Any form of machine intelligence is referred to as artificial intelligence (AI). In common parlance, the term artificial intelligence is used when a machine imitates “cognitive” performances such as learning and problem solving, which are otherwise associated with human mental performance. AI is an area of computer science research in which intelligent machines are developed and studied. An important area of modern research into artificial intelligence is machine learning: Here, the machine learns to classify images or predict time series with the help of huge amounts of data.

Generative Artificial Intelligence (GenAI) is an artificial intelligence that can generate media (e.g. texts or images) using generative models. GenAI models learn the patterns and structures of their training data and then generate new data with similar properties. Chatbots such as ChatGPT or text-to-image generators such as Midjourney and DALL-E are examples of such systems.

LLMs are an example of Generative AI. Basically, LLMs are prediction machines that generate a word sequence based on their input. The task of an LLM is to predict the next word in the sequence – nothing more. If you enter a few letters on your smartphone and the device suggests a word that matches the input, this is an example of a language model that predicts the end result (the word to be written) based on the context (i.e. the characters already entered).

These models are characterized by very intuitive and often unconscious “cognitive abilities”. Currently, they do not have true cognitive abilities – these would require conscious, slower and analytical thought processes. This limitation underlines the fact that LLMs can generate logical text but are not capable of deep thought or reasoning. Therefore, tasks that require reasoning and step-by-step analysis are a major challenge for LLMs, but one that is being actively researched.

It should also be noted that LLMs are not search engines. They have no real-time connection to the Internet or external databases. Their knowledge is limited to the data they have been trained with. The models therefore often lack currently available information or have adopted certain tendencies from the data with which they were trained. To circumvent these limitations, the model can be equipped with so-called “plug-ins”, which provide tools such as an internet search engine or a calculator. These can support the model in creating its answers.

LLMs are also seen as a way to democratize AI, as they allow people and companies to use powerful artificial intelligence without in-depth technical knowledge. After all, the “programming language” for these powerful AI systems is simply language! In addition, the knowledge contained in the huge training corpus becomes accessible to everyone in a much more intuitive, personalized and context-specific way.

Prompting is an important aspect when using LLMs. A prompt is the original input to the model. It generates its response on the basis of this input. Seemingly synonymous inputs can lead to significantly different results. The art of formulating these prompts in such a way that they lead to the desired results is known as “prompt engineering”. This is particularly important if the LLM is to reliably fulfill a specific task – for example in sentiment analysis.

Currently, LLMs are often integrated into the backend of other applications. The application then uses prompt templates that are customized with user data to perform the corresponding tasks. Since they do not have to formulate prompts, users are not even aware that they are interacting with an LLM behind the scenes.

In the context of LLMs, training refers to the process by which a model learns from a huge amount of text data in order to understand language patterns and generate texts or answer questions based on them. The model is not explicitly programmed to perform complex language tasks; rather, it learns to perform these tasks independently by analyzing the relationships between words and sentence structures in the training data.

During the training process, the LLM is presented with an extensive database of texts, which can include both the input texts and the desired outputs. By processing this data, the model adjusts its internal weights to replicate the nuances of human speech as accurately as possible.

The goal is to develop a model that is capable of responding accurately and coherently based on the learned patterns and structures not only on the training data, but also on new, unfamiliar texts.

Pre-training is a special step in the GenAI training process, especially for large language models. Here, the model is first trained on an extensive, generic collection of data before it is further trained for specific tasks. This preparatory step enables the model to gain a deep understanding of the language, its structures and patterns.

In fine-tuning, the pre-trained model is further trained with a more specific data set in order to adapt it to a very specific task or domain. During fine-tuning, the weights and parameters of the model are slightly adjusted to optimize performance on specific tasks such as chat, text classification, sentiment analysis or question-answer systems. Fine-tuning allows the extensive prior knowledge acquired by the model during pre-training to be used effectively for narrower use cases.

The goal of alignment in LLMs is to ensure that the outputs of the model are not only technically accurate, but also ethically responsible and in line with the goals and values of users and society. For LLMs, this requires careful monitoring and adjustment of training data, model settings and output checks to avoid bias, unwanted content and ethically problematic responses.

Tokens are the basic units of processing in an LLM. A token can be a word, part of a word or even a punctuation mark. The decomposition of text into tokens is an important step in making the input text processable for the model. This tokenization allows the LLM to analyze an input and generate meaningful output based on the context learned during training and pre-training.

One of the biggest challenges for LLMs are “hallucinations”, where the model’s response contains false or misleading information but is presented as fact. This usually happens because LLMs are trained to generate plausible-sounding texts based on certain patterns – and not to verify the truth of the information.

This problem is currently being tackled with the help of retrieval augmented generation (RAG). In short, this involves providing the model with documents that can be used to answer a query. More specifically, when users write a query, the system first searches its current database for passages that are likely to help answer the query. The LLM then generates an answer to the query based on the passages found. This procedure has the additional advantage that the source documents can be linked and used as added information for verification after the original training. Individual access rights for the respective documents can be taken into account.