- [人人能懂AI前沿] 从构建底层引力场、AI的“通感”到让笨徒弟超越师傅

 - [人人能懂AI前沿] 从构建底层引力场、AI的“通感”到让笨徒弟超越师傅

Log ind eller registrer dig for at anmelde eller tilføje dette element til din samling.

Filtype ikke understøttet.
UUID: 5iGc2mILRcrEYSCtmzqdc7
Class: podcastepisode
Category: podcast

/ 10

0 bedømmelser

Ingen nok bedømmelser
del af Podcast-program: AI可可AI生活
overblik

今天我们将为你拆解五篇极具启发的AI最新论文,看看AI如何摆脱既定路线构建“引力场”一步直达目标,又是如何靠“脑补”预测未来学会干活的。我们还会探讨为什么最聪明的AI拧不开矿泉水瓶,揭秘AI如何学会人类的“通感”建立万物坐标系。最后,我们将一起看看让“笨徒弟”掌握火候并超越师傅的进阶算法,带你从前沿科技中汲取成事心法!

00:00:33 绕开既定路线,直达目的地

00:05:45 不用手把手教,AI是如何靠“脑补”学会干活的?

00:09:59 最聪明的人工智能,为什么拧不开一瓶矿泉水?

00:15:14 万物互通的翻译官,AI如何学会人类的“通感”?

00:19:47 别只盯着“及格线”,一个让笨徒弟超越师傅的进阶算法

本期介绍的几篇论文:

[LG] Discrete Beckmann Transport Models for One-Step Language Modeling and Reasoning

[University of Pennsylvania & Harvard University]

https://arxiv.org/abs/2609.15903

---

[RO] ThinkWorld:World Imagination Reinforces Chain-of-Dynamics Reasoning

[NVIDIA & KAIST]

https://arxiv.org/abs/2610.37666

---

[RO] How High Can GPT-6 Astra Climb in Contact-Rich Manipulation? PhysLadder: A Hierarchical, Matched Sim-Real Benchmark for Physical Interaction

[Tuojing Intelligence]

https://arxiv.org/abs/2610.34527

---

[LG] EmbeddingGemma 2: Unified and Efficient Multimodal Representations

[Google]

https://arxiv.org/abs/2610.37364

---

[RO] Learning Quality-Controllable Robot Policies with Iterative Hindsight Relabeling

[KAIST & LGE]

https://arxiv.org/abs/2610.37393

在小宇宙查看该单集文稿

kommentarer
Anmeldelser
Noter