Abstract: The next generation of AI will not be defined by whether a model can produce a brilliant answer in one turn. It will be defined by whether an intelligent system can pursue long-horizon goals while the world changes, evidence becomes stale, agents disagree, and earlier actions cannot simply be undone. Meeting this challenge requires more than scaling models. It requires a System-2 architecture around them.

This keynote presents a research program connecting three stages on the path to artificial general intelligence: collaborative intelligence, System-2 reasoning, and wisdom. The first two stages are developed in my ACM Books volumes on Multi-LLM Agent Collaborative Intelligence and System-2 Reasoning; the third motivates the forthcoming final volume of the trilogy. Drawing on this program and our recent papers, I will show how multi-agent collaboration can be made operational through durable memory and transactional execution; how untrusted generative proposals can be independently validated, admitted, and locally repaired; how world-model predictions can be treated as auditable materialized views rather than self-validating truth; and how causal evaluation can distinguish genuine reasoning from answers that are correct for the wrong reasons.

Together, these systems point toward a common architecture built from semantic anchoring, independent validation, persistent state, bounded recovery, replayable provenance, and checks and balances. They also expose the limit of intelligence understood solely as optimizing a fixed objective. Wisdom begins when a system can reconsider the objective itself across longer temporal horizons, wider relational boundaries, and decisions with irreversible consequences.

I will argue that the path from today’s language models to operational AGI, and eventually to wisdom-oriented AI, lies not in a single larger model but in an auditable and adaptive system that knows what it may commit to, what it must verify, how it should recover, and when it ought to reconsider.

Related Books and Papers:

• Multi-LLM Agent Collaborative Intelligence: The Path to Artificial General Intelligence, Volume 1 (ACM Books) https://dl.acm.org/doi/book/10.1145/3749421
• System-2 Reasoning: From Semantic Anchoring to Causal Intelligence: The Path to Artificial General Intelligence, Volume 2 (ACM Books) https://dl.acm.org/doi/book/10.1145/3822380
• TRW: TRACE-RealWorld—An Auditable Consistency Contract for World Models as Materialized Views (Edward Y. Chang, arXiv:2607.21910) https://arxiv.org/abs/2607.21910

Bio: Edward Y. Chang is founder and CEO of SocraSynth, an AGI startup established at Stanford University in 2023. From 2019 to 2026 he served as Adjunct Professor of Computer Science at Stanford, directing AGI research and teaching reasoning, planning, and collaborative intelligence for AGI; he remains affiliated with Stanford as a Faculty Advisor in Clinical AI at the Graduate Schools of Business and Education and an invited guest lecturer. Since December 2025, he has served as Co-Editor-in-Chief of ACM Books. He is a Fellow of both ACM and IEEE, recognized for contributions to scalable machine learning and healthcare AI.

Chang has spent more than two decades engineering AI systems at three frontiers: scalable data-centric machine learning infrastructure adopted by industry and the open-source community; deployed healthcare AI honored with the Qualcomm Tricorder XPRIZE; and multi-agent System-2 architectures that unify his work on data quality, causal grounding, persistent memory, and collaborative intelligence into a framework for AGI. In Foundations of Large-Scale Multimedia Information Management and Retrieval (Springer, 2011), he argued for the primacy of data quality over model complexity, anticipating the data-centric AI movement by more than a decade. In The Path to AGI, Volume 1: Multi-LLM Agent Collaborative Intelligence (ACM Books, 2025) and Volume 2: The Quadrivium: A System-2 Architecture from AGI to ASI (ACM Books, 2026), he develops this thesis into a layered architecture for System-2 reasoning in multi-agent LLM systems. A third volume, Beyond Intelligence: From Operational AGI to Wisdom, is planned for 2027.

Additional honors include the NSF CAREER Award, the ACM SIGMM Test of Time Award for the 2001 paper Active Learning for Image Retrieval (awarded in 2020), and the Google Innovation Award. Before industry, Chang held a faculty appointment at UC Santa Barbara from 1999 to 2006, where he rose from assistant to full professor of computer science in six and a half years.

He holds a Ph.D. in Electrical Engineering and an M.S. in Computer Science from Stanford University, and an M.S. in Industrial Engineering and Operations Research from UC Berkeley.