Language remains
Computational linguistics already existed when I enrolled as a freshman at university more than half a century ago. In another life I would have liked to have chosen this subject, in this life I decided to study a different language.
I wrote my diploma thesis on an electric typewriter, but soon afterwards the first personal computers arrived in offices around the world.
Keyboards have replaced handwritten writing. Language and writing have not been replaced. Without wishing to wax philosophical, it is safe to say that both are aspects of what it means to be human.
Artificial intelligence will not change this, or vice versa: large-language models are not proving to be successful in generative AI from somewhere.
Humans and AI are learning
At the moment, it can be dizzying to try to keep track of the development of artificial intelligence. Rigid and rule-based approaches have long since evolved into systems that are constantly learning. Their huge advantage lies in the vast amounts of data they can process in a very short space of time.
Machine learning (ML) is the keyword for the method that data systems use to learn from the collected data.
This technical capability arouses human fears. A frequently voiced fear is that artificial intelligence will make human creativity superfluous in the long term.
We are still a long way off that. Let's take a step back and look at what LLM is currently doing.
What are Large Language Models?
LLMs are highly complex neural networks that are capable of both processing and generating human language. The term large actually refers to their enormous size: current models have billions of parameters. This scale enables them to recognize even subtle patterns in speech and generate context-dependent responses.
LLMs can not only help to write SEO-friendly blog articles. One possible use in linguistic research, for example, was recently presented in an episode of the podcast Sozusagen. It is about how centuries-old language textbooks can not only be scanned with the help of current technical possibilities, but above all made searchable. It is not simply a matter of listing vocabulary. The old sources are processed in such a way that grammatical features and word type assignments are also recognizable. Computational linguistics in 2025.
No LLM without training
Practice makes perfect - this old adage also applies to artificial intelligence. LLMs undergo training in several phases. First, the model is fed with existing text data from the internet, then fine-tuned and optimized for various tasks and guidelines. In the final step, human feedback is added (RLHF: Reinforcement Learning from Human Feedback) so that an LLM learns to generate useful and meaningful answers for humans.
Areas of application are in text creation and translation, in programming and coding, in learning programs and in context-related communication - with all justified reservations.
Limits and challenges of LLM
Artificial intelligence is still in its infancy. From technical challenges such as energy consumption during training and the risk of hallucinations, false reports and incorrect combinations, to ethical problems with data protection, copyright, bias in training data, potential misinformation, transparency and traceability, many important issues are still unresolved.
Nevertheless, none of this will slow down the rapid development of LLM. Work is constantly being done on more efficient training methods, which should lead to improved factual accuracy, among other things.
And of course, despite their name, LLMs are no longer limited to speech and text. New models are providing more and more options for generating images, audio and video. LLMs are increasingly specializing in very specific use cases.
And now?
Whatever you think of generative AI: LMAA is a bad attitude towards it. The Wiktionary https://de.wiktionary.org/wiki/leck_mich_am_Arsch explains what LMAA is.
It makes more sense to stay on the ball, i.e. to find out and try out where AI could actually be a useful addition to everyday business.
