The collection
Every exhibit has a permanent page, a short interpretation and documentation. Search by person, date or concept, then open the exhibit record to inspect the sources behind the story.
In the Iliad, the craftsman god has made golden assistants who resemble living young women. Homer says they have intelligence, voice and strength. It may be one of the earliest literary descriptions of “intelligent machines”.
Greek precursors
Aristotle systematises reasoning: rules that lead from valid premises to valid conclusions. Many centuries later, rule-based AI would build on this idea. In the Politics he also imagines tools that could perform their work by themselves, without human hands.
Greek precursors
In Alexandria, Ctesibius improves the water clock with a float that regulates the flow of water automatically. It is among the earliest known examples of automatic control through feedback: the mechanism corrects its own operation. Centuries later, feedback would become central to cybernetics and AI. He also developed pumps and the water organ.
Greek precursors
In Apollonius of Rhodes’ Argonautica, a bronze giant made by Hephaestus guards Crete by circling the island. A mythical robot-like guardian, with a single vulnerable point.
Greek precursors
Recovered from a shipwreck in 1901 and dated to roughly the second to first century BCE, the mechanism used dozens of gears to model the positions of the Sun and Moon and predict eclipses. It is often described as the first analogue computer. The original is held by the National Archaeological Museum in Athens.
Greek precursors
Hero describes automatic spectacles, including birds made to sing with air and water, and a cart that moves by itself. A rope wound around its axle, with pegs placed at selected points, determines when it travels straight and when it turns. In modern terms, it resembles a program stored in rope.
Greek precursors
McCulloch and Pitts describe a brain neuron as a simple mathematical switch: it either fires or it does not. Their abstraction is an ancestor of today’s artificial neural networks.
The beginnings
Alan Turing proposes a test: if, during a written conversation, you cannot tell whether you are speaking to a person or a machine, should the machine count as intelligent?
The beginnings
A group of researchers meets for a summer in the United States. The phrase “artificial intelligence”, proposed the previous year, becomes the name of a new scientific field.
The beginnings
Frank Rosenblatt builds a machine that learns to recognise simple patterns from examples instead of having every rule written for it.
The beginnings
Joseph Weizenbaum at MIT builds a program that imitates a psychotherapist. It understands nothing, yet many users feel that it understands them.
The beginnings
Minsky and Papert show mathematically that a simple perceptron cannot solve some basic classes of problems. Research on neural networks then loses momentum for years.
The AI winters
A report commissioned by the British government argues that AI research has failed to meet many of its goals. Funding is cut and the first AI winter begins in Britain.
The AI winters
Programs containing hundreds or thousands of “if… then…” rules enter business use. Digital Equipment’s XCON configures computer orders and is credited with major savings.
The AI winters
Expert systems prove expensive to maintain and brittle when faced with exceptions. At the same time, the market for specialised AI computers collapses.
The AI winters
IBM’s computer defeats the reigning world chess champion. It does not “think” like a human player: it searches enormous numbers of positions at high speed. AI returns to front-page news.
The AI winters
IBM researcher Arthur Samuel develops a checkers program that improves by playing repeatedly and uses the term “machine learning”.
How a machine learns
Fei-Fei Li’s team creates a huge database of millions of labelled images. It becomes clear that the quantity and quality of data can matter as much as the algorithm.
How a machine learns
Joy Buolamwini and Timnit Gebru show that commercial face-analysis systems make far more errors for darker-skinned women than for lighter-skinned men. The result exposes the consequences of unrepresentative data and uneven system performance.
How a machine learns
Reports reveal that Amazon dropped an experimental résumé-ranking system after it had learned from historical data to disadvantage women applicants.
How a machine learns
Rumelhart, Hinton and Williams show how a multi-layer network can correct its errors by propagating the error signal backwards. This helps overcome the limitation highlighted by Minsky and Papert.
Neural networks
Yann LeCun trains a network to recognise handwritten digits in postal ZIP codes. Related systems were later used for tasks such as reading bank cheques.
Neural networks
A deep neural network trained on graphics processors wins the ImageNet image-recognition competition by a large margin. The modern deep-learning era accelerates.
Neural networks
DeepMind’s system defeats Lee Sedol, one of the world’s leading Go players, in a game long considered exceptionally difficult for machines.
Neural networks
John Hopfield and Geoffrey Hinton are awarded the 2024 Nobel Prize in Physics for foundational discoveries and inventions that enabled machine learning with artificial neural networks.
Neural networks
Claude Shannon shows that text that resembles English can be generated by measuring which symbols or words tend to follow others.
Language and the first LLMs
With word2vec, each word can be represented as a point in a numerical space. Words used in similar contexts tend to end up close together.
Language and the first LLMs
Google researchers publish “Attention Is All You Need”. The attention mechanism allows the model to relate each token to other tokens in the sequence.
Language and the first LLMs
OpenAI and Google show that a model first pretrained on a very large body of text can later be adapted to many different tasks.
Language and the first LLMs
With 175 billion parameters, the model can write text, translate and answer questions, often from only a few examples in the prompt.
