Google DeepMind 的 Nana Tomashev 解释了 AI 智能体如何通过自主执行多步骤任务区别于语言模型，并探讨了它们创造新经济和通往 AGI 路径的潜力。
Google DeepMind's Nana Tomashev explains how AI agents differ from language models by autonomously performing multi-step tasks, and discusses their potential to create a new economy and path to AGI.
要点 · TL;DR
智能体作用于世界状态，而不仅是预测文本，能自动化复杂任务。 Agents act on world state, not just predict text, enabling automation of complex tasks.
安全需要分层防御：环境、智能体、模型和人类监督。 Safety requires layered defenses: environment, agent, model, and human oversight.
未来 AI 可能是专业智能体的分布式社会，而非单一通用人工智能。 Future AI may be a distributed society of specialized agents, not a single AGI.
核心观点 · Key points
智能体观察世界状态并执行动作，而语言模型仅延续提示。 Agents observe world state and perform actions, unlike language models that only continue prompts.
智能体通过自动化复杂任务加速进步，但因失败率需要人类监督。 Agents accelerate progress by automating complex tasks, but require human oversight due to failure rates.
安全需要纵深防御：在环境、智能体、模型和人类控制上设置多层缓解措施。 Safety requires defense in depth: multiple mitigations on environment, agent, model, and human controls.
未来的智能体经济将包含专才和通才，而非单一的超级智能AGI。 Future agentic economies will have specialists and generalists, not just one superintelligent AGI.