Every Adventure is a route over the concept graph. Pick one up where you left it, or author a new one.
The companion that writes back to you in this app is still a black box. This Adventure cracks it open from the bottom — and it starts with the one operation a neural network does over and over: a matrix meeting a vector.
Read through the idea, end on a worked example. No drills, no gates.
Now it asks you to prove it — a live, Claude-judged battle.
One clean synthesis to bank the session's XP.
A single linear layer takes an input vector $x$, multiplies it by a weight matrix $W$, and adds a bias $b$:
That’s it — no magic. The whole behaviour of the layer is stored in the numbers of $W$ and $b$. “Learning” is just nudging those numbers.
If $W$ has shape $(4,3)$ and $x$ is a column vector of shape $(3,1)$, the product is defined because the inner $3$’s line up, and the output has shape $(4,1)$:
Reading shapes is the single most useful debugging skill in deep learning. Most bugs are a shape that doesn’t line up.
Each entry of the output is one row of $W$ dotted with $x$. The dot product is just “multiply matching parts and add them up”:
Its sign reads the angle between the two vectors: positive → they point the same way, $0$ → they’re orthogonal, negative → they oppose. Hold onto that — it is the seed of attention later in this Adventure.
Take one row of $W$ and the input $x$, and compute that output entry.
Do that for all four rows of a $(4,3)$ matrix and you get the full $(4,1)$ output vector. That’s a forward pass through one layer.
You ran out of HP — but you saw where it slipped. Here's what to relearn before the rematch. The rematch asks different problems at the same difficulty.
Your route through this Adventure. Tap where you are to pick it up.