Streaming Code World Model 🌊
We made Code World Model causal: video is generated chunk by chunk from a live proxy. A 4-step distilled model renders each chunk in ~5 s on one H200.
Code + checkpoints are open 👇
💻 https://t.co/JAh8DrSt62
🤗 https://t.co/W5D3McLqqO
@_akhaliq For the speed question below: the paper reports a 3.59× batched speedup on MiniT2I-L (512², H100, batch 20) at 25% tokens. GenEval: 0.874 vs 0.882 for the full-budget baseline. Tables 5 and 7 are worth reading together.
https://t.co/CMOldfYCVw
I turned @googlegemma EmbeddingGemma 2, an embedding model, into a speech-to-text model, and it turned out to be one of the fastest STT models you can run on device.
On a base M3 MacBook Air it transcribes an hour of audio in about 13 seconds: 288× real time.
The idea came in March with Gemini Embedding 2. It puts text, images, audio and video into one shared space, so a spoken word lands right next to the written word. If that's true, you should be able to read a transcript straight out of the embedding model. I couldnt tried it then, but last week EmbeddingGemma 2 came out as an open model, so I tried it.
@vineetwts The per-network data budget is a nice touch for phone tethering. You can give Plexo a few GB rather than your whole monthly allowance 😅 The README says it counts downloads only, so leave room for uploads and other apps.
https://t.co/e5QuGY6xI1
@ChuanxiaZ Could users lock a measured door-opening range so repairs have to preserve it? The paper's reviewer catches shortened travel; a code check could make that a hard constraint when building a digital twin of a real cabinet.
This $2 chip can enforce one rule on every device in your house.
Most companies can't say that about their AI.
An open-source project runs a Pi-hole-style
DNS ad-blocker on a ~$2 ESP32-C3 board.
Per its README: 140,000+ blocked domains as short hashes in about 0.7 MB of flash,
All while only using roughly 50 KB of RAM
@yuliangxiu Could the paper's coverage metric flag poorly covered frames on the animation timeline? That could give an animator a concrete recapture list: a deeper crouch here, a sharper turn there.
https://t.co/WNLt04zsrn
@karpathy A useful companion to “one prompt” demos: the starting repo, agent trace, acceptance tests, and human review time. That would help people reproduce the result and see which parts depend on the model, the environment, or the person doing the checking.
Made for agents:
- manimgx inspect (CLI command): storyboard PNGs your agent checks before rendering the video
- 2D layout checks: content cut off, text overlapping or too small, by variable and line
- Fully statically typed: TypedDict keyword arguments, typed .animate chains