Language processing and communication
LLMs such as GPT-3.5 have achieved an unprecedented capability for natural language processing. These models not only understand the context of texts, they can also generate human-like text. These advances have far-reaching effects in a range of fields. In communication they make possible more precise chatbots that can understand complex enquiries and deliver appropriate answers.
Image and speech processing
The integration of LLMs into image and speech processing has led to ground-breaking advances. Machines can not only recognise images with impressive accuracy, they are also able to generate context-related descriptions. Voice assistants understand nuanced requests better and can offer personalised and precise answers.
Significance for research and development
Beyond that, LLMs also have an enormous influence on research and development. Through their ability to process huge amounts of data and recognise patterns, they help solve complex problems in fields such as medicine, science and engineering. The collaboration between researchers and these systems, which now appear almost intelligent, is accelerating progress in many disciplines.
Ethical and legal questions
These advances do, however, also raise ethical and social questions. The need for clear guidelines and standards for the responsible use of LLMs, referred to as AI, is becoming ever more pressing. The EU AI Act (Regulation 2024/1689) obliges providers of AI systems to ensure transparency and risk management. LLMs vs. classic AI – more on this in our FAQ. Integrating LLMs requires careful balancing between innovation and the protection of privacy, as well as the prevention of potential misuse.
What is a large language model?
A large language model, or LLM for short, is an AI model that has been trained on very large volumes of text and both understands and produces natural language. It breaks every input down into small units called tokens and computes an answer from them. Well-known examples are GPT from OpenAI, Claude from Anthropic, Llama from Meta and Gemini from Google. LLMs can write, translate and summarise texts, generate program code and answer complex questions.
How does an LLM differ from classic artificial intelligence?
Classic AI systems are trained for a single, clearly defined task, such as image recognition or fraud detection, and can do nothing beyond it. Large language models, by contrast, are general-purpose systems: one and the same model answers questions, drafts texts, classifies documents and writes program code without having been specially trained for each of these tasks. This flexibility is the reason why LLMs have changed so many workflows since 2023.
Can a language model be run inside your own company?
Yes. Open models such as Llama, Mistral, Qwen or Gemma can be run on your own hardware; in many cases a single GPU workstation is enough. Running a fully trained model in day-to-day operation, known as inference, requires orders of magnitude less computing power than training it. This means no request leaves your own network, there are no costs per request, and there is no dependency on a provider.
What does the EU AI Act require of users of AI systems?
Regulation 2024/1689, known as the EU AI Act, obliges providers and deployers of AI systems to ensure transparency and appropriate risk management, among other things. The requirements depend on the risk class of the particular application. For companies this means above all: document which system is used for which purpose, and disclose when content has been generated with artificial intelligence.
What does the abbreviation LoRA stand for in fine-tuning?
LoRA stands for Low-Rank Adaptation, a resource-efficient method for adapting existing language models. Instead of recalculating all model weights, which would require a data centre, only a few small additional matrices are trained and attached to the model. The result is a file of a few megabytes instead of many gigabytes. The QLoRA variant adds quantisation to the method, making fine-tuning possible on a single graphics card.
