demos

Drag a slider, pick an unlearning method, watch the forgotten knowledge come back under a fine-tuning attack. Every published method recovers within ~100 steps. Except the wrapper-axis ones.

• llm
• unlearning
• relearn-attack
• tofu / wmdp

Store binary patterns with Hebbian learning, corrupt them, then watch the network roll downhill into the nearest attractor. Explore the energy landscape, find spurious minima, and hit the 0.138 capacity wall.

• associative memory
• hebbian
• attractors
• pca

Train a Restricted Boltzmann Machine on tiny digits live in your browser. Watch filters emerge, paint corruption on a digit and watch Gibbs sampling clean it up, then explore the free-energy landscape in 3D.

• rbm
• contrastive divergence
• gibbs sampling
• energy-based

Kohonen SOM on a hex grid. Random RGB samples flow in, neurons compete for them, and a chaotic checkerboard self-organizes into smooth color regions. Toggle the U-Matrix to see cluster boundaries, then watch the grid morph into a 2D manifold inside the RGB cube.

• kohonen
• topology
• unsupervised
• manifold

Interactive paper on mean-field formation games. Read the math, then run agents on a 30×30 grid under any of six cost variants from §6 and watch them self-organise into a target pattern from random initial positions.

• mean-field
• multi-agent
• potential games
• simulation