Transformers, from the inside · Chapter 3
The MLP
By Tomáš Sikora · July 3, 2026
An interactive tour of the other two thirds of a transformer: the SwiGLU block, RMSNorm and pre- vs post-norm, MLPs as key-value memories, where facts live, how small arithmetic is computed, and why single neurons stop making sense.
The MLP block holds two thirds of the parameters yet gets a fraction of the attention. This chapter opens it up: normalization and its consequences for interpretability, neurons read as key-value memories, causal tracing of factual recall, arithmetic probes, layer ablation, and the polysemanticity that breaks single-neuron analysis. All quantities are precomputed from forward passes of Qwen3-1.7B.
The other two thirds
The SwiGLU block — gate, up and down projections and the SiLU nonlinearity.
Normalization, the part everyone skips
RMSNorm, pre- vs post-norm, and what normalization means for interpretability.
MLPs as key–value memories
Neurons as pattern → update pairs: what each key detects and each value writes.
Where facts live
Severed causal tracing and the three-step recall of factual associations.
How the model computes 8 − 6
Per-depth answer probes: where small arithmetic crystallises, MLP vs attention.
Ablation, again
Knocking out MLP layers to establish which are load-bearing.
Neurons: mostly not what you would hope
A census of neuron roles — a few clean detectors in a sea of polysemanticity.
The wrong basis
Why the neuron basis fails, and superposition and transcoders as the way out.
This is an interactive article — the figures are explorable and require JavaScript to load.
More in this series: The Residual Stream · Attention · all posts