[人人能懂] 如何打造一个我们能真正信赖的AI？

00:00:27  AI的“心术”：我们能给它装上一个诚实开关吗？00:05:08 比正确更重要的，是正确地思考00:10:18  省钱的艺术：如何让“实习生”干好“专家”的活？00:15:44 AI的“一键删除”，真的能删除吗？00:21:08 AI训练的“终局思维”：把答案直接告诉你本期介绍的几篇论文：[LG] Can LLMs Lie? Investigation beyond Hallucination  [Carnegie Mellon University (CMU)]  https://arxiv.org/abs/2509.03518  ---[LG] Beyond Correctness: Harmonizing Process and Outcome Rewards through RL Training  [Amazon]  https://arxiv.org/abs/2509.03403  ---[LG] Cut Costs, Not Accuracy: LLM-Powered Data Processing with Guarantees  [University of California, Berkeley]  https://arxiv.org/abs/2509.02896  ---[LG] Unlearning That Lasts: Utility-Preserving, Robust, and Almost Irreversible Forgetting in LLMs  [University of Tübingen & EPFL]  https://arxiv.org/abs/2509.02820  ---[LG] Imitate Optimal Policy: Prevail and Induce Action Collapse in Policy Gradient  [University of Sydney & King Abdullah University of Science and Technology]  https://arxiv.org/abs/2509.02737  在小宇宙查看该单集文稿