Miton AI Times

For most of their history, recommendation, search, and conversational interfaces have been built as separate systems, with distinct architectures, evaluation methods, and engineering teams. This separation is rapidly disappearing. Embeddings, foundation-model backbones, and large language models are pushing these three paradigms toward a single underlying stack, in which a query, a click, a scroll, and a natural-language request are increasingly different entry points into the same retrieval-and-ranking machinery. This talk examines what such convergence means in practice for the applications we build, and the new challenges it brings, both theoretical and practical. On the theoretical side, unifying interaction-based and semantic signals raises hard questions about evaluation, cold-start behavior, bias and alignment. On the practical side, deploying these systems in the real world forces an uncomfortable trade-off: they must be not only accurate, but also fast and economically viable at the scale of hundreds of millions of users — a constraint that reshapes architectural decisions far more than benchmark leaderboards suggest. Drawing on experience from both academic research and the production systems behind Recombee, the talk explores how to scale converged recommendation–search–conversation systems while keeping latency and cost under control, and how to ensure user satisfaction and AI alignment across very different domains, from news and media, through e-commerce, to education. The central argument is that the next generation of information systems will be defined less by raw model capability and more by how thoughtfully we engineer the interaction between accuracy, affordability, and alignment.
-------------------------------------------------------------