Latent Space · 4 min instead of 1 h 21 m · Oct 13, 2026
Robot data quality changes as the model improves
Jason Ma
Co-founder & chief scientist, Dyna Robotics
First author of Eureka, in which GPT-4 wrote robot reward functions that beat expert ones; previously at Google DeepMind and NVIDIA.
Shawn Wang (swyx)Host · Co-founder, AI Engineer
Wrote “The Rise of the AI Engineer” and founded Smol AI, maker of AI News.
Alessio FanelliHost · Founder, Kernel Labs
Partner at Decibel, backing technical founders who build developer tools.
About this segment
Ma explains why no single company can collect enough robot data and how Dyna pairs egocentric human video with small amounts of robot data. What counts as good data shifts as the model matures.
From “Robots That Predict the Future: Dyna-2 & Physical AI — Jason Ma, Dyna Robotics”, Oct 13, 2026. Watch the full episode
“I think data quality in robotics is a non-stationary problem in the sense that what is high-quality data for your model changes over time.”
Jason Maat 11:34 · Latent Space11:34Takeaways
- Any one company's in-house data is like building the internet from just Reddit
- Human video gives broad physical priors; robot data supplies low-level dexterity
- Early on any data helps; later you need targeted, task-specific data
In this episode
Full episode- 38:50Multi-site rollouts, not pilots, prove a robot creates value4:05
- 56:32Why predicting raw video beats VLAs, and where that bet could fail4:16
10:12Robot data quality changes as the model improves3:36