Posts

Reading notes on dexterous RL: reset distributions, goal-conditioned pretraining, structured exploration, motion tracking, contact priors and simulation.

Reading notes on Robo-centric VLMs: data construction, spatial grounding, multi-task training, and closed-loop robot execution.

Reading notes on Mega-Scale RL: batches, multiple policies, reset distributions and off-policy methods.

Reading notes on ELT, Looped-DiT and LiFT, with comparisons to hidden states across steps and DEQ.

My reading notes on diffusion / flow post-training, comparing the approaches used in image/video generation and robotics.

When the future can be watched in advance, what is left of living? A short science-fiction story about world models, hindsight, and the unknown.

A reflection on how my personal AI usage has evolved from pasting error messages to full-blown life logging, and why I believe this trajectory — giving agents ever more context about your life — is how the future will unfold.

A summary of my personal opinions on world models in 2025, covering their current state, future prospects, and implications for embodied AI. Discusses 3D modeling approaches, data challenges, research directions, and the role of JEPA-style architectures in the evolution of world models.

Lecture notes and Cheat Sheet for Course: Algebra and Computation, which is a course offered to Freshman students at Yao Class.

This article introduces Titans, a novel architecture that as a meta in-context learner, learns to memorize at test time. Through designing a long-term memory module, and proposing three variants of Titans (MAC, MAG, MAL), the model achieves superior performance compared to Transformers and other baselines, especially in long-context tasks.