Yash Thube
I care about making machines intuitively understand space, time & physics through world models. Moving beyond next-token prediction, mental simulation engines that learn from mistakes, plan with imagination, and adapt to our physical world.Works shaping my thinking: LeCun’s JEPA/V-JEPA , NVIDIA Cosmos , Sakana AI’s work in open-endedness, DeepMind’s Dreamer series & Genie , World Labs, Physical Intelligence (π), and Genesis.
Currently exploring how latent dynamics can improve VLMs/VLAMs spatial, casual understanding and predictive planning. Always open to collaborations!