Model zoo
Models
A ladder of networks, each one introducing the problem the next exists to solve. Every one is trained from scratch and runs in your browser; each page shows the run behind its numbers and the command that reproduces it.
- 01
Multilayer perceptron
MLP in NumPy
784 → 128 → 64 → 10 with backpropagation written by hand and checked against finite differences. The network in the hero.
98.69%test accuracy
MNIST · 109,386 parameters · NumPy only
Runs in your browserOct 2026
- 02
Convolutional network
Small CNN
Five convolutions and no dense layer: 14x fewer parameters than the MLP and fewer mistakes. Shows its feature maps live.
99.41%test accuracy
MNIST · 7,578 parameters
Runs in your browserOct 2026
- 03
Convolutional network
CNN on CIFAR-10
Colour photos in ten classes. Pick a test image or upload your own, and see which part of it decided the answer.
89.69%test accuracy
CIFAR-10 · 288,746 parameters
Runs in your browserOct 2026
- 04
Residual network
ResNet on CIFAR-10
Plain and residual nets at 20 and 56 layers, same recipe: why skip connections make depth trainable, measured.
+9.8 ptsfrom skip connections at 56 layers
CIFAR-10 · 91.70% vs 81.90%
Runs in your browserOct 2026
- 05
Generative
Variational autoencoder
Every digit squeezed into two numbers. Drag through that space and watch one digit turn into another.
-137.0nats, test ELBO
MNIST · 1,068,820 parameters · 2-D latent
Runs in your browserOct 2026
- 06
Generative
DCGAN
A generator and a discriminator trained against each other. New digits on every click, judged by the CNN above.
3.30 bitsspread over 10 digits, of 3.32
MNIST · 546,177 parameters
Runs in your browserOct 2026
- 07
Recurrent
LSTM forecaster
Delhi's electricity demand an hour ahead from the last 24 hours, weather and the sun's position. Pick any day of 2024.
1.06%error, an hour ahead
Delhi 2024 · 204,300 parameters
Runs in your browserOct 2026
- 08
Transformer
Mini-GPT
A character-level transformer that writes in your browser. Click any character to see what each attention head looked at.
2.11bits per character
Shakespeare · 818,048 parameters
Runs in your browserOct 2026
- 09
Vision transformer
ViT-Tiny
Attention instead of convolution, trained from scratch on the same data as the ResNets, and where it falls short.
82.68%test accuracy
CIFAR-10 · 1,806,538 parameters
Runs in your browserOct 2026