General introduction

Ever since humanity began telling stories, it has dreamed of breathing thought into inert matter: statues that come to life, clay golems, automatons of copper. For a long time this dream belonged to myth. Then, in the middle of the twentieth century, it became a scientific discipline. And since the turn of the 2020s, it has become an industrial reality that is transforming, week after week, the way we work, heal, create and wage war.

Artificial intelligence refers to the set of techniques that allow machines to perform tasks which, if carried out by humans, would require "intelligence": understanding a text, recognizing a face, driving a car, proving a theorem, holding a conversation. It is a deliberately vague definition, and that vagueness is not a flaw but a feature of the subject. What we call "intelligent" keeps shifting: yesterday, playing chess at the highest level seemed the pinnacle of intelligence; today, it is a feature you download for free. Researchers call this the "AI effect": as soon as a machine can do something, we stop regarding that thing as intelligence.

This course is built around a historical turning point. For sixty years (1956-2017), AI advanced in fits and starts, alternating between heady promises and "winters" of disillusionment. Two great families of approaches succeeded and clashed with one another: symbolic AI, which sought to reproduce reasoning through explicit rules, and connectionist AI, which sought to imitate it by drawing inspiration from the brain, using neural networks fed on data.

In 2017, a scientific paper with a provocative title ("Attention Is All You Need") introduced a new architecture, the Transformer. No one at the time grasped its significance. Five years later, in November 2022, a product built on this architecture, ChatGPT, reached a hundred million users in two months, the fastest adoption in the history of any consumer technology. The world discovered large language models (LLM): systems capable of writing, coding, translating and reasoning at a level that commands respect and inspires unease in equal measure.

This entire course can be read as the story of that shift and its consequences:

How could a simple idea (predicting the next word) produce machines that seem to understand? And what does that change for science, work, war, the planet and, perhaps, the future of our species?

A single thread runs through the whole document. As these systems become more capable, the question of their control grows more acute. They are two sides of the same coin:

• On the bright side: proteins folded by computer opening the way to new medicines; earlier medical diagnoses; unprecedented mathematical discoveries; productivity multiplied tenfold.
• On the dark side: jobs under threat; a colossal energy bill; autonomous weapons; industrialized disinformation; and, on the horizon, the dizzying question of whether systems more intelligent than us will remain aligned with our intentions.

Neither the purely enthusiastic narrative ("AI will solve everything") nor the purely catastrophist one ("AI will destroy us") does justice to reality. This course strives to hold both ends at once.

To set the scene (each point is developed in detail in the relevant chapters):

• The state of the art is held by a handful of laboratories. On the American side, Anthropic (Claude models), OpenAI (GPT), Google DeepMind (Gemini) and xAI (Grok); on the Chinese side, DeepSeek, Qwen (Alibaba), Moonshot and others, often in open weights and at very low cost; on the European side, France's Mistral carries the banner of sovereignty. The ranking changes every month, and the lesson of 2026 is that there is no longer one best model, but a best model for each task.
• AI has become agentic. We no longer simply chat with a model: we entrust it with objectives, and it acts: it writes code, browses the web, manipulates software. This is the move from the "copilot" to the "digital worker".
• Robotics has left the laboratory. In 2026, humanoid robots are genuinely working in factories (Figure at BMW, Boston Dynamics' Atlas at Hyundai), while Chinese platforms such as Unitree sell them for 16,000 dollars. This is no longer a demonstration: it is an emerging market.
• Regulation is taking shape. The European Union's AI regulation (AI Act) reaches a decisive milestone on 2 August 2026, the date on which most of its obligations come into force.
• The debate over AGI has intensified. The leaders of the major laboratories publicly state that they are aiming for "artificial general intelligence" within the coming years. Detailed forward-looking scenarios, such as AI 2027, describe a possible dizzying acceleration that others consider overstated. We will return to this without complacency or doom-mongering.

We will start from the foundations (Part I): the history before the LLMs, then the mechanics of machine learning, up to the Transformer revolution. We will then enter the era of large models (Part II): LLMs, world models, agents, and the map of the players. We will descend into the engine room with hardware and infrastructure (Part III): chips, data centers, local AI, energy. We will explore the great convergences (Part IV): blockchain, quantum, robotics. We will see AI at work in the real world (Part V): health, work, law, defense. Finally, we will confront the existential stakes (Part VI): alignment, safety, governance and possible futures.

Enjoy the read.

---