Like all forms of AI, language models are programs trained on large sets of data. The term "big data" is commonly used in English. Language models are trained on large amounts of text, called corpora. The language models are created by the AI system predicting the next word in the input and then trying to guess it. In this way, the system trains itself to produce better output by finding patterns and connections between words in the corpus.

At first, the model will "guess" completely at random, but as it receives positive and negative feedback, it gradually builds a statistical picture of relationships. With this statistical foundation, the model can calculate connections with much higher accuracy. The larger the dataset, the more accurate it will be, even on data the model has not been trained on.

Illustrasjon: Sammenhengen mellom et KI-system, operatør og omgivelser. Fra Nasjonal strategi for kunstig intelligens. (s.12), av Kommunal- og moderninseringsdepartementet, 2020

Illustration: Kommunal-og moderniseringsdepartementet. (2020). Nasjonal strategi for kunstig intelligens. s. 12. regjeringen.no

As the illustration shows, language models are not only trained using unsupervised learning and deep learning. Supervised learning and reinforcement learning are often used as a next step to fine-tune the models. In reinforcement learning, an operator provides feedback to improve a model by rewarding (reinforcing) or punishing the answers it gives. Thus, responses from the language model will be influenced by this adjustment as well as the original dataset.

Researcher and author Inga Strümpke offers a simple explanation to how the large language models behind advanced chatbots work, using MyAI integrated into SnapChat as an example. Regrettably, this video is in Norwegian only and lack subtitles.

Video: Tenk. (2023). MyAI og språkmodeller. youtube.com

Social media News media Academic literature Artificial intelligence