An essay on why scalable data ecosystems reshape research directions, why 3D feels structurally hard to scale, and why video currently looks like the more natural pre-training frontier.
Blog
Writing
Longer-form notes on research, scaling, video, 3D, and whatever else feels worth writing down.
A longer reflection on error accumulation in causal world models, why LLMs seem much less affected, and why rollout length, teacher quality, and base-model scale are the real bottlenecks.
A collection of views on why learnable sparse attention is still underrated for video, what context should really mean in world models, and why data and unified training may matter more than new memory algorithms.