Luca Dan Șerbănați at the baccalaureate, 1961
Luca Dan Șerbănați in the early 1970s
Luca Dan Șerbănați giving his last fifth-year lecture, April 1989

Luca Dan Șerbănați

Emeritus Professor at Politehnica University of Bucharest

Research, teaching, industry and memoirs

RO | EN | IT
Luca Dan Șerbănați in Venice, 1990
Luca Dan Șerbănați in New York, 2005
Luca Dan Șerbănați

A timeline of artificial intelligence, 1943–2025

From the artificial neuron model and symbolic AI to deep learning, foundation models, generative AI and the first global regulations.

This page is a concise timeline, not an exhaustive history of artificial intelligence. Its purpose is to provide an orientation framework for my texts on AI: the 1985 book, later research on ontologies, agents, digital ecosystems and digital twins, and recent reflections on conversational models.

I chose 1943 as the starting point because the formal model of the artificial neuron proposed by McCulloch and Pitts anticipates one of the deep lines of AI: the attempt to connect thought processes with logical or computational structures. 1956 remains, however, the institutional founding moment, through the Dartmouth Conference, where the term “artificial intelligence” entered scientific circulation.

A personal thread: R2V2R, avatars and digital twins

Before the chronology itself, it is useful to recover a conceptual thread specific to my own research: the transition from R2V2R – Real to Virtual to Real – to avatars and digital twins in healthcare. This direction does not belong to the general chronology of AI, but it shows how some ideas from artificial intelligence, multi-agent systems and digital modeling entered my own research path.

In its simplest form, R2V2R describes a circuit: medical reality is captured and projected into a virtual space, and that virtual space, once structured, interpreted and activated through intelligent services, returns to reality by supporting decision and action. This thread includes the concepts of Virtual Healthcare Record, Digital Healthcare Ecosystem, LuMiR, the agentification of EHR systems and, finally, Health Digital State and Smart EHR.

Retrospectively, the health avatar proposed through Smart EHR can be read as an early form of digital twin: not a passive copy of the patient, but an active digital representation, fed by clinical data, able to interact with the physician, organize information and participate in diagnostic and therapeutic processes.

Articles connected with this thread

1943–1956 – The prehistory of artificial intelligence

Before the official name

Before the field received the name “artificial intelligence”, several converging lines prepared the ground: mathematical logic, cybernetics, information theory, programmable electronic computers and the first formal models of the neuron.

  • 1943 – Warren McCulloch and Walter Pitts propose a formal model of the artificial neuron, based on Boolean logic.
  • 1946–1949 – The first programmable electronic computers and the stored-program architecture appear.
  • 1948 – Norbert Wiener publishes Cybernetics, a work that influences thinking about feedback, control and self-regulation.
  • 1950 – Alan Turing publishes Computing Machinery and Intelligence and proposes what will become the Turing Test.
  • 1951 – Dietrich Prinz implements a simplified chess program on the Ferranti Mark I.
  • 1955–1956 – Allen Newell and Herbert Simon develop Logic Theorist, capable of proving logical theorems.

1956–1966 – The founding enthusiasm

The birth of symbolic AI

The Dartmouth Conference of 1956 marks the establishment of AI as a research field. The first decade is dominated by optimism: intelligence appears modelable through symbols, rules, search in state spaces and programs capable of solving problems.

  • 1956 – The Dartmouth Conference establishes the term “Artificial Intelligence”.
  • 1958 – John McCarthy develops Lisp, which becomes an essential language for AI research.
  • 1957–1958 – Frank Rosenblatt proposes the perceptron, a simple early form of neural network.
  • 1959 – Newell and Simon develop General Problem Solver, a symbol of the general problem-solving approach.
  • 1961 – Unimate becomes the first industrial robot installed in a factory.
  • 1965–1966 – ELIZA, Joseph Weizenbaum’s program, reveals the psychological force of text-based human–machine interaction.

1966–1976 – First limits and the first winter

From optimism to caution

After the founding enthusiasm, AI research encounters theoretical and practical difficulties: computers are still limited, the real world proves difficult to formalize, and the initial promises do not materialize at the expected pace.

  • 1968–1970 – SHRDLU, Terry Winograd’s system, combines natural language, reasoning and a symbolic blocks world.
  • 1969 – Minsky and Papert publish Perceptrons, contributing to the decline of interest in simple neural networks.
  • 1971 – STRIPS introduces an influential formalism for automated planning.
  • 1972 – Prolog appears in France and becomes one of the key languages of logic programming.
  • the 1970s – DENDRAL and MYCIN demonstrate the potential of expert systems in specialized domains.
  • 1973 – The Lighthill Report criticizes AI results and contributes to funding cuts in the United Kingdom.
  • 1976 – Joseph Weizenbaum publishes Computer Power and Human Reason, bringing the ethical and philosophical limits of simulated intelligence into the debate.

1977–1987 – Expert systems and the revival of symbolic AI

The age of explicit knowledge representation

AI regains visibility through expert systems, symbolic languages and ambitious projects. Prolog, Lisp, production rules, inference engines and knowledge engineering become central landmarks.

  • late 1970s – MYCIN becomes the classic example of a medical expert system.
  • 1978–1980 – XCON/R1 is implemented at Digital Equipment Corporation for computer system configuration.
  • 1979 – OPS5 becomes an important tool for rule-based systems.
  • 1980–1981 – AI-specialized companies such as IntelliCorp and Teknowledge appear.
  • 1982 – Japan launches the Fifth Generation Computer Systems project, based on logic programming and parallel architectures.
  • 1986 – Backpropagation is popularized by Rumelhart, Hinton and Williams, renewing interest in multilayer neural networks.
  • the 1980s – Mature directions in distributed artificial intelligence and multi-agent systems begin to emerge.

