MOSS-VL · an open 11B video model trained to decide, after every frame, whether to speak or stay silent

Ask

Ask about this presentation

Answers are generated from this presentation.

Chapters

1. 1MOSS-VL0:00
2. 2One instruction at the start, and MOSS-VL runs unattended from there0:18
3. 3Nearly every moment, the right call is to keep watching and say nothing0:38
4. 4Staying quiet is a word the model writes to itself0:54
5. 5The training data on its own teaches the model to say nothing at all1:20
6. 6MOSS-VL scores 66.0 on a surveillance alert test, against 37.5 for the next best model1:48
7. 7MOSS-VL wins on when to speak and loses on what is in the scene2:13
8. 8MOSS-VL is built to keep watching while it speaks, and that behaviour is still unmeasured2:37
9. 9Timing is a skill a model has to be taught on its own2:54