Nando de Freitas 探讨 AI 作为通用工具的潜力、他在 DeepMind 的工作，以及通过非洲 Indaba 和拉丁美洲 Khipu 等倡议推动 AI 民主化的努力。
Nando de Freitas discusses AI's potential as a universal tool, his work at DeepMind, and his efforts to democratize AI through initiatives like Indaba in Africa and Khipu in Latin America.
要点 · TL;DR
AI 应增强人类能力，而非取代。 AI should augment human capabilities, not replace them.
AI 社区的多样性对构建有益技术至关重要。 Diversity in AI community is key to building beneficial technology.
从视频学习的通用智能体是通往 AGI 的路径。 Generalist agents learning from video are a path to AGI.
核心观点 · Key points
AI 应成为扩展人类能力的工具，如同显微镜和望远镜。 AI should be a tool to extend human capabilities, like microscopes and telescopes.
AI 社区的多样性和包容性对于构建有益的技术至关重要。 Diversity and inclusion in AI community are essential for building beneficial technology.
大模型令人兴奋但并非全部，基础创新可以来自任何地方。 Large models are exciting but not everything; fundamental innovation can come from anywhere.
从视频（如 YouTube）中学习是训练智能体的有前途的方法，无需昂贵的强化学习。 Learning from video, like YouTube, is a promising way to train agents without expensive RL.
像 Gato 这样将所有输入标记化为序列的通才智能体是迈向 AGI 的一步。 Generalist agents like Gato that tokenize all inputs into sequences are a step toward AGI.
未来的 AI 应增强科学家和劳动者，解决能源、气候和经济问题。 Future AI should augment scientists and laborers, solving energy, climate, and economic problems.
反共识 · Contrarian takes
单一通用智能算法的想法可追溯到 1970 年代的神经科学。 The idea of a single universal algorithm for intelligence dates back to neuroscience in the 1970s.
通过梯度下降学习的学习如今在大语言模型的少样本提示中实现。 Learning to learn by gradient descent is now realized in few-shot prompting of large language models.
机器人应像人类一样从 YouTube 视频学习，而不仅仅从遥操作数据。 Robots should learn from YouTube videos like humans, not just from teleoperation data.
AI 中语言的涌现可能需要智能体相互欺骗，如同自然界中。 The emergence of language in AI may require agents to deceive each other, as in nature.
AlphaCode 的代码常包含未使用的循环，让人联想到学生添加无关内容。 AlphaCode's code often contains unused loops, reminiscent of students adding extraneous content.
最终目标是 AI 能创造新符号和知识，如同爱因斯坦的思想实验。 The ultimate goal is AI that can create new symbols and knowledge, like Einstein's thought experiments.
本期章节 · Chapters（共 14）
开场与引言Introduction and Opening
社区建设与多样性Community Building and Diversity
AI 助力全球健康等难题AI for Global Health and Other Problems
从专家到通用智能体From Specialist to General Agents
序列建模与统一模型的早期构想Early ideas on sequence modeling and single model for all tasks
通用智能体与视频学习Generalist agents and learning from video
机器人学习的视频方法Learning from video for robotics
通过梯度下降学会学习Learning to learn by gradient descent
学会学习与涌现能力Learning to Learn and Emergent Capabilities
智能中的欺骗与抽象Deception and abstraction in intelligence
未来 5-10 年 AI 的激动方向Exciting directions for AI in 5-10 years
AI 愿景：气候、经济与劳动Dreams for AI: Climate, Economics, and Labor