你有没有想过，让一个复杂的AI黑箱像线性代数一样清晰可控？或者，我们根本不用动它的大脑，只要教它一个新词，就能让它和我们心有灵犀？甚至，让AI原地反复“琢磨”就能变聪明，或者只动一个“微创手术”就能完成一次完美进化？今天，我们就来聊聊这些脑洞大开的最新论文，看看科学家们如何用最聪明的“笨办法”，撬动AI的智慧极限。00:00:32 AI黑箱的新“坐标系”：当非线性遇上线性代数00:06:59 给AI造个新词，就能跟它心有灵犀？00:12:07 AI学习的“错题本”智慧00:16:47 大模型微调的新思路：为什么动大手术还不如做个“微创”？00:21:45 聪明的笨办法：如何让模型原地变聪明本期介绍的几篇论文：[CL] WHO SAID NEURAL NETWORKS AREN’T LINEAR? [Preprint] https://arxiv.org/abs/2306.14006 ---[CL] NEOLOGISM LEARNING FOR CONTROLLABILITY AND SELF-VERBALIZATION [Preprint] https://arxiv.org/abs/2401.01401 ---[CL] Agent Learning via Early Experience [Preprint] https://arxiv.org/abs/2309.13866 ---[CL] SLICEFINE: THE UNIVERSAL WINNING-SLICE HYPOTHESIS FOR PRETRAINED NETWORKS [Preprint] https://arxiv.org/abs/2307.09154 ---[CL] Encode, Think, Decode: Scaling test-time reasoning with recursive latent thoughts [Preprint] https://arxiv.org/abs/2305.14194

Ask this episode

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

Transcript

Related