Global AI MapAI RESEARCH AND ROBOTICSBrowse AI history by date. Each section includes films, publications and a description of later developments.
McCulloch and Pitts connected idealized neuron behavior with formal logic.
This abstract model helped establish the idea that networks of simple units can implement computation. Modern neural networks use different training methods and richer units.
Wiener’s cybernetics and Shannon’s information theory supplied distinct mathematical languages for control and communication.
Feedback remains central to robotics; information theory underlies coding and communication. These are related foundations, not interchangeable theories.
Turing proposed a behavioral test and discussed learning machines.
The paper set a durable agenda for evaluating machine behavior. Modern benchmarks operationalize narrower capabilities and do not settle consciousness.
Whirlwind demonstrated interactive computing; early robots and Shannon’s maze machine made learning and feedback visible.
Real-time displays, control loops and networked command systems informed later interactive computing and robotics.
The Dartmouth proposal brought learning, language, abstraction and self-improvement into a common research program.
AI developed into multiple traditions: symbolic reasoning, statistical learning, neural networks and embodied systems.
Rosenblatt studied the perceptron while McCarthy developed Lisp for symbolic computation.
Learned numerical representations and symbolic programs became two influential approaches, later combined in many systems.
Systems engineers explored shared resources, parallel processing and common interfaces.
The lasting contribution is architectural: software and computation can be organized across interchangeable resources. Particular 1960 designs were not modern cloud systems.
Sketchpad and early vector displays supported direct manipulation of shapes and engineering models.
Interactive graphics influenced CAD, user interfaces and the visual tools later used to build and inspect AI systems.
The DENDRAL project used chemical knowledge and search to propose molecular structures from experimental data. It became an early example of a domain-specific expert system.
The work helped establish knowledge engineering: collecting expert knowledge and using it in a program. Later expert systems used this approach in medicine, configuration and other fields.
ELIZA explored conversational interaction; Shakey combined perception, planning and movement.
ELIZA exposed how readily people attribute understanding. Shakey’s planning work influenced later autonomous systems.
Interactive systems brought pointing, linked documents and collaborative editing into a shared workspace.
The legacy runs through graphical interfaces, networked collaboration and today’s agent workspaces.
Terry Winograd developed SHRDLU at MIT from 1968 to 1970. The program answered questions and followed instructions in a simulated world of blocks.
SHRDLU linked language, planning and a model of its environment. Its success depended on a small, defined world. Understanding unrestricted language remained a separate problem.
Stanford’s MYCIN project used rules to recommend treatments for bacterial infections. It also studied how a program could explain its recommendations and acquire new rules.
MYCIN was a research system, not a general-purpose doctor. Its rule-based methods and explanation tools influenced later expert-system software. Clinical reliability and deployment required more than a successful research demonstration.
MIT built computers designed to run Lisp efficiently. The CADR combined specialized hardware with an interactive software environment for developing and debugging programs.
The Lisp machine became a base for commercial workstations. Its integrated environment, memory management and support for symbolic programs remain useful subjects in programming-system design.
Symbolics was formed in 1980 to commercialize Lisp-machine technology. Its later 3600 systems combined Lisp-oriented hardware with a substantial software environment.
Lisp workstations supported AI development, engineering and graphics. General-purpose workstations later changed the economics of specialized machines. The demonstrations show the working environment, not only the hardware.
Expert-system projects applied rules and domain knowledge to practical tasks such as configuring computer systems. This work moved knowledge-based AI beyond laboratory demonstrations.
These systems could be useful within a defined domain. Maintaining the rules, handling exceptions and acquiring knowledge became major engineering costs. Later AI systems continued to face related problems of scope and maintenance.
Danny Hillis and colleagues founded Thinking Machines Corporation in 1983. The company developed Connection Machine computers that used large numbers of processors.
The machines supported scientific computing and research in parallel programming. Later CM-5 work used a different architecture from the early SIMD machines. Hillis’s recorded lecture explains the CM-5 design choices.
Doug Lenat began Cyc in 1984 to build a large base of common-sense knowledge and reasoning rules. The project addressed knowledge that people often leave unstated.
Cyc continued as a long-term knowledge-engineering project. Its approach differs from learning a language model from text. The 1992 documentary includes Lenat alongside Minsky, Feigenbaum and other AI researchers.
The 1984 Computer Chronicles episode presented AI research, expert systems and Lisp. Guests included John McCarthy, Edward Feigenbaum and researchers from SRI.
The programme records how researchers explained AI during the expert-system period. It is useful for comparing their demonstrations and expectations with later results. The 1985 episode provides a further view of the field.
Backpropagation made it practical to train useful internal representations in multilayer networks.
Combined with larger datasets, improved hardware and architectural advances, gradient-based learning became a foundation of modern deep learning.
Minsky’s Society of Mind proposed that complex thought can arise from the interaction of many specialized processes.
The book remains an influential conceptual account of cognition. Its internal agents are not identical to today’s tool-using software agents; the connection is a useful question rather than a claim of direct equivalence.
IBM recruited members of the Deep Thought chess team in 1989. Feng-Hsiung Hsu, Murray Campbell and colleagues worked on search, evaluation and specialized chess hardware.
Deep Blue won a game against Garry Kasparov in 1996, but lost that match. The team improved the system before the 1997 rematch. The two results should not be treated as the same event.
The Machine That Changed the World included a 1992 episode about AI. Participants included Doug Lenat, Marvin Minsky, Edward Feigenbaum and other researchers.
The film provides a record of symbolic AI, attempts to represent knowledge and debates about machine intelligence. The related papers and technical records let viewers examine the methods as well as the film’s account.
Deep Blue defeated the reigning world chess champion in a match.
The system demonstrated specialized search and evaluation at scale. Later game-playing systems increasingly learned representations and strategies.
IBM’s Watson combined language analysis, information retrieval, evidence scoring and other methods to answer Jeopardy! questions. It defeated Ken Jennings and Brad Rutter in the televised 2011 contest.
Watson demonstrated a question-answering system built for a specific task. Its result did not establish unrestricted language understanding. Ferrucci’s presentation and the DeepQA paper describe how the system worked.
AlexNet combined convolutional networks, GPU training and ImageNet-scale data.
Its results accelerated adoption of deep learning in computer vision, followed by wider use in speech, language and multimodal systems.
Attention-based sequence modeling removed recurrent structure from the core architecture.
Transformers became a common architecture for large language models and later appeared in vision, scientific modeling and robot policies.
AlphaFold demonstrated substantial advances in protein-structure prediction.
The work broadened access to structural hypotheses. Predicted structures remain starting points for experimental and biological interpretation.
Robot learning connected demonstrations, visual perception and learned control across a wider range of tasks.
Open robot datasets and reproducible platforms made it easier to compare methods. Generalization, safety and reliability remain active research problems.
Inference runtimes, tool-using harnesses and memory systems bring more AI work to smaller machines.
Results depend on model, quantization, context, hardware and task. The directory connects the models to the systems used to run them.
27m27s
20m16s
17m13s
26m28s
28m56s
9m43s
11m3s
21m24s
14m31s
28m48s
12m12s
5m28s
7m6s
24m24s
26m39s
9m32s
17m41s
46m57s
21m33s
21m29s
28m31s
14m49s
15m5s
24m46s
23m16s
20m56s
14m39s
17m19s
2m19s
8m36s
9m24s
10m45s
4m50s
6m42s
19m1s
1m56s
1h36m46s
1m9s
2m17s
13m30s
3m19s
3m1s
2m9s
10m43s
5m3s
5m43s
4m4s
44m54s
3m21s
2m27s
