https://mayulu.co/tags/rl/Recent content in RL on mayulu的AI笔记HugozhcnCopyright © 2025–2026 mayulu All Rights ReservedSat, 08 Aug 2026 10:48:06 +0800https://mayulu.co/posts/the-way-to-agi-viewpoints/Thu, 06 Aug 2026 10:48:06 +0800https://mayulu.co/posts/the-way-to-agi-viewpoints/行业曾寄望依靠扩大算力、数据与离线 RL 规模实现 AGI，该范式如今遭遇越来越多质疑。本文结合 Jerry Tworek 与梁文锋的观点，剖析离线强化学习的固有局限，拆解从大模型、CoT、Agent、持续学习、自我迭代到具身智能的六级递进 AGI 技术阶梯。指出持续学习是工具 AI 迈向通用智能的分水岭，AGI 竞赛的关键不再是规模军备竞赛，而是原生自主成长机制的架构重构。