Chinese Startup Livestreams Two Weeks of Physical AI Training

Shengshu Technology’s world-model team opened a public livestream of its Motubrain 2.0 training process on October 21, and plans to keep it running for half a month. The broadcast follows a similar move by Xiaomi’s MiMo team, whose lead disclosed progress from the reinforcement-learning stage of its MiMo-V2.6 model, giving outsiders a look at the gritty details of large-model training.

Physical AI is the branch of the field concerned with models that learn from motion, sensor data and the physical world rather than text alone. Training such systems is usually closed-door work, so watching a full run unfold in public is unusual for a Chinese lab.

Shengshu Technology is the company behind Motubrain, a world-model effort aimed at teaching machines to predict and simulate physical environments. The livestream turns a normally private engineering process into something closer to a spectator sport, with all the uncertainty that implies.

The significance is less about any single benchmark than about transparency as a competitive tool. As Chinese AI firms race on embodied systems, showing the training process in public may become a way to attract talent, partners and credibility, not just attention.