- Camera screens
- One short question to Cosmos 3 Nano about one camera frame: does it show a possible incident? Routine screens take the cameras in turn; extra screens of cameras that something drew attention to are not counted here. Brief events between screens can be missed. It is not a count of incidents.
- Median time since a camera’s last screen
- Half of the live cameras with a screen on record were last screened at most this long ago. A screen stays on record for 14 days; live cameras without one are left out and counted separately.
- Routine scene reviews
- Fuller descriptions of a camera’s scene, made in rotation, in the last complete hour.
- Possible incidents
- Flagged: a screen, a detector hint or a review proposed one. Confirmed: Cosmos 3 Nano and an external AI model both agreed. Neither is checked by a person. Counting starts again after a gap in the server’s records, so “since” can be later than seven days ago.
- Motion puzzles
- Questions about real footage from our cameras that we reversed, shuffled, sampled more sparsely or paired with a still. Because we made the change, the right answer is known exactly. They test whether motion and time reach Cosmos 3 Nano, not whether it recognizes events. They are not incident accuracy.
- Chance
- What random guessing among the allowed answers scores on average: 50% for arrow of time (2 answers), 16.7% for order (6 orders), 33.3% for span (3 lengths) and 50% for sync (2 answers). These come from the training code. A set of 100 clips can lean toward one answer; when we know that lean, we also show what always giving the most common answer would score.
- 95% range
- The spread expected from the clips scored alone. When it reaches down to chance, or to what always giving the most common answer would score, the result cannot be told apart from guessing. It does not adjust for looking at every round.
- Validation and test cameras
- Validation cameras never train the model and are used to follow its progress. Test cameras, a separate set, are read only by the go-live checks. Both are held-out camera views, not new places: some look at streets that training cameras also see.
- Loss against machine labels
- Answer-token loss: how unlikely the model found the label’s words, averaged per word piece. Lower means its answers are closer to labels written by an external AI model (from OpenAI). It is not accuracy, and machine labels can be wrong.
- Examples per training hour
- Examples divided by the time spent in training steps since the trainer last started. Time spent testing and saving is not counted, so it runs faster than examples per clock hour.
- Examples today
- Counted by the trainer when each round ends or it stops, since midnight Pacific time.
- From the last 24 hours
- Share of a round’s examples cut from video captured in the past day. A round does not start below 25%.
- Rounds and stages
- A round is 250 steps of 32 examples, then a test on validation cameras. Stage 1a trains only the bridge that turns V-JEPA 2.1’s reading of an 8-second clip into 96 tokens for Cosmos 3 Nano; Stage 1b also trains a small adapter inside Cosmos. In both, V-JEPA 2.1’s own network stays frozen: what learns is the bridge and the adapter. A qualification stage then trains two copies side by side, one that also tunes the top blocks of V-JEPA 2.1’s own network and one that keeps it frozen, to check whether tuning helps; its rounds skip the validation test. Stage 2 continues with the copy qualification picked, and its rounds can become candidates for live use.
- Why not hours of video
- Each V-JEPA example is one 8-second clip (up to 32 seconds for span), clips are reused across examples, and old clips are removed over time, so hours of video would overstate what the model has seen. Cosmos fine-tuning learns from labelled frames, not video.
- Live use
- Automatic promotion moves a V-JEPA version into scene reviews only after 13 automatic checks on test cameras and a canary of up to 3 hours on GPU 0, and its first 24 hours in use are a probation that rolls it back on any failure. An operator can also put a version in use by hand, which the checks do not gate, or roll one back; the page says so when that is how the version in use got there. Until a version is in use, live reviews use the released Cosmos 3 Nano. Cosmos fine-tuned runs are kept, but nothing loads them live or promotes them automatically.
- What we do not publish
- No images, video, camera IDs, locations or tracks; no list of held-out cameras; no names or addresses; only totals. At the owner's direction, signed-in live video and alert clips preserve faces. Models on our machine analyze unblurred video: the people-and-vehicle detector does, and the motion model trains partly on unblurred clips from saved recordings. Analyzed stills and images sent to outside AI models remain face-blurred, and blur can miss faces.