Yipeng Wang

MTS @theworldlabs. prev. @pika_labs @meta @RealityLabs
Video Diffusion & World Models | Seeking a way to connect all the dots

[Yipeng Wang @chill_chio]

I can see a lot of unlocks.

[gokaygokay @gokayfem]

World to World

I gave Claude Opus 5.5 a Gaussian splat of a garden and asked it to rebuild the scene in Blender

Procedural objects, Geometry Nodes, Shader Nodes, then captured textures mapped onto selected surfaces

[Yipeng Wang @chill_chio]

&& it's going to be easier to RL next-token prediction tasks than diffusion. I think we will see some paradigm shift

[zeji yi @ZejiYi]

Common belief: Robot policies need continuous action heads; discrete actions are 𝐥𝐞𝐬𝐬 𝐧𝐚𝐭𝐢𝐯𝐞 𝐚𝐧𝐝 𝐥𝐞𝐬𝐬 𝐩𝐫𝐞𝐜𝐢𝐬𝐞.

❗We find policies w/ the simplest discrete representations 𝐦𝐚𝐭𝐜𝐡 𝐨𝐫 𝐛𝐞𝐚𝐭 continuous ones—in sim & real, in single-step & diffusion/flow settings.

Here's why 🧵

[Yipeng Wang @chill_chio]

Life is a symphony of models operating at different timescales

[OFF SCRIPT @offsrambledegg]

Speed is life

[Yipeng Wang @chill_chio]

very cute launch video! reminds me of pika old days

[Dream Labs @dreamlabs_ai]

Today, Dream Labs is coming out of stealth with Dream-1, a robot foundation model that controls bimanual robots it has never seen, on tasks it was never shown, in places it has never been, from $3/hour.

Dream API is live for everyone, today. 🧵

[Yipeng Wang @chill_chio]

not all history repeats but i think this one will 100%

[a16z @a16z]

Every AI app gets asked whether the labs will crush it

The mobile era offers one answer: chips led first, then devices, and over time software and services built on top pulled ahead

[Yipeng Wang @chill_chio]

@wenhaocha1 Im mostly bounded by my attention span tbh

[Yipeng Wang @chill_chio]

it's a good model sir :D

[Ziwei Liu @liuziwei7]

🎉 Native, encoder-free VLMs are an important step toward more unified multimodal intelligence.

We’re also pioneering this paradigm with our NEO & SenseNova-U series:

- NEO: https://t.co/qooHtOdaM6
- SN-U: https://t.co/U2uOLI1ISL

Excited to see the field move forward together!

[Yipeng Wang @chill_chio]

If really climbing from the bottom... it's quite more compute heavy than pre-training

[SemiAnalysis @SemiAnalysis_]

GLM 5.3’s biggest leap came from post-training alone.

“GLM 5.3 was built entirely by post-training the previous model. For me, that was a wake-up call.”

“Post-training is far more GPU-efficient than pre-training from scratch. Plenty of companies can’t afford to train a model from scratch, but they have proprietary data. They can post-train themselves and beat some of the closed models.”

“There’s still so much juice left to squeeze from post-training. We compared GLM 5.3’s cyber capabilities with 5.2, and it’s a night and day difference. That’s all just post-training.”

[Yipeng Wang @chill_chio]

why

[amit @amitisinvesting]

$TSLA

We just witnessed history folks.

[Yipeng Wang @chill_chio]

How much tactile data do u pretrain on？is the demo mostly overfitting？

[Gordon Wetzstein @GordonWetzstein]

A secure grasp needs continuous feedback on contact forces. At Rhoda, we’re giving robots tactile sensing and teaching our models to predict what they’ll feel next—a step toward more efficient, reliable manipulation, even when vision is available.

[Yipeng Wang @chill_chio]

as close-loop as possible, as long-context as possible. and making fast decisions, instead of pure imitations. maybe these are the keys.

[Gordon Wetzstein @GordonWetzstein]

Diffusion models spend the same compute on a blank wall as on a face. But you often know in advance where the detail will be.

Introducing Level-of-Token (LoT) Diffusion: we turn that knowledge into a multiresolution token layout, with fine tokens where detail is needed and coarse tokens elsewhere.

1/9🧵

[Wenhao Chai @wenhaocha1]

New research (165 pages!)

We thought there were too many benchmarks for video understanding, so we squeezed the benchmarks dry and got video-index

(cover inspiration: gradient canopy 😜

[Cong Wei @CongWei1230]

Let's remove VAEs and ViTs from video models!

🚀 Introducing 𝗣𝗶𝘅𝗲𝗹𝗨𝗠𝗠: an encoder-free unified multimodal model for image and video understanding and generation, directly in pixel space.

Code and model available today!

🌐 https://t.co/RXUObhsSIZ
📄 https://t.co/lA7VaJoVws

[Yipeng Wang @chill_chio]

Now, with these tools in hand, and a posttaining recipe, can we get to the NeRF level of 3D reconstruction quality? Reminded that NeRF used to also take 10hrs to run for a single scene optimization, even longer than a long-horizon LLM run.

[Siyuan Bian @siyuan_bian]

GPT-6 Astra can build 3D scenes from video. But how can we make them more accurate, physically plausible, and richly interactive?

We tackle this by integrating 3D vision tools (Pi3X, SAM3, GVHMR/Kimodo, and Blender sim) into Astra’s agentic workflow.

👇
https://t.co/veb0aqcpqR

[Yipeng Wang @chill_chio]

@cuijiaxingfb Feels kinda like it in the recent jump of 3D understanding capabilities of VLM, being primarily driven by posttraining on 3D/CAD tasks

[Yipeng Wang @chill_chio]

Question:
To the researchers training coding/cua agents to draw images, make videos, and create 3D objects/scenes, are we now seeing the kind of "generation helps understanding" story coming true, of which the multimodal researchers has been pursuing for years?

[Yipeng Wang @chill_chio]

Definitely one of the most insightful paper that I've seen in a while. Congrats to @HaoZhang623 and @gengshanY !

[Hao Zhang @HaoZhang623]

World Tracing is a NeurIPS 2026 Spotlight! 🎉 It’s also our first NeurIPS paper at World Labs. Grateful to everyone who worked on it with me. @jcjohnss @BenMildenhall @_mbanani @KeunhongP @AndyCheng_JH @pzpzpzp1 @Hawaii271828 @chlassner @gengshanY

And there’s more to come: World Tracing Pro, with code and enhanced checkpoints, is on the way 👀

https://t.co/s4Ek8HvSxf
https://t.co/bkYxCO0y5Y

[Microsoft AI @MicrosoftAI]

MAI is hill-climbing 61% faster than any other lab in image generation. We've gained +283 Elo on text-to-image in 10 months to optimise the pareto frontier.