How a language model works
The model is trained on an enormous amount of text and learns to predict which word (more precisely, token) comes next. From this simple goal, at a huge scale, abilities emerge such as translation, summarising, programming and answering questions. The model is then fine-tuned to be helpful and safe.
What it can and cannot do
- It writes, translates, summarises and helps with code very well.
- It can make up facts that sound convincing. These are called hallucinations.
- It may not know about the latest events if it has no access to search.
- Important information should therefore be checked.
The race
- 2022ChatGPT
- 2023GPT-4, Claude, Llama, Gemini
- 2024Models understand images and voice in real time
- 2025Models that “think” in steps, and AI agents
- 2026New versions arrive within weeks: at the end of September GPT-6.1 Sol, Claude Sonnet 5.5 and Gemini 4 Argon
Regulation
In 2024 the European Union adopted the Artificial Intelligence Act (AI Act), the world's first comprehensive law on AI. Its obligations are being phased in gradually and cover, for example, labelling content created by artificial intelligence.