue dates the Mark I Perceptron to 1958 and describes sensory, association, and response units distributed across cabinets, switches, a plugboard, potentiometers, and meters. The surviving equipment makes an abstract learning proposal inspectable as a manufactured system.
What to notice
This is not a modern deep network or evidence that the machine fulfilled contemporary predictions about human-like intelligence. The Smithsonian record is still subject to revision, says the object is not currently on display, and applies separate reuse conditions to its images; no image is reproduced here.
- Creator
- Frank Rosenblatt · Cornell Aeronautical Laboratory team
- Date
- 1958
- Publisher / holder
- National Museum of American History, Smithsonian Institution
- Collection
- National Museum of American History, Smithsonian Institution · 1987.0819.01
- Physical size
- part 1: 199.4 × 194.3 × 64.1 cm · part 2: 200.6 × 146 × 64.7 cm
- Display note
- Metadata and commentary only
-
06 PAPER / AIM-DOC-1966-001
ELIZA—a computer program for the study of natural language communication between man and machine
The paper behind a conversation that felt richer than its mechanism.
Why it is here
Weizenbaum described ELIZA as a program for studying natural-language communication. Its operation depended on decomposition rules, keyword ranking, and reassembly—not on a general understanding of the conversation.
What to notice
Compare the modest mechanism with the human tendency to supply meaning. The gap between a system’s operation and a visitor’s interpretation remains a live museum question.
- Creator
- Joseph Weizenbaum
- Date
- 1966-01
- Publisher / holder
- Communications of the ACM
- Display note
- Metadata and commentary only
-
07 PHYSICAL INSTRUMENT / AIM-ROB-1969-001
Shakey
A wheeled robot linked cameras and touch sensors to a large remote computer for bounded room tasks.
The surviving Shakey robot photographed on display at the Computer History Museum on 10 January 2007.
“Shakey.png” — Marshall Astor; Sanchom (source crop) Image rights: CC-BY-SA-2.0 · Creative Commons Attribution-ShareAlike 2.0 Generic ↗Why it is here
The Computer History Museum dates Shakey to 1969, records an overall size of 69 × 47 × 47 inches, and names SRI as manufacturer. The object makes a distributed early robot system visible: cameras and touch sensors on a mobile body, computation elsewhere, and instructions translated into planned actions.
What to notice
Do not read this as general or fully onboard autonomy. The catalogue explicitly places control in a large remote computer, while the surviving body carried sensors and motion hardware. Its current display status was checked on 11 September 2026, not promised permanently. The copyrighted catalogue image is not reproduced; the separate 2007 visitor photograph shown here carries its own CC BY-SA 2.0 record.
- Creator
- Stanford Research Institute (SRI)
- Date
- 1969
- Publisher / holder
- Computer History Museum
- Collection
- Computer History Museum · X279.83
- Physical size
- Overall: 69 in × 47 in × 47 in
- Display note
- Licensed photograph · CC-BY-SA-2.0
-
08 PHYSICAL INSTRUMENT / AIM-HUM-1973-001
WABOT-1 humanoid robot
A 1973 Waseda research machine brought bipedal walking, hands, and simple Japanese communication into one full-body humanoid.
Why it is here
Waseda University's history describes WABOT-1 as widely considered the first full-scale humanoid robot. Its record joins mechanisms and information systems in one inspectable body: the machine walked on two legs, grasped objects with its hands, and communicated in simple Japanese.
What to notice
This institutional retrospective does not settle an uncontested priority claim, demonstrate general autonomy, or show that the machine understood language. The Building 63 display status was checked from Waseda's 2026 record and is not permanent. Dimensions are not recorded here, and the credited page image is not reproduced because no open licence is stated.
- Creator
- Waseda University WABOT Project · Ichiro Kato
- Date
- 1973
- Publisher / holder
- Waseda University
- Collection
- Waseda University · WABOT-1
- Physical size
- Not recorded by source
- Display note
- Metadata and commentary only
-
09 RESEARCH BOOK / AIM-BOOK-1984-001
Rule-Based Expert Systems: The MYCIN Experiments
A 754-page critical account of a rule-based expert-system research programme.
Why it is here
Buchanan and Shortliffe edited a retrospective analysis of nearly a decade of MYCIN-related experiments. Its contents expose the work around the rule engine: acquiring knowledge, handling uncertainty, explaining conclusions, and evaluating a system.
