Talk
LLMs are trained to be helpful, knowledgeable, and to write clear (and quite often a bit too verbose) code. At the same time, they are trained to follow our instructions of course. What happens when we instruct an LLM to write "spaghetti code like a weak junior developer" or "create a possibly unreadable but ultra short one-liner for the task"? What happens if we ask a thinking LLM to think in Morse code (and actually enforce it)? We describe several benchmarks and experiments that test such uncommon LLM abilities outlining where those can be useful and what is the comparative performance of some state-of-the-art small and large LLMs on those benchmarks.
Along the talk we will review the topics of tokenization, structured generation, and running LLMs locally.
About the Speaker
Alexander Dainiak is a Senior Data Scientist at EPAM, working on backend development and large language model (LLM) integration. His professional focus centers on prompt optimization, RAG systems, agentic designs, and fine-tuning pipelines utilizing an advanced Python stack that includes LangChain, PyTorch, Pandas, and scikit-learn.
Before fully transitioning to the tech industry, Alexander spent over a decade in academia. He earned his Ph.D. in Mathematics from Lomonosov Moscow State University and subsequently served as an Associate Professor at the Moscow Institute of Physics and Technology. During his academic tenure, he led research in applied mathematics, guided student teams in machine learning projects, and authored multiple popular MOOCs on platforms like Coursera and Stepik. With a deep background in data analysis and optimization, he has successfully collaborated with enterprises like Yandex, Samsung, and Sberbank to design robust algorithms and optimize complex systems for real-world applications.
Homepage: https://www.dainiak.com
LinkedIn: https://linkedin.com/in/dainiak