Self-Evolving Data: How AI Learns to Create Its Own Data

Session Overview

AI is no longer just a system that learns about the world — it's starting to learn how to build it. The era of consuming data is ending, and the era of evolving data is beginning. The biggest shift in AI research right now isn't about building bigger models — it's about how to keep generating better data. This session explores Self-Evolving Data, a new learning paradigm in which AI generates its own data and then grows from it, and why it's emerging as a defining technology for the next generation of AI.

Human-generated data has already hit a ceiling. Most of the high-quality text and images available on the web have already been used for training, and human data production simply can't keep pace with how fast AI is advancing. The world's top research teams are all converging on the same answer: instead of relying on humans to produce data, AI now generates, evaluates, and retrains on its own data — a self-evolving loop that's becoming the new standard. This session looks at how that shift is already changing model performance today, and what it means for the coming era of Agentic AI and Physical AI.

Key Takeaways

• Why AI's competitive edge is now determined by data-generation capability, not model size.
• How the paradigm is shifting from human-made data to AI-evolved data.
• The core principle behind how Self-Evolving Data lifts the performance of reasoning models and AI agents.
• Why choosing and evolving better data matters more than simply training on more of it.
• What recent research suggests about where AI development is headed next.