San Francisco’s streets, measured live.
Street cameras counting · 15-min averagesOrchestra’s AI counts the people and vehicles in view on San Francisco street cameras and reviews sampled scenes with Cosmos 3. The map shows live aggregates by neighborhood.
No imagery, tracks or camera positions are published here. How camera video is handled
Loading neighborhood numbers…
- People in view
- —per camera
- Loading…
- Vehicles in view
- —per camera
- Loading…
- Cameras counting
- —
- Loading…
- Cameras AI-reviewed
- —in 5 min
- Loading…
15-minute averages per counting camera.
Color and label: live count per cameraDot size: cameras countingHatched: 1–2 cameras, values withheld
How to read this map
Positions are neighborhood centers, not cameras. Labels show 15-minute averages per counting camera; dot size shows how many cameras are counting. Hatched areas have 1–2 counting cameras, so their numbers are withheld.
Outlines: DataSF Analysis Neighborhoods (public domain).
Select a neighborhood outline or card to open it in the live city.
What’s new
All updatesAI
Two world models, live and in training
Cosmos 3 Nano · live on GPU 0
Screens each live camera for possible incidents.
—routine screens in the last full hour
Loading…
V-JEPA 2.1 · training on GPU 2
A link we are training so Cosmos 3 Nano can read motion.
—training examples today
Loading…
AI review
What the AI is doing
Cosmos 3 Nano reviews sampled scenes from counting cameras; possible incidents get a second review by an external model. Sampled, not continuous: brief events can be missed.
Current alerts
Checking…
Review activity
- Cameras reviewed in the last 5 min
- —
- Median time since a frame was sent for review
- —
- Reviews completed in the last hour
- —
Loading…
Possible incidents, last 7 days
Model-written insights
Neighborhoods · 15-minute averages
Neighborhoods
Live 15-minute averages are shown where at least 3 cameras are counting. Lines show the last 24 hours.
- Loading neighborhoods…
How it works
How the live map works
Observe
Street cameras across San Francisco stream to Orchestra’s servers, where object detection counts the people and vehicles in view, minute by minute.
Review
Cosmos 3 Nano, a vision-language model, reviews sampled face-blurred scenes. A possible incident becomes an alert only when an external model’s second review agrees; alerts are not verified by a person.
Original camera recordings are kept privately on our servers for counting and model training. Models on our servers analyze unblurred video, and frames kept privately for them may be unblurred. At the owner's direction, signed-in live video and alert clips preserve faces. Analyzed still images and images sent to outside AI models remain face-blurred, though blur can miss faces. Audio is never decoded. Face-blurred frames and crops go to OpenAI for demographic estimates, incident second reviews and training labels, and at most five per hourly analysis for city insights; alert clips are emailed privately to the owner and two team members. No imagery, tracks or camera positions are published here.