![[人人能懂AI前沿] AI的智慧升级：元认知、B计划与组合式决策](https://image.xyzcdn.net/FuDP4HpAp8ezgVZMmEel3mblKCmJ.jpg@middle)
主播
节目简介
来源：小宇宙
想知道AI如何学会给自己准备“B计划”以防不测，又如何像个聪明的财务顾问一样在预算内做出最优决策吗？本期我们将一探究竟，从让AI拥有“元认知”能力，到学会“两步一回头”的智慧工作法，再到化身“艺术家与工匠”的完美结合体。这些最新的AI论文，正在教AI如何更聪明地思考和工作，而不仅仅是更努力地计算。
00:00:30 你的“外挂”，也需要一个“外挂”
00:06:39 你的AI助手，需要一个“Plan B”
00:13:33 预算有限，如何做出最优决策？
00:20:07 为什么顶尖高手，都懂得“两步一回头”？
00:26:01 AI制药，高手对决还是联手坐庄？
本期介绍的几篇论文：
[AI] Meta-Harness: End-to-End Optimization of Model Harnesses
[Stanford University]
https://arxiv.org/abs/2603.28052
---
[LG] Next-Token Prediction and Regret Minimization
[Google Research]
https://arxiv.org/abs/2603.28499
---
[LG] Multiple-Prediction-Powered Inference
[MIT & Google Research]
https://arxiv.org/abs/2603.27414
---
[LG] High dimensional theory of two-phase optimizers
[Google DeepMind]
https://arxiv.org/abs/2603.26954
---
[LG] Scaling Atomistic Protein Binder Design with Generative Pretraining and Test-Time Compute
[NVIDIA]
https://arxiv.org/abs/2603.27950
00:00:30 你的“外挂”，也需要一个“外挂”
00:06:39 你的AI助手，需要一个“Plan B”
00:13:33 预算有限，如何做出最优决策？
00:20:07 为什么顶尖高手，都懂得“两步一回头”？
00:26:01 AI制药，高手对决还是联手坐庄？
本期介绍的几篇论文：
[AI] Meta-Harness: End-to-End Optimization of Model Harnesses
[Stanford University]
https://arxiv.org/abs/2603.28052
---
[LG] Next-Token Prediction and Regret Minimization
[Google Research]
https://arxiv.org/abs/2603.28499
---
[LG] Multiple-Prediction-Powered Inference
[MIT & Google Research]
https://arxiv.org/abs/2603.27414
---
[LG] High dimensional theory of two-phase optimizers
[Google DeepMind]
https://arxiv.org/abs/2603.26954
---
[LG] Scaling Atomistic Protein Binder Design with Generative Pretraining and Test-Time Compute
[NVIDIA]
https://arxiv.org/abs/2603.27950
