1.1. What is a text-based AI model?

Text-based AI models (chatbots) are AI-powered programs capable of generating text on any topic.

Text-based AI models (chatbots) can solve school assignments, take exams, write essays, and generate computer programs. Popular text-based AI models such as ChatGPT, Gemini, and Claude are based on large language models (LLMs) and fall under the field of natural language processing (NLP), which focuses on the interaction between computers and human language.

These models are trained to simulate conversations, answer questions, generate content, and perform other language-based tasks based on the input (a prompt) they receive.

How do text-based AI models differ from other AI systems?

The distinctive feature of text-based AI models compared to other AI systems is evident from their name – text-based AI models create and process text. While text-based AI models are designed using natural language processing (NLP), AI models for image generation, for example, rely on entirely different foundational models, specifically machine vision models. Computer vision systems, for instance, are built to interpret and process visual data, enabling machines to identify objects, analyse images, and make decisions based on visual input.

The primary function of text-based AI models is natural language processing—they generate and refine text, unlike other AI systems that focus on processing visual or spatial data.

In the field of robotics, AI is used for physical interaction with the environment, enabling machines to control movement, respond to sensory feedback, or solve spatial tasks. These systems make real-time decisions based on data from sensors, cameras, or other input streams.

Text-based AI models rely almost exclusively on textual data, both during training and operation. They are trained on massive amounts of written content, allowing them to learn grammar, syntax, and word meanings in different contexts. This text-based training makes them highly effective for language-related tasks but less suited for processing other data types, such as audio or visual inputs. Despite their focus on text, many widely used text-based AI models (e.g., ChatGPT) are integrated with other foundational models capable of recognizing text in images and audio. This expands their functionality beyond language tasks—for example, they can solve a math problem based on a photo taken with a smartphone.

The learning mechanisms of text-based AI models also differ from those used in other AI fields. These models are typically trained using unsupervised or semi-supervised learning, leveraging large text-based datasets. Their goal is to identify linguistic patterns – how words and sentences are structured – so they can predict the next word in a sequence. This predictive capability enables them to generate coherent responses to user input.

Text-based AI models are trained on extensive textual datasets, which helps them develop an understanding of grammar, syntax, and language patterns, making them highly effective for text-based tasks.

Other AI systems, such as those utilising reinforcement learning, operate under a different paradigm. Instead of training on static datasets, these models learn by interacting with their environment and receiving feedback in the form of rewards or penalties. This approach is commonly used in robotics and game AI, where the model refines its strategies or behaviours over time to maximise performance or rewards. For example, AI models trained via reinforcement learning can learn to play chess through trial and error, improving their decision-making process with each game.

Although text-based AI models fall under the category of narrow AI, they are often perceived as broader in scope than other generative AI models, such as those designed for image creation. This is because text-based AI can handle an exceptionally wide range of tasks, including writing research papers, finding sources, language learning, translation, and providing expertise across various academic fields. Even though models like ChatGPT can recognize text in images and respond through spoken output, they still remain within the domain of narrow AI.

How does a text-based AI model use data to improve its models?

To train an AI model, it is necessary to have as much high-quality data as possible. The GPT-4o model has 100 trillion parameters. The AI uses data for learning by identifying patterns and relationships within it, optimising parameters, and adjusting them so that the model can make the most accurate predictions possible.

Self-assessment