Zhen

pm operator (scale ai, amazon, robinhood) turned vc (ince)

[Zhen @zhen_do_ob]

Boom goes the Dyna-mite 🧨

[Dyna Robotics @DynaRobotics]

Today we are introducing Dyna-2, a world-action model pre-trained on one million hours of human video. At this scale, for the first time, we discovered several new scaling laws:

• world-action models exhibit scaling law on human data across four orders of magnitude, from 1000 to 1,000,000 hours,
• this human data scaling law implied a scaling law on never seen robot data,
• both data and objective matter; world modeling and scaling on video data are essential for cross-embodiment scaling transfer to emerge

🧵

[Zhen @zhen_do_ob]

Simple is beautiful

[Stijn Noorman @stijnnoorman]

Dumb people make things complex.

Smart people keep things simple.

[Zhen @zhen_do_ob]

@pronounced_kyle I was a fixed income portfolio manager who ran $20bn. not all that it’s cracked up to be. venture is way more fun and aum is not a proxy for how much $ upside there is.

[Zhen @zhen_do_ob]

Being a VC is basically nonstop negotiations (so glad I took both negotiations and advanced negotiations at booth).

Single case negotiations can be I win you lose but often times are zero sum, and hurt the long term business relationship.

Most real life negotiations are multi-case and all parties need to come out feeling like they’ve won for things to be truly successful long term.

The way to do that is to give more and take less or come up with outside the box solutions based on belief gaps.

And above all, when push comes to shove, always make decisions based on the business relationship, not so much the numbers.

[Zhen @zhen_do_ob]

even the OG sim folks doing real world egocentric

[Lukas Ziegler @lukas_m_ziegler]

Training robot hands at scale! 🪬

@LightwheelAI released EgoSuite-Open100K, a 100,000-hour open dataset of egocentric video with hand and body pose annotations.

Everything is available on @huggingface.

Multiple variants for different needs.

→ Head-view only for hand-centric tasks.
→ Wrist-view for close-range contact and grasping detail.
→ Full-body pose variants for whole-body context.
→ Raw video variants without annotations for lightweight inspection.

15,000+ distinct tasks across 15,000+ different scenes.

Egocentric data is gold for imitation learning. Show a robot how humans actually manipulate things from the perspective humans have, from their eyes, from the wrist. Hand pose annotations transfer directly to robot hands. Wrist-camera view shows contact dynamics most robots need to learn.

Multiple data formats, LeRobot v3 and MCAP, so researchers can use whatever their pipeline expects. Semantic annotations at the subtask level, not just raw video.

Congrats @bgxc, and team! 🦾

Find it on HF: https://t.co/y0oZyyo514

~~

♻ Join the weekly robotics newsletter, and never miss any news → https://t.co/GoA3ZuwoPB

[Zhen @zhen_do_ob]

besides me there was only one other SPM (my direct report) who did both fine grained sensor fusion computer vision and gen AI. she’s now at OAI.

[Zhen @zhen_do_ob]

yups superbullish and that would’ve been my plan had I stayed at scale. the obvious next thing that was brewing two years ago. and now it’s finally happening.

[Anto Patrex @antopatrex1]

the best ML researchers in the world are leaving pure software AI for embodied AI. google deepmind, meta FAIR, openAI all expanded robotics teams this year. physical AI is where the frontier moved.

[Zhen @zhen_do_ob]

@danprimack damn - the goalposts have moved

[Zhen @zhen_do_ob]

@sam_brin your bro was 🐐 even in hs

[Zhen @zhen_do_ob]

ERHS mafia represent :)

I think I also did calc in 10th grade and did linear equations at UMCP over the summer.

[Jon Erlichman @JonErlichman]

Sergey Brin was born this day in 1973.

His high school called him a young Einstein.

[Zhen @zhen_do_ob]

@eddybuild @buildpbc wait you should just ask L and see if he can sublet some space to u lol

[Zhen @zhen_do_ob]

@eddybuild @buildpbc Bro go peninsula (RWC/San Mateo/San Carlos) - it’s where all the robotics companies are (Dyna, Generalist, Genesis etc)

[Zhen @zhen_do_ob]

@eddybuild @buildpbc no but ask around - Lindon when he’s more free from yolo’ing.

[Zhen @zhen_do_ob]

@eddybuild @buildpbc also Sunday

[Zhen @zhen_do_ob]

linear algebra*

[Zhen @zhen_do_ob]

The meta pattern that hasn’t changed over multiple cycles between US and China is actually a 🪃:

1) 🇺🇸 0->1 true creative innovation. Examples: YouTube, Pinterest, Uber, ChatGPT, Atlas/Tesla/Figure humanoids
2) 🇨🇳 1->100 stand on shoulders of US innovation and polish down on cost. Examples: ByteDance, Rednote, Meituan, Deepseek and other models, Unitree etc.
3) 🇺🇸 100->100000000 resell back to US to try to enter a more lucrative market. Examples: TikTok, DoorDash (learned from Meituan), Enterprise AI companies like Harvey fine tuning off of Kimi etc., Unitree robots

In short, US and China are largely interdependent when it comes to technology. AI is not really different except for the dual use nature and national security concerns. The workaround is obviously to segregate the uses and organizationally ringfence commercial vs public sector like what AWS does.

Startups don’t really get this and cry about things but more mature generational tech companies (Apple, Nvidia, Tesla, Microsoft, Amazon) and many others find a way to thread the needle. It’s at the heart of American capitalism.