如何让AI更聪明、更可靠？这期节目，我们将颠覆你的好几个固有认知。我们会发现，让小模型拥有大师风范的最佳方式，竟是引入一场“鉴赏家”参与的博弈；而AI最好的记忆方法，有时反而是那个最“笨”的。接着，我们将探讨如何用一张“考试大纲”驯服AI，又如何给它内置一个“苏格拉底”进行自我纠错。最后，我们还会揭秘，AI是如何从仅仅“听到”音乐，进化到能够“听懂”音乐背后的高级情感与故事的。00:00:37 让你的小模型，拥有宗师风范00:05:09 为什么说，最笨的方法，是AI最好的记忆方法？00:10:30 AI的“考试大纲”：我们如何让它更听话？00:15:54 如何让AI少犯错？给它一个内置的“苏格拉底”00:21:06 从“好听”到“高级”：AI如何学会聊音乐？本期介绍的几篇论文：[CL] Black-Box On-Policy Distillation of Large Language Models [Microsoft Research] https://arxiv.org/abs/2511.10643 ---[CL] Convomem Benchmark: Why Your First 150 Conversations Don't Need RAG [Salesforce AI Research] https://arxiv.org/abs/2511.10523 ---[CL] Rubric-Based Benchmarking and Reinforcement Learning for Advancing LLM Instruction Following [Meta Superintelligence Labs & Princeton University] https://arxiv.org/abs/2511.10507 ---[CL] SSR: Socratic Self-Refine for Large Language Model Reasoning [Salesforce AI Research] https://arxiv.org/abs/2511.10621 ---[AS] Music Flamingo: Scaling Music Understanding in Audio Language Models [NVIDIA & University of Maryland] https://arxiv.org/abs/2511.10289

Ask this episode

“What did the council decide on item 42?” — answered with the speaker and the moment.

Transcript

Related