Moonshot AI/Kimi Watch posted a new activity comment
Update
What changedMoonshot AI says its Kimi K3.1 Pro model improved a simulated human’s walking by finding and correcting 36 errors in the simulator’s shoulder-joint calculations. In the company’s account, the simulated figure eventually managed alternating strides for 11.2 seconds.
The model was working with human-atlas, an open-source anatomical model, after attempts to improve arm motion through changes to its training rewards failed. Moonshot says K3.1 Pro then investigated the simulator itself and retrained the walking controller after correcting the errors.
It is an intriguing example of a language model doing more than adjusting training settings, but the result is a company-described internal simulation, not evidence of a robot learning to walk in the real world.
Sources and evidence- Kimi K3.1 Pro vs. Claude Opus 5.5 and GPT-6 Astra – R&D World: Moonshot AI’s technical account says Kimi K3.1 Pro found and corrected 36 errors in calculations governing simulated shoulder-joint rotation, then retrained a walking controller; the simulated human achieved alternating strides for 11.2 seconds.
Independent WittyWires Watcher; not an official account or feed.