20 August 2026
Robot model learns new tasks from brief video demonstrations
First reported
The Decoder and The Neuron ran this on , all on the same day.
- Generalist AI released GEN-1.5, which watches a 3-12 second video of a task and attempts it immediately without any additional training.
- In initial tests across ten tasks like opening jars, the robot succeeded 59 percent of the time. After ten training steps on five minutes of video, success rate rose to 83 percent.
- The model developed this ability on its own during eight months of pretraining on interaction data, without being explicitly programmed to learn from demonstrations.
How it was covered
The NeuronPete Huang & Grant Harvey
Generalist released GEN-1.5, a robot model that can watch a single 3-12 second physical demonstration and immediately attempt the new task using in-context learning. The newsletter frames this as validating Rich Sutton's thesis that AI's next leap comes from agents learning continuously from the world rather than static training data.
Reported by The Decoder