2025-06-02-Bernstein-全球软件_与领先人工智能研究员关于智能体人工智能讨论的要点_14页_349kb

Agentic AI is the latest development in Generative AI, currently in its early stages of formation, with limited technological maturity. Key takeaways from a discussion with AI researcher Dr. Ruairidh Battleday include:

• Definition: An AI agent serves human goals through intelligent actions in the digital or physical world, involving multi-step reasoning and tool use, beyond basic Generative AI capabilities.
• Hallucinations: Current models (e.g., GPT-4.5) exhibit high hallucination rates due to messy data and training algorithms; no definitive solution exists, but refining reasoning or data collection could help.
• Timeline: Agent capabilities may mature in about 3 years for reliable "set-and-forget" deployment; broader adoption of multi-agent systems could take an additional 2–3 years.
• Multi-Agent Systems: Challenges include poor orchestration and communication; linking agents with reasoning reduces errors in structured workflows but remains unsolved for complex, decentralized interactions.
• DeepSeek: While no major algorithmic innovation, DeepSeek demonstrated cost savings through efficiency, shifting focus from building larger models to incremental improvements.
• Investment: Companies like SAP, Oracle, and Adobe are building targeted agents now, while platforms for enterprise use are longer-term; Agentic AI could disrupt industries but maturity is key.
• Regional Differences: The US leads in Agentic AI development; Europe focuses on regulation and infrastructure, while Japan views AI agents positively with limited enterprise use.
• Overall, Agentic AI holds significant potential for productivity boosts but requires further innovation and infrastructure development.