[SAIF2025] Keynote: Advancing Agentic Intelligence скачать в хорошем качестве

[SAIF2025] Keynote: Advancing Agentic Intelligence

Professor, Joseph E. Gonzalez | University of California, Berkeley ※ 본 세션은 영어로 진행되었습니다. 한국어 동시 통역 버전은 여기서 확인하실 수 있습니다:    • [SAIF2025] Keynote: Advancing Agentic  Int...   (This session is originally in English. A Korean simultaneous interpretation version is also available at link above.) Large language models (LLMs) are evolving from passive completion engines into persistent agents capable of executing complex tasks over extended interactions. This talk presents recent efforts to enhance the agentic capabilities of LLM-based systems. One key concept introduced is "sleep-time compute," a paradigm that allows agents to reason, learn, and plan between interactions. This innovation unlocks new modes of intelligence beyond just reactive prompt completion. The presentation also discusses how to enable agents to mimic human behavior and reason about human preferences, paving the way for more socially aligned and intuitive AI. Additionally, it highlights early efforts on advisor agents—meta-agents designed to guide and critique other agents as they work—establishing the foundation for self-improving AI systems. Together, these advancements are bringing us closer to a future where LLMs operate as adaptive, collaborative, and trustworthy digital agents.