The Physical World Data Flywheel: Hierarchical Vision AI System Design and Optimization
Processing continuous physical AI video streams is limited by edge compute constraints and long-horizon temporal recognition. We present a hierarchical cascade system combining a lightweight encoder for anomaly screening with a VLM for on-demand verification, reducing compute overhead while maintaining high recall and precision.
Read More
Modeling Physical AI: From VLA to World Model
Echoing the evolution of LLMs, a similar wave is happening in physical AI: VLAs equipped robots with internet-scale semantic understanding for the first time; World Models are now enabling robots to truly understand the world and model how actions change the physical environment.
Read More
Agentic RL (Part III): Architecture Analysis of Verl and SkyRL to Retool-RL Case Practice
Exploring how to build an efficient and stable training closed-loop for Agentic RL from the perspective of system and algorithm synergy. It analyzes the HybridFlow architecture and SkyRL's modular design, and introduces the practice of using process rewards and a posteriori teacher signals in Retool-RL.
Read More
From Data Processing to the Experience Flywheel: The Next Stage for LLM Data Engineering
A quick overview of the new book LLM Data: Principles, Technology and Practice, and exploring how to move from static data processing to building closed-loop agentic data flywheels.
Read More
Agentic RL (Part II): RL Systems for Real-World Tasks
Exploring how real-world open-ended tasks like financial forecasting and scientific discovery are transformed into trainable environments through environment engineering, verifiers, and data flywheels.
Read More
Agentic RL (Part I): A New Paradigm for Self-Evolving LLMs
A deep dive into Agentic Models, exploring the necessity of Reinforcement Learning, the evolution of reward engineering, and the progression of RL algorithms for self-evolving systems.
Read More