What to notice
This is a research record, not medical guidance or a deployment certificate. Read how rules, uncertainty, explanation, evaluation, and human use were separated into design problems.
- Creator
- Bruce G. Buchanan · Edward H. Shortliffe
- Date
- 1984
- Publisher / holder
- Addison-Wesley edition; Columbia University author archive
- Display note
- Metadata and commentary only
-
10 PAPER / AIM-DOC-1986-001
Learning representations by back-propagating errors
A concise account of adjusting network weights by propagating output error backward.
Why it is here
Rumelhart, Hinton, and Williams described repeatedly changing connection weights to reduce the difference between actual and desired outputs, allowing hidden units to acquire useful task features.
What to notice
The object documents one influential 1986 formulation. It does not establish that every later learning system uses the same procedure, or that the authors invented every ingredient.
- Creator
- David E. Rumelhart · Geoffrey E. Hinton · Ronald J. Williams
- Date
- 1986-10-09
- Publisher / holder
- Nature
- Display note
- Metadata and commentary only
-
11 PAPER / AIM-DOC-1989-001
Backpropagation Applied to Handwritten Zip Code Recognition
A Bell Labs paper records one trainable network reading normalized images of handwritten postal digits.
Why it is here
LeCun and colleagues describe putting task constraints into the network architecture and training a single network for the whole recognition operation, from a normalized character image to its final classification.
What to notice
The museum checked the abstract, not the full paper. This record supports the stated postal-digit task and image-to-classification path; it does not support calling the system the first convolutional network or a general-purpose vision system.
- Creator
- Y. LeCun · B. Boser · J. S. Denker · D. Henderson · R. E. Howard · W. Hubbard · L. D. Jackel
- Date
- 1989-12-01
- Publisher / holder
- Neural Computation, MIT Press
- Display note
- Metadata and commentary only
-
12 PHYSICAL INSTRUMENT / AIM-CHP-1997-001
Deep Blue custom chess chip, version 2
A surviving version 2 processor made Deep Blue’s bounded chess search inspectable as dedicated hardware.
Why it is here
The Computer History Museum dates this custom chip to 1997, credits its design to Feng-Hsiung Hsu, and records 1.5 million transistors running at 24 MHz. IBM Research’s publication record adds system context: the 1997 Deep Blue used 480 custom chess chips, which supplied most of its computational power.
What to notice
This is one chip, not the complete Deep Blue system, and specialised chess search is not general or human-like intelligence. The collection page records neither dimensions nor current display status, so both remain explicitly unrecorded here. Its image is copyrighted and is not reproduced.
- Creator
- Feng-Hsiung Hsu · IBM Research
- Date
- 1997
- Publisher / holder
- Computer History Museum
- Collection
- Computer History Museum · 102645415
- Physical size
- Not recorded by source
- Display note
- Metadata and commentary only
-
13 PAPER / AIM-DOC-1997-001
Long Short-Term Memory
A journal paper proposes a gated recurrent method intended to keep error signals available across long time intervals.
Why it is here
Hochreiter and Schmidhuber frame decaying error backflow as an obstacle to learning across long delays, then describe constant error flow through special units controlled by multiplicative gates.
What to notice
The museum checked the official bibliographic record and abstract, not the full paper. This object records the authors’ stated mechanism and artificial long-time-lag experiments; it does not establish a universal solution to memory or describe every later LSTM implementation.
- Creator
- Sepp Hochreiter · Jürgen Schmidhuber
- Date
- 1997-11-15
- Publisher / holder
- Neural Computation, MIT Press
- Display note
- Metadata and commentary only
-
14 PAPER / AIM-DOC-2009-001
ImageNet: A Large-Scale Hierarchical Image Database
A conference paper records a shared visual database built with a WordNet hierarchy, web collection, and human verification.
Why it is here
Deng and colleagues report 3.2 million images organized across 5,247 WordNet synsets and describe the collection and cleaning workflow, including internet search and human verification. The object makes data organization and label review visible as part of the AI system.
What to notice
The counts are a 2009 snapshot, not the current size of ImageNet. The paper does not show that the database alone caused the 2012 result, and this museum record does not reproduce its images.
- Creator
- Jia Deng · Wei Dong · Richard Socher · Li-Jia Li · Kai Li · Li Fei-Fei
- Date
- 2009
- Publisher / holder
- IEEE Computer Society / ImageNet
- Display note
- Metadata and commentary only
-
15 PAPER / AIM-DOC-2012-001
ImageNet Classification with Deep Convolutional Neural Networks
A conference paper joining a deep convolutional network, a large image benchmark, and GPU implementation.
