Not Everything Needs LLM: Boosting, Embeddings and Rules Win in Production

At the UWDC 2026 conference in Yekaterinburg, a discussion was held on the choice between Large Language Models (LLMs), classic machine learning (ML), and traditional development. Participants—Kamil Shakirov (AI Engineer at DAR), Nikita Ovchinnikov (ML Engineer at RUTUBE), and Yan Anisimov (Tech Lead DS/DS+ at Yandex Practicum)—agreed that businesses often do not need a universal model but rather a system that effectively solves a specific problem. Classic ML wins with predictability, determinism, and ease of validation: models can be tested with metrics that are understandable to the business. LLMs, on the other hand, provide impressive answers but hallucinate, require constant monitoring and updates every three months, and their inference is expensive. In practice, hybrid architectures with rules and heuristics at the lower level, ML at the middle level, and LLMs at the upper level reduce costs and improve quality. Validation of LLM results remains a major challenge, requiring manual labeling and additional tooling. Experts noted that it is incorrect to pit ML against LLMs—each approach is suitable for its own tasks.