Interactive Lab
Standalone tools to see, step by step, how a model processes a sentence, how a tokenizer can be manipulated and how to harden an agent. Everything runs in your browser; nothing is sent anywhere.
Model architectures
Modern LLM standard
The Transformer (self-attention)
Step through Query, Key and Value projections, attention weights, residual connections and token prediction. Every word looks directly at every other word, in parallel.
Open the tool →
Sequential recurrence
The LSTM (gated recurrent network)
Step through the forget, input and output gates that regulate the memory conveyor belt (cell state C) across time steps.
Open the tool →
Parallel local filter
The 1D CNN (sliding convolutions)
Slide a kernel across neighbouring tokens to extract local features, in full parallel.
Open the tool →