Where a model's facts actually live. · Source

Attention gets the credit. This layer keeps the memories.

Every transformer block has two halves. Attention decides what to look at — and takes all the credit. The other half, the feed-forward layer, sits right after it and quietly holds most of what the model actually knows. When it recalls that a lemon is sour or a river runs through a city, that fact most likely lives here — not in the looking, but in the remembering.

The famous half is the smaller one.

\text{attn: } 4d^2 \quad\text{vs}\quad \text{FFN: } 2\,d\,d_{ff} = 8d^2 \;\;(d_{ff}=4d)\;\Rightarrow\; \frac{8d^2}{12d^2}=\frac{2}{3}

Like an iceberg: attention is the bright tip above the water, but the feed-forward layer is the mass below. Count the big weight matrices in one block and roughly two-thirds of them sit in this single layer. The part everyone studies is the smaller part. And unlike attention, it works on each word alone — no mixing, no neighbors, just one token passing through.

First it tests each word against thousands of patterns.

h = \max(0,\; xW_1 + b_1)\in\mathbb{R}^{d_{ff}}, \qquad d_{ff} = 4d

Like a tray of compass needles: sweep a magnet over them and only the ones aligned to it swing. The feed-forward layer does this with each word — it widens it into a space about four times larger, where thousands of learned patterns wait. Each pattern measures how well the word matches it; a simple threshold keeps the strong hits lit and zeros the rest.

Every pattern is wired to a stored answer.

\text{FFN}(x) = \sum_{i=1}^{d_{ff}} f\!\left(k_i^{\top} x\right)\, v_i

Here's the secret: each lit unit is a key bound to a value. The key is the pattern it hunts for; the value is the slice of meaning it writes back. Like a set of inked stamps: whichever ones get pressed leave their mark, and the page ends up carrying all of them at once. A word's output is simply every fired value, added together.

So this is the model's memory — and it can be edited.

x \;\leftarrow\; x + \text{FFN}(x)

If a fact is a key wired to a value, then it has an address. And the layer never overwrites the word — it adds its recalled values onto the running stream. Like punching a piano roll: each hole triggers one stored note, so re-punching a single hole changes that one alone — the rest of the song plays on. Find the right values and a single fact can often be located and changed, leaving the rest untouched.

Modern models add a smarter gate.

\begin{aligned} \text{Swish}(z) &= z\,\sigma(z), \quad \sigma(z)=\tfrac{1}{1+e^{-z}} \\ \text{SwiGLU}(x) &= \big(\text{Swish}(xW_1)\otimes xW_3\big)\,W_2 \end{aligned}

Newer layers split the job in two. One branch proposes the content; a second branch becomes a gate that decides, channel by channel, how much of it gets through — then the two are multiplied. Like a stencil over spray paint: the paint is the content, the stencil decides where it lands. This gating, called SwiGLU, adds a third matrix, so models trim the hidden width to two-thirds and pay nothing extra.

Two moves, every block: gather, then remember.

\begin{aligned} x &\leftarrow x + \text{Attn}(x) && \text{(gather)} \\ x &\leftarrow x + \text{FFN}(x) && \text{(recall)} \end{aligned}

Now the whole block clicks together. Attention gathers context — it asks what here is relevant? Then the feed-forward layer answers from memory — here's what I know about it. Like a baton handoff: one runner carries the context in, the next carries the knowledge home. Stack that pair dozens of times — with the stabilizing norms tucked in between — and you have the engine: look, recall, look, recall.

🌱 Is remembering a fact different from looking it up?

We found where knowledge sits: keys wired to values, patterns wired to answers — a library that replies the instant the right question brushes past. But a library doesn't understand the books it holds. If everything the model knows is a value waiting for its key, is recalling a fact a kind of thought — or only a very fast lookup? And somewhere in all that wiring, where does knowing become understanding?