重點導讀
PAN 2 預測動作後連續且一致的未來世界狀態,而非只預測下一個影格。
- 示範中,衝浪者必須在第 6 秒出現,提早或延後都會失敗。
- 原文指出,這類模擬可供AI 助理規劃並評估「如果……會怎樣」的結果,避免昂貴的真實世界試驗。
AI 伴隨問答
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@LinusEkenstam
·12 天前
AI thats grounded in TIME.
This surfer needs to pop at the 6 second mark, any sooner or later, and he will crash.
Watch the demo and see the versions where the surfer fails.
why is this important?
PAN 2 predicts consistent future world states after actions rather than isolated next frames, which matters because agents need reliable multi-step simulations to plan and evaluate "what if" outcomes without costly real-world trials.
推文來源 · X (Twitter)在 X 參與討論
來源: Linus ✦ Ekenstam 🔜 Adobe MAX Miami 2026 (x.com)
在 X 查看推文