Emergence describes properties of a system that arise from the interaction of many simple parts but cannot be seen in any single part. In AI research, abilities are called emergent when large language models above a certain size solve tasks such as multi-step arithmetic or translation markedly better than smaller models. Whether such jumps are genuine or partly depend on how they are measured is debated among researchers.
A large language model (LLM) is a neural network trained on very large amounts of text, enabling it to understand and generate language. It works by calculating the most likely next token, step by step. LLMs write, summarise, translate, answer questions and generate code; well-known examples include Claude, GPT, Gemini and Llama. Without access to current sources, they only know what was in their training data.
A foundation model is a large AI model pre-trained on very extensive, wide-ranging data that can then be used for many different tasks. The term was coined at Stanford University in 2021. Well-known examples include large language models such as Claude or GPT, as well as image models. Businesses usually build their applications on such models rather than training their own from scratch; the EU AI Act refers to general-purpose AI models.
Artificial general intelligence (AGI) refers to a hypothetical AI system that can handle a broad range of intellectual tasks at human level and flexibly transfer what it has learned to new areas. Today's systems perform well on many tasks but still show limitations, for instance in reliability. There is no generally accepted definition or way of measuring it, and whether and when AGI will be achieved is disputed among experts.
Multi-hop reasoning is an AI system’s ability to answer a question through several successive inferences, often combining information from different sources. Example: “Which department is headed by the person responsible for project X?” first requires identifying the responsible person and then their department. Such questions are demanding for search and RAG systems because every intermediate step must be found and linked correctly.