Why it is here
Krizhevsky, Sutskever, and Hinton described training a large deep convolutional network on ImageNet, including architectural choices, a GPU implementation, and measures intended to reduce overfitting.
What to notice
Keep the claim inside the experiment: one architecture, one competition dataset, and reported test errors. “Deep learning won” is a later story, not a result printed by this object.
- Creator
- Alex Krizhevsky · Ilya Sutskever · Geoffrey E. Hinton
- Date
- 2012
- Publisher / holder
- Advances in Neural Information Processing Systems 25
- Display note
- Metadata and commentary only
-
16 CASE STUDY / AIM-CAS-2014-001
SAFFiR aboard ex-USS Shadwell
A bipedal firefighting research robot photographed during a bounded shipboard test.
SAFFiR during testing aboard the Naval Research Laboratory's ex-USS Shadwell in Mobile, Alabama, on 6 November 2014.
“Making Sailors 'SAFFiR' - Navy unveils Firefighting Robot prototype at Naval Tech EXPO [Image 7 of 8]” — John F. Williams; Wikimedia Commons (archival mirror) Image rights: DVIDS-PD-US · Public domain in the United States ↗Why it is here
SAFFiR makes a 2014 human–robot control boundary visible in one machine: a humanoid body, sensors intended for smoke-filled spaces, hose manipulation, and an operator who could intervene. The official record lets visitors distinguish a demonstrated research loop from a finished autonomous service.
What to notice
This is a research-test record, not evidence of autonomous deployment. ONR says SAFFiR still took instructions from researchers at a console and retained human intervention. The record does not establish where the prototype is preserved today; the image's public-domain statement is limited to the United States and carries separate non-endorsement guidance.
- Creator
- Virginia Tech researchers
- Date
- 2014-11-06
- Publisher / holder
- Office of Naval Research
- Display note
- Public-domain photograph · DVIDS-PD-US · US
-
17 PAPER / AIM-DOC-2016-001
Mastering the game of Go with deep neural networks and tree search
A Nature paper records a Go system that combines learned policy and value networks with tree search.
Why it is here
Silver and colleagues describe training policy networks first from expert games and then through self-play, learning a value network from self-play positions, and using both to guide Monte Carlo tree search.
What to notice
Keep the claim inside this paper and the game of Go. Its reported five-game result against Fan Hui does not establish general intelligence, and this record does not cover the later Lee Sedol match or reproduce the paper’s figures.
- Creator
- David Silver · Aja Huang · Chris J. Maddison · Arthur Guez · Laurent Sifre · George van den Driessche · Julian Schrittwieser · Ioannis Antonoglou · Veda Panneershelvam · Marc Lanctot · Sander Dieleman · Dominik Grewe · John Nham · Nal Kalchbrenner · Ilya Sutskever · Timothy Lillicrap · Madeleine Leach · Koray Kavukcuoglu · Thore Graepel · Demis Hassabis
- Date
- 2016-01-28
- Publisher / holder
- Nature / Google DeepMind
- Display note
- Metadata and commentary only
-
18 PAPER / AIM-DOC-2017-001
Attention Is All You Need
The paper that introduced the Transformer architecture.
Why it is here
The authors proposed an encoder–decoder architecture built around attention, without recurrence or convolution. The object matters because it gives a compact technical account of the architecture rather than a retrospective origin story.
What to notice
Start with Figure 1 and the model architecture section. Separate what the paper actually proposes from everything later products built on top of it.
- Creator
- Ashish Vaswani · Noam Shazeer · Niki Parmar · Jakob Uszkoreit · Llion Jones · Aidan N. Gomez · Lukasz Kaiser · Illia Polosukhin
- Date
- 2017-06-12
- Publisher / holder
- arXiv
- Display note
- Metadata and commentary only
-
19 RELEASE RECORD / AIM-WEB-2022-001
Introducing ChatGPT
The launch record for a conversational research preview.
Why it is here
OpenAI presented ChatGPT as a research preview and described a dialogue format that could answer follow-up questions, acknowledge mistakes, challenge incorrect premises, and reject some inappropriate requests.
What to notice
Read the page as a launch document. Its examples, limitations, and deployment language show how the product was framed on release day—not how later versions behaved.
- Creator
- OpenAI
- Date
- 2022-11-30
- Publisher / holder
- OpenAI
- Display note
- Metadata and commentary only