CIKM26 Industry Day

Nihil Sub Sole Novum: AI Is (Still) a Data Management Problem

Abstract: As reasoning becomes mostly undifferentiated through commoditized models, what’s left to build? Drawing on our production and research experience building Bauplan, I’ll discuss AI systems through the lens of the data management lifecycle: the data that goes into an agent, during pre-training and at runtime, and the data an agent produces, including reasoning traces and changes to data systems.

We bridge the gap between these two perspectives through the powerful lexicon of data systems: versioning, isolation, consistency, and provenance. We argue that real-world AI system failures are no longer primarily due to a “lack of intelligence,” but to a lack of proper infrastructure. By analyzing data mixtures, context management, and evals through tried-and-tested database mental models, we show that what’s old is indeed new again.

Bio: Educated in several acronyms across the globe (UNISR, SFI, MIT), Jacopo is the co-founder of Bauplan, a data and AI infrastructure company. Previously, he co-founded Tooso (acquired by TSX:CVO) and built Coveo AI Labs, whose libraries, models, and datasets have garnered thousands of citations and millions of downloads. His research spans AI, NLP, information retrieval, data management, and computer systems, with collaborations at Netflix, NVIDIA, Stanford, and Wisconsin-Madison. While building Bauplan, he moonlights as a professor of AI Systems at NY, which is notable because it’s the only job he’s ever had that his parents understand.

Abstract: Over the past two decades, AI has moved from explicit representations of meaning and knowledge to increasingly powerful Large Language Models (LLMs) capable of remarkable generalization. Yet many of the challenges that motivated computational semantics in the first place, i.e. multilinguality, grounding, interpretability, reliability, and controlled access to structured knowledge, remain highly relevant in the LLM era. In this talk, I will revisit this evolution through the lens of BabelNet, the Italian LLM Minerva, and its AI assistant ChatMinerva, highlighting what we have learned from building large-scale semantic resources and generative AI systems. I will argue that the future is not simply about replacing symbolic semantics with LLMs, but about bringing meaning, knowledge, and reasoning back into the design of next-generation AI.

Bio: Roberto Navigli is Professor of Natural Language Processing at Sapienza University of Rome, where he leads the Sapienza NLP Group. He is a Fellow of ACL, AAAI, ELLIS, and EurAI. He has received two ERC grants on multilingual lexical and sentence-level semantics, highlighted among the 15 projects through which the ERC has transformed science, as well as several prizes and awards from leading AI and NLP journals and conferences. He is the creator of BabelNet and leads the Minerva LLM and ChatMinerva projects - the first family of LLMs and AI assistants pretrained with a focus on Italian, and among the very few initiatives of this kind in Europe. He is also Scientific Director and co-founder of Babelscape, a successful deep-tech company with ten years of experience in research-driven multilingual neuro-symbolic NLP. He is an Associate Editor of JAIR and an Action Editor of TACL, served as General Chair of ACL 2025, and previously served as an Associate Editor of the AI Journal.