Snowflake's Cortex AISQL is a production SQL engine that integrates native semantic operations directly into SQL. This integration allows users to write declarative queries that combine relational operations with semantic reasoning, enabling them to query both structured and unstructured data effortlessly. However, making semantic operations efficient at production scale poses fundamental challenges. Semantic operations are more expensive than traditional SQL operations, possess distinct latency and throughput characteristics, and their cost and selectivity are unknown during query compilation. Furthermore, existing query engines are not designed to optimize semantic operations. The AISQL query execution engine addresses these challenges through three novel techniques informed by production deployment data from Snowflake customers. First, AI-aware query optimization treats AI inference cost as a first-class optimization objective, reasoning about large language model (LLM) cost directly during query planning to achieve 2-8× speedups. Second, adaptive model cascades reduce inference costs by routing most rows through a fast proxy model while escalating uncertain cases to a powerful oracle model, achieving 2-6× speedups while maintaining 90-95% of oracle model quality. Third, semantic join query rewriting lowers the quadratic time complexity of join operations to linear through reformulation as multi-label classification tasks, achieving 15-70× speedups with often improved prediction quality. AISQL is deployed in production at Snowflake, where it powers diverse customer workloads across analytics, search, and content understanding. 

Bio: Anupam Datta is a Principal Research Scientist and Snowflake AI Research Lead at Snowflake. He joined Snowflake as part of the acquisition of TruEra where he served as Co-Founder, President, and Chief Scientist from 2019-2024. Datta was on the faculty at Carnegie Mellon University from 2007-2022, most recently as a tenured Professor of Electrical & Computer Engineering and Computer Science. Datta's current research focuses on Trustworthy AI, especially evaluation, optimization, and security for data and coding agents. Prior results include early work on Shapley Values & gradient-based explanations, fairness assessments, robustness of classical machine learning and deep learning models for natural language processing and computer vision. These research results have had a significant impact on products at TruEra and Snowflake. Datta has published over 100 research papers, served as Chair of the National Academies Workshop on Assessing and Improving AI Trustworthiness, on the Steering Committee of of the ACM Conference on Fairness, Accountability, and Transparency, and the IEEE Computer Security Foundations Symposium, and as an Editor-in-Chief of Foundations and Trends in Privacy and Security. He received the 2018 David P. Casasent Outstanding Research Award from the CMU College of Engineering, a 2020 Young Alumni Achiever Award from IIT Kharagpur, a 2021 Google Faculty Research Award, and several awards for top papers at conferences. Datta obtained a B.Tech. from IIT Kharagpur, and Ph.D. and M.S. degrees from Stanford University in Computer Science, where he currently teaches a course on Trustworthy AI. 

Data ecology studies how dataflows shape data ecosystems. As AI agents become active participants in organizations, the design of dataflows increasingly determines what information agents can access, how efficiently they can act, and how organizations govern their behavior. Today, however, these dataflows evolve largely organically rather than by design. This talk presents a systems vision for the emerging infrastructure of agentic organizations. I will describe Pneuma, which enables fast and cost-efficient structured data work for AI agents; internal data markets, which allocate valuable context to the agents and people who need it most; and ongoing work on "existential documents", a project that provides the foundation for governing organizational dataflows. Together, these efforts illustrate how data ecology can guide the design of the socio-technical infrastructure organizations need to fully realize the potential of AI agents.

Bio: Raul Castro Fernandez is an assistant professor in the Department of Computer Science at the University of Chicago, where he is also affiliated with the Committee on Data Science. His research agenda, data ecology, asks how data shapes our technological, economic, social, and cultural systems, and how to design interventions that let us realize the benefits of data while reducing its harms. He studies dataflows as the unit of analysis and build interventions that are primarily technical — including data escrows, data-sharing protocols, and agentic systems for data discovery and processing — complemented by economic, social, and institutional mechanisms. He received the SIGMOD Test-of-Time Award in 2023, an NSF CAREER Award in 2024, and was named a Sloan Research Fellow in 2025.

Scaling foundation models appears to be offering us a free lunch, but I argue that it is a bitter free lunch. The models are making broad one-size-fits-all competence cheap, but this leaves the last-mile information that adapts AI to a downstream environment extremely expensive to elicit and learn from. In this talk, I’ll describe some of our recent efforts developing methods for updating the behavior of computer systems from qualitative downstream information, whether this information takes the form of arbitrarily long user prompts, a large corpus of documents, or an interactive environment. I’ll first discuss Recursive Language Models, which treat long prompts as a symbolic object and train LLMs to process it fundamentally via recursion, and thereby achieve remarkable performance and generalization. I’ll then present a different interface to the problem of information retrieval, namely Machine Studying, and use it to argue that current data agents are severely limited. Finally, I’ll introduce Pedagogical RL, a paradigm of optimizing models to teach themselves, allowing far more efficient learning than the dominant algorithms oriented around on-policy trial and error. Together, these projects extend our earlier work on declarative and self-improving AI systems toward a broader view of how to express complex needs and align AI systems with them.

Bio: Omar Khattab is an Assistant Professor at MIT EECS and CSAIL. His research develops models, abstractions, and algorithms for building reliable and scalable AI systems and for learning from sparse and qualitative feedback. He created the ColBERT retrieval model, which has helped shape the modern landscape of neural information retrieval, and the DSPy framework, which is the first and most widely used declarative programming model for LLM systems. His lines of work on ColBERT, DSPy, GEPA, and RLMs each form the basis of influential open-source projects, together downloaded over one hundred million times to date, have sparked applications at dozens of organizations, were supported by the Apple Scholars in AI/ML PhD Fellowship, and received a Best Paper Award at SIGIR 2025.