Positive activation Negative activation Active token (highlighted)
⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.

Transformer Feed-Forward Block: Per-Token Expand & Project

Attention gets most of the attention, but roughly two-thirds of a transformer's parameters live in its other sublayer: the position-wise feed-forward network. This simulator renders every token's vector as a column of bars and runs it through the real two-layer MLP — expand to dff, apply GELU or ReLU, project back to dmodel, add the residual — independently for each position, with live parameter counts and per-token sparsity readouts so you can see exactly why this sublayer, not attention, holds most of the weights in models like GPT and BERT.