跳到正文
𝕏 Linus ✦ Ekenstam 🔜 Adobe MAX Miami 2026(x.com)·· 9 天前品質評分72
重點導讀

機器要在真實世界試錯,可能會造成損壞;PAN 2 提供一種先在模擬世界練習的方式。

  • MBZUAI 的基礎模型研究所推出這個世界模型,只要給它一個場景和一個動作,它就會模擬接下來可能發生的事。像是從疊高的積木中抽出中間一塊,模型會模擬積木倒下。
  • 換成慢慢抽出靠近頂端的一塊,積木則可能站住。AI 助理可以反覆嘗試、檢視模擬結果,再調整動作。研究團隊表示,測試中的 AI 助理也曾透過模擬互動發現看不見的世界特性。PAN 2 的模型權重、訓練程式碼和推論程式碼以 Apache 2.0 授權開放。

AI 伴隨問答

自備 Key 深度分析

針對這篇報導有疑問?可直接點選快捷問題或向 AI 追問背景脈絡與技術細節:

Linus ✦ Ekenstam 🔜 Adobe MAX Miami 2026
Linus ✦ Ekenstam 🔜 Adobe MAX Miami 2026@LinusEkenstam
·9 天前
𝕏

This is a MUCH bigger deal than people realize.

MBZUAI's Institute of Foundation Models just unveiled PAN 2, a world model for physical intelligence that predicts the next world: give it a scene and an action, and it simulates what happens next.

So you can ask it "what if?" Take a Jenga tower and pull a block from the middle, and it falls. What if you slowly pull one near the top instead? Run it again, and the tower stands.

But here's where the team really blew my mind: "what if?" again and again.

And you get an environment where an AI agent could learn: try an action, see the simulated result, adjust, try again.

In IFM's tests, an agent even discovered a property of the world it couldn't see, purely through simulated interaction. Try first in the simulated world, act in the real one after.

This is where world models should be heading: not just generating what the world looks like, but learning to predict how it behaves.

And it's open under Apache 2.0: weights, training code and inference code.

This is absolutely nuts.

推文來源 · X (Twitter)在 X 參與討論

讀者討論

理性交流 · 即時討論