You may have to wait a bit for some to load up from hibernation.
Computer Vision · CNN
Iris
A convolutional neural network trained on MNIST. Draw any digit and Iris will identify it with ~1.68M parameters.
Reinforcement Learning · SAC
Myrmidon V3
A Soft Actor–Critic agent trained from scratch on Ant-v5 — eight joints, twin critics, a self-tuning temperature. Scores ~39% above the published SAC baseline.
Reinforcement Learning · PPO
Hermes 2
A PPO actor–critic trained from scratch on HalfCheetah-v5 — six continuous torques, no early termination. Watch a full 1000-step episode with live telemetry.
Reinforcement Learning · REINFORCE vs PPO
Sisyphus & Atlas
Two RL agents trained on CartPole-v1 — Sisyphus using REINFORCE and Atlas using PPO. Watch them balance side by side and compare their performance.
Language Model · Transformer
Romeo
A GPT-style transformer language model trained on TinyStories. Enter a seed and Romeo will continue the story.
Language Model · LSTM
Mercutio
An LSTM character-level language model trained on Shakespeare. Generates text in the style of the Bard.
Language Model · RNN
Benvolio
A vanilla RNN character-level language model trained on Shakespeare. A simpler cousin to Mercutio.
Classification · NLP
Spamton
A 3-layer neural network built from scratch using NumPy, trained on 5,500 emails to detect spam with ~386K parameters.