AI可可AI生活

[人人能懂] AI的动态思维、可塑信念与捷径法则

你真的了解那个天天与你对话的AI吗?这一期,我们来当一回“AI读心师”,带你换个全新的视角看AI。我们会潜入AI思考的“时间之河”,揭示它那颗会悄悄“变心”的内在。更重要的是,我们将看到几篇最新论文,是如何教会AI聪明地“抄近道”、真正地“辨因果”,并最终找到那个任你怎么问都不会动摇的“坚固答案”的。

00:00:30 AI的“时间盲区”:我们看懂它的方式,可能一开始就错了

00:05:35 那个天天陪你聊天的AI,正在悄悄“变心”

00:11:23 高手过招:如何聪明地“抄近道”?

00:16:33 想用AI解决数据难题?你得先学会给它“立规矩”

00:23:46 换个姿势再问一遍:如何找到最可靠的答案?

本期介绍的几篇论文:

[LG] Priors in Time: Missing Inductive Biases for Language Model Interpretability

[Goodfire AI & Harvard University]

https://arxiv.org/abs/2511.01836

---

[CL] Accumulating Context Changes the Beliefs of Language Models

[CMU & Princeton University]

https://arxiv.org/abs/2511.01805

---

[RO] SLAP: Shortcut Learning for Abstract Planning

[Princeton University & CMU]

https://arxiv.org/abs/2511.01107

---

[LG] A Technical Exploration of Causal Inference with Hybrid LLM Synthetic Data

[UC Berkeley]

https://arxiv.org/abs/2511.00318

---

[CL] Self-Harmony: Learning to Harmonize Self-Supervision and Self-Play in Test-Time Reinforcement Learning

[The University of Tokyo & RIKEN Center for Advanced Intelligence Project]

https://arxiv.org/abs/2511.01191