回应
按立场分组 · 点上方立场只看那一组
靠学习智能来自在世界中持续从经验和数据里学，抽象也得自己学到
仅靠海量人类生成的语言和图像数据训练，就足以让人工智能达到今天的能力水平
▶ 3:31梅拉妮·米切尔Are We Thinking Correctly About AI Intelligence? | PODCAST: The Joy of Why2026
《苦涩的教训》否定的不是精巧算法，而是无法随算力扩展的算法
▶ 9:29理查德·萨顿Rich Sutton and Khurram Javed: Why AI Models Stop Learning, and How to Start It Again2026
创造新概念的权重学习不该只发生在预训练一次，而应在模型使用过程中持续进行
▶ 23:21理查德·萨顿Rich Sutton and Khurram Javed: Why AI Models Stop Learning, and How to Start It Again2026
没有任何动物是靠监督学习来学习的，没人给得出肌肉该如何抽动的样例
▶ 26:52理查德·萨顿Rich Sutton and Khurram Javed: Why AI Models Stop Learning, and How to Start It Again2026
持续深度学习是通往通用人工智能最关键的一步，它能解锁其余所有能力
▶ 37:26理查德·萨顿Rich Sutton and Khurram Javed: Why AI Models Stop Learning, and How to Start It Again2026
不存在客观上正确的抽象，智能体必须自己在所处世界里学到什么抽象才算对
▶ 37:50理查德·萨顿Rich Sutton and Khurram Javed: Why AI Models Stop Learning, and How to Start It Again2026
靠先有关键在学之前就有的结构与世界模型，而不是数据量
从「苦涩的教训」推出「不要做架构研究」是错的，因为可用数据始终有限
▶ 40:14约翰·江珀He won a Nobel here for AlphaFold. Then he left. - John Jumper2026
要以类人的方式理解世界，AI 可能需要具备类似人类的核心知识系统及由此产生的概念
▶ 43:32梅拉妮·米切尔The Debate Over “Understanding” in AI’s Large Language Models2024
人类在学语言之前就具备的知识，是人能用远少于大语言模型的数据学会语言的关键
▶ 21:23乔什·特南鲍姆Josh Tenenbaum | How to grow a mind from a brain: From guessing and betting to thinking and talking2024
是心智创造了语言，而不是语言训练出心智
▶ 22:55乔什·特南鲍姆Josh Tenenbaum | How to grow a mind from a brain: From guessing and betting to thinking and talking2024
大脑最根本的计算功能不是学习，而是借助世界模型做出好的猜测与好的下注
▶ 31:01乔什·特南鲍姆Josh Tenenbaum | How to grow a mind from a brain: From guessing and betting to thinking and talking2024
刻画直觉物理不能只靠高维高斯或弱非线性，得把物理引擎式的模拟器放进概率模型的核心
▶ 40:54乔什·特南鲍姆Josh Tenenbaum | How to grow a mind from a brain: From guessing and betting to thinking and talking2024
要合起来通向类人智能要靠多种成分的综合，单一路线不够
通向类人智能的技术路径是神经、符号与概率三者的综合
▶ 1:08:22乔什·特南鲍姆Josh Tenenbaum | How to grow a mind from a brain: From guessing and betting to thinking and talking2024
人若只靠自身原始经验学习、从未被他人告知任何事，很可能算不上文明人
▶ 16:16苏巴拉奥·坎帕蒂[ICAPS 2020] Polanyi vs. Planning (Planning around AI's New Romance with Tacit Knowledge)2020
人类会做规划和长期的显性知识推理，所以推理路线与学习路线必须结合起来
▶ 19:55苏巴拉奥·坎帕蒂[ICAPS 2020] Polanyi vs. Planning (Planning around AI's New Romance with Tacit Knowledge)2020
纯自下而上的学习或许原则上能达成通用智能，但等待时间太长，满足不了产业界当下的需求
▶ 54:27苏巴拉奥·坎帕蒂[ICAPS 2020] Polanyi vs. Planning (Planning around AI's New Romance with Tacit Knowledge)2020
对隐性知识任务强求显式模型，与在显性知识任务上拒绝现成的部分模型，是同样愚蠢的错误
▶ 55:18苏巴拉奥·坎帕蒂[ICAPS 2020] Polanyi vs. Planning (Planning around AI's New Romance with Tacit Knowledge)2020
苏菲周报 · THE WEEKLY
每周一封，
追问一个大问题。 苏菲拉底的每周来信，写这一周在追问的问题和看到的回应。
追问一个大问题。 苏菲拉底的每周来信，写这一周在追问的问题和看到的回应。
苏菲拉底
ASK THE BIG QUESTIONS · THINK DEEPLY · SEE THE WORLD DIFFERENTLY
微信公众号
© 2026 苏菲拉底 · 内容仅供学习
[email protected]