Take a sad song and make it better: Intriguing use cases for applying LLMs  

Vladan Deved?i? is a Professor of Computer Science and Software Engineering at the University of Belgrade, Faculty of Organizational Sciences. He is the founder and Head of the research group focused on Artificial Intelligence (GOOD OLD AI research network). He is also the founder of the Artificial Intelligence Laboratory at his home faculty. Since 2021, he has been a corresponding member of the Serbian Academy of Sciences and Arts (SASA) at the Department of Technical Sciences. According to the list of world's top scientists, published by Stanford University, he is among 0.6% of the most cited researchers in the "career" category in the field of Artificial Intelligence (for the period 1996-2025). His current interests include Artificial Intelligence, Programming Education, Software Engineering, and Educational Technologies. He has authored/co-authored numerous research papers, published in international and national journals or presented at international and national conferences, as well as six books on Intelligent Systems and Software Engineering. See the list of his most important research publications. He has also given dozens of keynote talks, invited talks, plenary talks and tutorials at international research conferences. His international research cooperation and collaboration record includes universities and research institutes from EU, USA, Canada, Japan, Australia, New Zealand, India and China. More recently, he has also initiated collaborations with universities in South America. See the list of research and development projects that have resulted from these cooperation and collaboration activities. Reference: This is a talk about practical, both-feet-on-the-ground applications of Large Language Models (LLMs).  It moves past the general hype to showcase where they truly deliver value, and where they still "need polishing". It explores several intriguing usage areas where LLMs come handy, but are still not completely changing the game. First, in the field of Software Engineering (SE), LLMs are great for coding assistance. Still, it is also important to focus on the real-world pitfalls: trying to use an LLM to solve all programming tasks can actually be detrimental to code quality, security, and developer skill. The talk shows examples of why LLMs in SE should still be better used as a co-pilot, not an autopilot. Next up is Music Information Retrieval (MIR). It is actually a good idea to carefully look at how LLMs simplify the "sad song" of data collection, but also to what extent. They can be great assistants in terms of automating the messy work of creating and cleaning complex music datasets from the web – extracting structured details about songs, artists, and industry players that traditional scraping methods miss. However, it still comes only with a great deal of human oversight and cleanup. Finally, in medical research and applications, LLMs can truly help in covering practical tasks like synthesizing vast amounts of literature, drafting patient information in plain language, and aiding in diagnostic support. The talk illustrates very intriguing recent and novel applications of LLMs in these domains. Still, strictly addressing the issue of hallucinations and the non-negotiable need for human oversight are extremely critical points in medical research and applications.  This talk is a look at what LLMs do for us today, not what they might do tomorrow.