Lab
Play with it, then see why.
- The task
- An agent sees 27 numbers (distances to walls and body, the apple, free space) and picks one of three moves: turn left, go straight, turn right.
- How it learned
- Two ways, same inputs: a small network evolved by selection and mutation (genetic algorithm), and a reinforcement-learning agent trained by trial and reward (PPO). A third PPO model trained for longer.
- What to look at
- The bars are the model's raw scores for each move. Inside the network shows which inputs drive the choice. Click a cell to move the apple.
- The task
- Read a digit you draw. The network gets the prepared image, cropped, scaled and centred like the digits it learned from.
- How it learned
- A small network (784, 64, 64, 10) written in numpy with backpropagation coded by hand, trained on the MNIST handwritten digits. It runs entirely in your browser.
- What to look at
- The ten bars are its live probabilities. It always picks a digit, even for a scribble: confidence is not the same as being right.
Track the satellites
Satellite tracking temporarily unavailable.
Each dot is a real satellite, positioned from its public orbital elements for the current time.
Drag to rotate the globe, tap a dot to select it. The thin line is the selected satellite's path over the next 100 minutes.
- The data
- Public orbital elements of real satellites, fetched live from CelesTrak.
- How it works
- Every position is computed in your browser from those elements, and drawn in 3D.
- What to look at
- Drag to rotate, tap a dot to select a satellite and see its path for the next 100 minutes.
- The data
- Public highway cameras of the California transport department (Caltrans), streamed live.
- How it works
- A small YOLO detector runs in your browser on each image; a tracker follows the vehicles and counts them. No server.
- What to look at
- Switch camera, move the confidence slider, hide parked vehicles. The curve is the traffic over the last three minutes.