AI可可AI生活

[人人能懂] 从万能骨架、老司机到一副新眼镜


Listen Later

想让AI更聪明,为什么它有时反而会“学傻”?本期节目,我们将一起揭开AI训练中“差不多”哲学的代价,并探索如何为所有大模型打造一副省钱又省力的“万能骨架”。我们还会看到,有时只需给机器人加一点“噪声”,或者校准一下它看世界的“眼镜”,就能让它从新手秒变老司机。最后,我们将见证一个奇迹:如何让机器人看懂我们天马行空的“梦境”,将想象力直接翻译成物理世界的行动。

00:00:35 驯服AI这匹野马,问题出在了“差不多”上

00:07:43 给机器人加点“噪声”,它就变聪明了?这事没那么简单

00:14:08 怎么让机器人听懂你的想象力?

00:19:18 AI大模型们的“万能骨架”:省钱省力的秘密

00:23:03 机器人换个角度就犯傻?问题可能出在你没想到的地方

本期介绍的几篇论文:

[CL] Stabilizing Reinforcement Learning with LLMs: Formulation and Practices

[Qwen Team, Alibaba Inc.]

https://arxiv.org/abs/2512.01374

---

[RO] Much Ado About Noising: Dispelling the Myths of Generative Robotic Control

[CMU]

https://arxiv.org/abs/2512.01809

---

[RO] From Generated Human Videos to Physically Plausible Robot Trajectories

[UC Berkeley & Johannes Kepler University]

https://arxiv.org/abs/2512.05094

---

[LG] The Universal Weight Subspace Hypothesis

[Johns Hopkins University]

https://arxiv.org/abs/2512.05117

---

[RO] VLA Models Are More Generalizable Than You Think: Revisiting Physical and Spatial Modeling

[Sun Yat-sen University]

https://arxiv.org/abs/2512.02902

...more
View all episodesView all episodes
Download on the App Store

AI可可AI生活By fly51fly