Artificial intelligence,
without unnecessary jargon.
The concepts, systems, and technologies that are changing the way we work, create, and interact with machines.
Artificial intelligence
Systems that perform tasks associated with capabilities such as recognizing patterns, understanding language, making predictions, or generating content. There is no single “AI”: methods, data, capabilities, and limitations vary.
LLMs and language models
Large language models process and generate text by predicting sequences of words or tokens based on learned examples. They can summarize, translate, write code, and converse; they can also make mistakes, invent details, or appear confident without having verified anything.
Multimodal models
They integrate multiple types of information, such as text, images, audio and video. Multimodality enables richer interactions, but it does not guarantee that the system will always correctly interpret what it sees or hears.
AI agents
An agent receives an objective, plans steps and can use tools such as a browser, a repository or software. Greater autonomy also means a greater need for permissions, limits, action logs and the ability to stop it.
Multi-agent systems
Multiple agents can divide a task: one searches for information, another compares it, and another prepares a result. Collaboration can help with complex activities, but coordination, shared errors and costs grow with the number of components.
RAG and source retrieval
Retrieval-augmented generation combines a language model with search across documents or databases. The system can ground its response in specific materials and cite them, but quality depends on the selection, updating and correct interpretation of sources.
Robotics and physical AI
Robotics brings perception and decision-making into the material world: industrial robots, mobile systems and devices that interact with objects and people. Physical environments are less predictable than a chat: safety, sensors and motion control are central.
Generative AI
Generative models produce text, images, audio, video or code from instructions and learned data. They are tools for creating and transforming content; attribution, rights, provenance and verification remain essential issues.
Open models and open-weight models
Some models make weights or technical components available that can be run and adapted by third parties. “Open” can indicate different degrees of access: licenses, training data and freedom of use must be evaluated separately.
Evaluation and safety
Evaluations measure behavior on defined tasks and conditions. A good result on a benchmark does not on its own prove that a model is reliable in every environment: relevant tests, continuous checks and observation of real-world effects are needed.
Chips, cloud and infrastructure
AI requires computing, energy, networks and storage. Chips and data centers influence costs, availability and energy impact; infrastructure is part of the product and of competition among companies and countries.