🤝 Open to Collab

Dipankar Sarkar PRO

dipankarsarkar

7 1957 383

https://www.dipankar.cc

AI & ML interests

Building the AI-native stack. Agents as infrastructure, safety as architecture, performance as plumbing. I publish the receipts: papers, datasets, demos.

Recent Activity

reacted to kumarkjha's post with 🔥 about 7 hours ago

I measured four interactive video world models against their published frame-rate claims — on the hardware each claim names, wherever I could match it. Every number is bound to a digest ledger anyone can re-verify. Matrix-Game 2.0: claims 25 fps (1× H100) → 19.7 fps on the same H100. Reproduces. ABot-World-0: claims 16 fps (RTX 5090) → 5.85 fps on the same 5090 (optional fp8/sage off). Oasis-500M: claims 20 fps (their engine) → 1.06 fps (the open 500M weights, vanilla). HY-World 1.5: claims 24 fps (streaming) → 1.36 fps (released pipeline runs batch, 8× A100). None of these contradict a vendor's claim — the finding is that most headline numbers can't be reproduced from the released artifacts, and the harness shows exactly which condition differs. Try it in 60 seconds on CPU: pip install "liveframe[world]" then liveframe world run synthetic --suite smoke Full methodology + calibration + limitations: https://huggingface.co/blog/kumarkjha/what-interactive-video-world-models-actually-deliv Code + evidence bundles: https://github.com/kkjcodes/liveframe I haven't built a world model — the teams above know far more than I do. I built the measuring stick. If you find a wrong number or a bad assumption, I'd genuinely love to hear it.

liked a dataset about 8 hours ago

karanxa/agent-action-safety-dataset

liked a dataset about 8 hours ago

Agentic-Payment/agent-overspend-benchmark-2026-10

View all activity

Organizations

dipankarsarkar 's datasets 3

Dipankar Sarkar PRO

AI & ML interests

Recent Activity

Organizations

dipankarsarkar 's datasets 3 Sort: Recently updated

dipankarsarkar 's datasets 3