Activity Feed

AI & ML interests

None defined yet.

Recent Activity

Smooke 
posted an update 12 days ago
lorraine2 
posted an update 3 months ago
view post
Post
1,769
🧠 New NVIDIA paper: WorldTrace 🧠

Long-horizon video world models can store prior scenes in their KV cache and still fail to retrieve them once RoPE offsets leave the training range. We show this is an addressability problem, not just a compression problem.

WorldTrace assigns compressed memory slots fixed, in-distribution positions. WT-Field improves temporal consistency by +15.5%, while WT-Landmark improves episodic recall by +19.5% on LoopBench, without retraining.

🏆 Best Paper + Oral, ICML 2026 From Frames to Stories Workshop
🔎 Project page: https://research.nvidia.com/labs/sil/projects/WorldTrace/
📄 Paper: Addressable Memory for Video World Models (2608.07408)

Led by Xindi Wu, with Sven Elflein, James Lucas, Olga Russakovsky, Laura Leal-Taixé, Despoina Paschalidou, Aljoša Ošep, and me.
Smooke 
posted an update 3 months ago
view post
Post
125
What the Cloud Actually Drinks from "AI’s Water Problem Is Smaller Than You Think" on @hackernoon https://hackernoon.com/ais-water-problem-is-smaller-than-you-think

For years, the environmental conversation about computing was about carbon. Water was an afterthought, buried in sustainability reports under "operational metrics." That was a mistake, because the physics were never subtle.

Here's how a conventional data center stays alive. Heat from the chips is carried out of the server hall by a closed loop of coolant, which dumps it into a second, open loop — the cooling tower — where water is sprayed into moving air and deliberately allowed to evaporate. It's a spectacularly efficient way to move heat, the same trick your body uses when it sweats. It is also, by definition, consumptive: that water doesn't go back to the river. It goes into the sky. Depending on climate and tuning, this costs between 1 and 9 liters of water per kilowatt-hour of server energy.