This period is also important for my personal trajectory: in 1985 I published, together with Cristian Giumale, the book Inteligența artificială, written within the symbolic paradigm of the time.

1987–1997 – The second winter and probabilistic methods

The decline of expert systems and a shift in direction

High costs, the difficulty of maintaining knowledge bases and the collapse of the Lisp machine market lead to a new period of skepticism. At the same time, probabilistic and statistical methods consolidate, preparing the following changes.

  • late 1980s – The Lisp machine market collapses and expert systems lose part of their commercial credibility.
  • 1988 – Judea Pearl publishes fundamental work on Bayesian networks and probabilistic reasoning.
  • 1989 – ALVINN, a Carnegie Mellon project, demonstrates autonomous driving assisted by neural networks.
  • the 1990s – Hidden Markov Models become dominant in speech recognition.
  • 1993 – Quinlan’s C4.5 algorithm consolidates decision trees as a machine-learning tool.
  • 1995 – ALICE anticipates one line of chatbot development.
  • 1997 – Deep Blue defeats Garry Kasparov, a symbolic moment in the relationship between massive computation and human intelligence.

1997–2007 – Machine learning, the internet and data

AI enters the web infrastructure

The growth of the internet changes the scale of the problems. Search engines, e-commerce, recommendation systems, machine translation and autonomous vehicles increasingly depend on statistical and machine-learning methods.

  • 1997 – The Deep Blue victory has major public impact and brings AI back into the media spotlight.
  • 1998 – Google Search uses PageRank, an algorithm that reorganizes access to web information.
  • the 2000s – SVMs, Random Forests, Boosting and other statistical methods become standard tools in classification and prediction.
  • 2004–2005 – The DARPA Grand Challenge gives a decisive impulse to autonomous vehicle research.
  • 2006 – Amazon Web Services facilitates access to scalable storage and computing infrastructure.
  • 2006–2007 – Research by Hinton, Bengio, LeCun and others prepares the return of deep neural networks.

2007–2016 – The deep learning revolution

Data, GPUs and deep networks

The availability of large volumes of data, GPU computing power and algorithmic advances turn deep neural networks into a dominant technology. AI enters commercial products, digital services, visual recognition, speech recognition and machine translation on a massive scale.

  • 2009 – ImageNet creates an essential benchmark for computer vision research.
  • 2012 – AlexNet wins the ImageNet competition and marks the visible beginning of the deep learning revolution.
  • 2014 – Ian Goodfellow proposes GANs, opening an important path for generative AI.
  • 2014–2016 – Assistants such as Siri, Alexa and Google Assistant become part of everyday AI use.
  • 2015–2016 – TensorFlow, Keras and PyTorch contribute to the democratization of AI model development.
  • 2016 – AlphaGo defeats champion Lee Sedol at Go, a moment comparable in symbolic impact to Deep Blue in 1997.
  • 2016 – Neural machine translation becomes a major industrial landmark.

2017–2025 – Generative AI and foundation models

From specialized models to cognitive infrastructures

The Transformer architecture decisively changes natural language processing. Large language models, multimodal models and generative AI move from specialized applications to systems able to write, summarize, program, generate images and enter into dialogue with the user.

  • 2017 – Attention Is All You Need introduces the Transformer, the architecture that will dominate large language models.
  • 2018 – BERT establishes bidirectional Transformer models for language understanding.
  • 2019–2020 – GPT-2 and GPT-3 amplify interest in large-scale generative language models.
  • 2020 – AlphaFold demonstrates the power of AI in protein-structure prediction.
  • 2021 – GitHub Copilot brings AI-assisted code generation into programming practice.
  • 2022 – DALL·E, Midjourney and Stable Diffusion popularize image generation; ChatGPT publicly transforms conversational interaction with AI.
  • 2023 – GPT-4, LLaMA, Claude and other models confirm the move toward a global competition of foundation models.
  • 2024–2025 – Multimodal AI, autonomous agents, enterprise applications and regulations, including the European Union AI Act, become central themes.

This is the stage in which AI ceases to be perceived only as background technology and becomes interlocutor, co-author, decision-support tool, knowledge infrastructure and a political, legal and moral problem.

Stage-by-stage synthesis

Period Characterization Landmarks
1943–1956 AI prehistory Artificial neuron, cybernetics, the Turing Test, programmable computers, Logic Theorist.
1956–1966 Founding enthusiasm Dartmouth, Lisp, perceptron, General Problem Solver, ELIZA.
1966–1976 First limits SHRDLU, Prolog, STRIPS, DENDRAL, MYCIN, the Lighthill Report.
1977–1987 Revival through expert systems XCON, OPS5, Prolog, Lisp machines, FGCS, backpropagation.
1987–1997 Second winter and probabilism Decline of expert systems, Bayesian networks, HMM, ALVINN, Deep Blue.
1997–2007 Machine learning and the internet PageRank, SVM, Random Forests, DARPA Grand Challenge, AWS, early large-scale web applications.
2007–2016 Deep learning ImageNet, AlexNet, GAN, TensorFlow, PyTorch, AlphaGo, neural translation.
2017–2025 Generative AI and foundation models Transformer, BERT, GPT, ChatGPT, DALL·E, AlphaFold, LLaMA, AI Act.

Internal links

Back to Artificial Intelligence • Back to Research