GEN-1.5 can in-context learn a new task from as little as 3 to 12 seconds of demonstration data — or adapt with 1 to 10 gradient steps on minutes of data. These capabilities emerge from pretraining on large-scale physical experience.

The bitter lesson, physical edition?

  • unpossum@sh.itjust.worksOPEnglish
    4·
    11 days ago

    This video of the robot getting a lego brick stuck on a finger, then removing it, stuck out to me as a very natural-looking motion.

    ETA: figured out how to embed!