Language and the first LLMs
On 30 November 2022, a conversational language-model interface is released publicly as a research preview. Generative AI begins to enter everyday life at unprecedented scale.
Language and the first LLMs
Huge quantities of text are gathered from sources such as websites, books and code. Selection and cleaning influence what a model can learn, how well it performs and which biases it may reproduce.
How an LLM is made today
Text is split into smaller units called tokens, which may be whole words or pieces of words. The model processes numerical identifiers rather than written characters directly.
How an LLM is made today
The model repeatedly learns to predict the next token across enormous text corpora and large amounts of computing. This stage gives it broad language ability and general patterns, but not necessarily assistant-like behaviour.
How an LLM is made today
With examples of instructions and high-quality answers, the model is trained to follow requests and behave more like an assistant.
How an LLM is made today
People, and sometimes other models, evaluate answers. Training then pushes the model toward responses judged more helpful, honest or safe. A well-known approach is reinforcement learning from human feedback (RLHF).
How an LLM is made today
Before release, models are evaluated for errors, harmful behaviour and misuse. Testing and monitoring can continue after deployment.
How an LLM is made today
When you type a prompt, a basic language model does not retrieve a ready-made answer from a database. It generates a response token by token, although some modern systems can also call search engines, tools or external knowledge sources.
How an LLM is made today
The term becomes associated with synthetic videos in which one person’s face is replaced or generated using AI. Today voice, image and video can all be fabricated convincingly, with consequences for fraud, elections and public information.
Society and ethics
The 2023 Hollywood writers’ strike includes negotiated rules on the use of AI. The broader question remains: which jobs change, which disappear, and which new roles emerge?
Society and ethics
Authors, artists and news organisations, including The New York Times, take disputes over the use of copyrighted works in model training to court. Many legal questions remain unsettled.
Society and ethics
The European Union adopts the first comprehensive horizontal law on artificial intelligence, introduced in stages. It bans certain practices and imposes stricter duties on high-risk uses such as some employment and education systems.
Society and ethics
A small Greek team of linguists and computer scientists joins the European EUROTRA project to build machine translation between Community languages. This work becomes a starting point for organised Greek language technology.
AI in Greece
SKEL is founded at NCSR Demokritos and grows into one of Greece’s largest AI research labs, working on intelligent information systems, data analysis and human-computer interaction.
AI in Greece
The EUROTRA team gains a permanent institutional home with the creation of ILSP. Its work expands from translation to speech, language resources, culture and human-machine interaction.
AI in Greece
Researchers release a BERT model trained specifically on Greek text, showing why a language with far fewer digital resources than English benefits from dedicated models and evaluation.
AI in Greece
Athena Research Center establishes a unit for basic and applied research in AI, data science and algorithms, linking researchers in Greece and abroad.
AI in Greece
The gov.gr digital assistant brings generative AI into a public service used by citizens at national scale.
AI in Greece
ILSP introduces Meltemi, an open large language model designed for Greek, highlighting the challenge of languages with fewer digital resources.
AI in Greece
Llama-Krikri extends the effort with more Greek training data, a larger context window and broader capabilities, while remaining openly available.
AI in Greece
Greece joins the European AI Factories network with PHAROS, focusing among other areas on Greek language and culture, health and sustainability.
AI in Greece
The DAEDALUS supercomputer brings new national computing capacity for science and AI. In the June 2026 TOP500 it ranks 31st with measured performance of 85.69 petaflops.
AI in Greece
AI systems are beginning to use computers and tools to complete sequences of actions. On OSWorld, agent performance rose from roughly 12% to 66.3% in a year, yet agents still fail about one in three structured attempts.
The future of AI
Robotics is increasingly connected with vision-language models. Yet the gap between a controlled lab and an unpredictable home remains large: the 2026 AI Index reports only 12% success on real household tasks.
The future of AI
Self-driving laboratories combine AI, robotics and automation: a system proposes an experiment, runs it and uses the result to choose the next step. Today they mostly operate in narrow, well-defined domains.
The future of AI
AI is already used in diagnosis, drug discovery and research on medical digital twins. Early results are promising, but clinical reliability and generalisation require rigorous evaluation.
The future of AI
Generative AI is already widespread in study and education. The question is shifting from whether students will use it to what humans should learn when machines can write, solve and explain.
The future of AI
The ILO estimates that one in four workers is in an occupation with some GenAI exposure. In the near term, task transformation is more likely than the automatic disappearance of whole occupations.
The future of AI
The future is not only about ever larger models. Smaller systems can run locally, specialise in specific domains and offer lower cost or better privacy.
The future of AI
AI depends on physical infrastructure: chips, data centres, cooling and electricity. In the IEA base case, global data-centre electricity use roughly doubles to around 945 TWh by 2030.
The future of AI
As AI enters higher-stakes decisions, requirements for transparency, evaluation, provenance and accountability grow. In Europe the AI Act is being phased in over several years.
The future of AI
There is no scientific consensus on whether or when artificial general intelligence will exist. Current systems show increasingly autonomous capabilities, but that does not establish an inevitable path to AGI or loss of control.
The future of AI
No exhibits match those criteria.