Hallucination by Design - How Embedding Models Misunderstand Language

Explore how embedding models fundamentally misunderstand language and contribute to AI hallucinations in this 29-minute conference talk from Conf42 ML 2025. Begin with foundational concepts of embeddings and their evaluation methods, then dive into the embedding process and key underlying concepts. Examine practical use cases for embeddings and learn techniques for comparing different embedding approaches. Work through hands-on examples using OpenAI's embedding models to understand their practical applications. Investigate the critical challenges and hallucination phenomena that arise when embedding models process language, uncovering the systematic ways these models can misinterpret meaning and context. Discover fine-tuning techniques for embedding models, including step-by-step approaches to improve model performance and reduce misunderstandings. Gain insights into the limitations of current embedding technologies and understand why these models struggle with accurate language representation, ultimately learning how to identify and mitigate embedding-related hallucinations in AI systems.