EPISODE · Jun 3, 2026 · 29 MIN
[人人能懂AI前沿] 自我复盘、拥抱不确定与事前清醒
from AI可可AI生活
你有没有想过,一个AI要怎样才能超越过去的自己?本期节目,我们将从几篇最新的AI论文出发,探讨AI认知升级的奇妙路径。我们会发现,顶尖的AI也需要像人一样“睡觉”来巩固记忆,也需要一位“总设计师”来指挥“泥瓦匠”完成复杂工程。更重要的是,我们会看到AI如何学会从自己的成功经验里提炼方法论,如何跳出“标准答案”的陷阱拥抱创意,以及如何从“事后辟谣”进化到“事前清醒”。00:00:35 从自己的脚印里,发现通往未来的地图00:07:10 AI的“选择困难症”与“标准答案”陷阱00:12:54 AI说话,怎么从“后期辟谣”升级到“事先清醒”?00:17:54 为什么顶尖的AI,也需要“睡一觉”?00:23:02 AI界的“泥瓦匠”和“总设计师”本期介绍的几篇论文:[CL] Inducing Reasoning Primitives from Agent Traces [CMU] https://arxiv.org/abs/2606.02994 ---[LG] Using Reward Uncertainty to Induce Diverse Behaviour in Reinforcement Learning [Google DeepMind] https://arxiv.org/abs/2606.03962 ---[LG] Conformal Language Modeling via Posterior Sampling [MIT] https://arxiv.org/abs/2606.03731 ---[LG] Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories [Google Research & Cornell University] https://arxiv.org/abs/2606.03979 ---[AI] LEAP: Supercharging LLMs for Formal Mathematics with Agentic Frameworks [Google Cloud AI Research & Google Cloud & Google DeepMind] https://arxiv.org/abs/2606.03303 在小宇宙查看该单集文稿
Embed this episode
Ready to play
[人人能懂AI前沿] 自我复盘、拥抱不确定与事前清醒
No transcript for this episode yet
Similar Episodes
No similar episodes found.