谷歌 DeepMind CEO Demis Hassabis 探讨了过去一年 AI 的进展、通往 AGI 的路径，以及当前大语言模型在持续学习和推理方面是否需要新的突破。
Google DeepMind CEO Demis Hassabis discusses AI progress over the past year, the path to AGI, and whether current LLMs need new breakthroughs for continual learning and reasoning.
要点 · TL;DR
AGI 需要所有人类认知能力，还需 5-10 年。 AGI requires all human cognitive abilities, 5-10 years away.
持续学习和世界模型是关键缺失环节。 Continual learning and world models are critical missing pieces.
带通用助手的智能眼镜是下一个形态。 Smart glasses with universal assistant are next form factor.
核心观点 · Key points
AGI 需要包括创造力和物理智能在内的所有人类认知能力，还需 5-10 年。 AGI requires all human cognitive abilities including creativity and physical intelligence, 5-10 years away.
持续学习和更好的记忆是当前 AI 系统缺失的关键部分。 Continual learning and better memory are critical missing pieces for current AI systems.
世界模型对长期规划和 AGI 至关重要，视频生成已展示这一点。 World models are essential for long-term planning and AGI, as shown by video generation.
配备通用助手的智能眼镜是下一个形态，预计夏季推出。 Smart glasses with a universal assistant are the next form factor, likely launching by summer.
AI 将改变科学和医学，AlphaFold 已改变蛋白质折叠。 AI will transform science and medicine, as AlphaFold already has for protein folding.
信息是宇宙最基本的单位，是解决复杂问题的关键。 Information is the most fundamental unit of the universe, key to solving complex problems.
反共识 · Contrarian takes
AGI 还很遥远；当前系统缺乏真正的创造力和物理智能。 AGI is not near; current systems lack true creativity and physical intelligence.
仅靠 Scaling 现有技术可能无法实现 AGI；需要新的突破。 Scaling existing techniques alone may not achieve AGI; new breakthroughs are needed.
LLM 并非死胡同；它们是关键组件，可能是 AGI 的唯一组件。 LLMs are not a dead end; they are a key component, possibly the only one for AGI.
尽管市场猜测，Google 目前没有在 Gemini 应用中添加广告的计划。 Google has no current plans to add ads to Gemini app, despite market speculation.
AI 行业部分存在泡沫，但 Google 无论哪种情况都处于有利位置。 Parts of the AI industry are in a bubble, but Google is well-positioned either way.
人类的目标和意义将随 AI 自动化演变，不一定会变糟。 Human purpose and meaning will evolve with AI automation, not necessarily worsen.
本期章节 · Chapters（共 17）
引言与过去一年 AI 进展Introduction and AI progress in the past year
对 LLM 局限性的怀疑与突破需求Skepticism about LLM limitations and need for breakthroughs
混合系统与 AGI 中学习的作用Hybrid systems and the role of learning in AGI
持续学习问题与潜在解决方案Continual learning problem and potential solutions
AGI 定义与时间线AGI Definition and Timeline
世界模型与视频生成World Models and Video Generation
AI 眼镜与形态因素AI Glasses and Form Factor
智能眼镜作为下一代形态Smart glasses as next form factor
编码能力与竞争Coding capabilities and competition
AI 行业泡沫与基础设施AI industry bubble and infrastructure
对知识工作的影响与游戏类比Impact on knowledge work and games analogy
人类对 AI 的适应Human Adaptation to AI
信息作为基本单元Information as Fundamental Unit
AlphaFold 发布决策AlphaFold Release Decision
AlphaFold 对药物发现的影响AlphaFold's impact on drug discovery
谷歌支持与科学方法Support from Google and scientific approach
从 AlphaGo 到 AlphaZero：超越人类知识From AlphaGo to AlphaZero: beyond human knowledge