Neural Trainer

A 2→8→8→1 network, hand-written backprop, zero libraries. Running right now in your browser.

Every framework hides the same small miracle: a pile of random numbers becomes a decision-maker through nothing but repeated nudges. This page removes the framework so you can watch it happen. The network below has two inputs (the x and y of each dot), two hidden layers of eight tanh neurons, and one sigmoid output that says "orange or purple". The colored background is the network's current opinion about every point on the plane. At step zero, that opinion is noise.

Press train. The loss curve falls, the boundary stops being a blur, and somewhere around the spiral dataset you'll catch the network genuinely struggling before it figures out that the world is curved. That struggle is the honest part. Crank the learning rate too high and watch the boundary thrash instead of settle, which is the same failure you'd debug in production, just at a scale where you can see it.

The same mechanics run every model I've shipped, just with more zeros in the parameter count. If you want the production-scale version of this story, start with One Encoder, Three Jobs →