High Bit ZeroEntropy: The Hidden Bottleneck in AI. Retrieval, Not Models
Jan 30, 2026
Ghita Houir Alami, cofounder and CEO of ZeroEntropy and retrieval specialist, explains why retrieval—not bigger models—is the real bottleneck for production AI. She discusses why embeddings fall short, how rerankers provide a crucial second pass, and training via pairwise comparisons and Elo scoring. The conversation covers distillation for fast rerankers, cross-domain generalization, recency and personalization, and what trustworthy retrieval unlocks.
AI Snips
Chapters
Transcript
Episode notes
Information Retrieval Is The Core Productivity Problem
- Most productive work is about finding and using the right information at the right time.
- ZeroEntropy focuses on improving retrieval accuracy and efficiency across messy knowledge stores.
Medical Coding Project Led To Retrieval Focus
- Building an automated medical coding system taught them the real bottleneck was finding the right context.
- That experience pushed them to build connectors and focus on retrieval rather than only models.
Embeddings Alone Don't Solve Ranking
- Embeddings find semantic clusters but often fail to provide fine-grained ordering for specific queries.
- Embeddings are not designed to guarantee perfect ranking for every potential question.