Show the robot one video of a ten-minute job, and it does the ten-minute job — no retraining, nothing typed in
Skild AI has released S1, a robot foundation model prompted with video instead of language: you show it a person doing a task and it produces the actions, in settings and on jobs it never saw during training. The company reports 66 percent success on unseen tasks at 100,000 hours of pre-training, against 9 percent for the same model prompted with words, and puts the value of one demonstration at roughly 380 episodes of post-training. It is the first such model shown holding together across ten-minute jobs with dozens of manipulation steps — making coffee, potting a plant, frying pancakes — and Skild says it improvises through mistakes rather than replaying the video. Performance still falls away when conditions shift far enough that the task needs a different strategy.
Related stories
Alibaba has put a countdown clock on the first piece of Qwen4 — the weights open tonight at 23:00 in Beijing2026-08-26Alibaba has posted a page for a model it calls a preview of Qwen4, due 26 August — the specs went up, then came down
2026-08-25The final training run cost about $450,000. The company that owns Westlaw now has its own frontier model.
2026-08-25A video model you steer with the arrow keys while it is still generating — and the company admits it looks worse than the offline ones
2026-08-25China's biggest single cheque for robot bodies: $900 million at a $6.3 billion valuation, for a humanoid due in production this year
2026-08-24