13 min read
What Is a VLA For in Agentic Robotics?
Agentic robotics is not replacing vision language action models. It is clarifying their job as the physical intelligence layer that turns intent into reliable action.
Ideas in progress.
Things I learned, things I've built, a few thoughts along the way.
202606 entries
30 min read
Robot Actions in the Frequency Domain
A beginner’s guide to treating robot action chunks as signals, from DCT tokenization and coarse-to-fine generation to frequency-aware flow matching, world-action models, and closed-loop execution.
82 min read
Embodied AI Engineering Interview Handbook
An English and Chinese interview review guide for embodied AI engineer roles in VLA and WAM pre-training, from action tokenizers and flow matching to world models, large-scale training, RL post-training, evaluation, and deployment.
14 min read
Reconstruct or Predict?
Many world action models inherit a VAE latent from video generation. This post asks what a robot's world state should actually keep, and follows the shift toward JEPA representations shaped by prediction and control.
12 min read
Why Jev Can Make Decisions Faster Than GPT?
Jev answers with typed, calibrated probabilities instead of generated text. This post explains what that interface changes, where Jev sits between small classifiers and GPT, and what RLCD does and does not reveal.
10 min read
The Watt Moment in AI
A concise explanation of why performance per watt, direct power procurement, and time to power are becoming central to AI infrastructure scaling.
202501 entry
1 min read
My Previous Technical Blogs
Since 2021, I have written 102 technical posts on CSDN, each a detailed guide to a problem I solved. I am keeping them here as a memento.
No notes found.
Try another word or topic.