hawkes.aiKey ideas
- The one idea — Softmax converts raw scores into a budget summing to 1 — exponential sharpening of winners, nothing ever exactly zero.
- How it is shown — The messy skyline entering the prism; a clean spending-budget emerging; the sharpen-vs-flatten personality.
- The trap to avoid — Reading attention weights as explanations — they show where information flowed, not why the model answered as it did.
- What it sets up — You'll meet this machine again at the output.
Deep in the machine, nothing is ever fully ignored. Every word keeps a filament of connection to every other — a blessing for learning, and a door that never quite closes.
The one idea
Softmax converts raw scores into a budget summing to 1 — exponential sharpening of winners, nothing ever exactly zero.
Raw scores are messy. Softmax turns them into a spotlight. Two moves inside the machine — and a personality worth knowing. Move one: exponentiate every score. That does two things at once — gaps become gulfs, so leaders pull dramatically ahead, and even negative scores become small positive numbers. Nothing dies. Move two: divide everything by the total. Now the numbers are fractions of one whole. A budget. Why force a sum of one? Because attention is a fixed budget being divided.
How it works — the demo
The messy skyline entering the prism; a clean spending-budget emerging; the sharpen-vs-flatten personality.
The token has one reading-visit to spend: seventy percent on "river," fifteen on a neighbor, slivers everywhere else. Every weight is now comparable, spendable, and interpretable as a share — which is exactly what the blending step next episode needs. Now the personality. Softmax never outputs an exact zero — exponentials can shrink forever but never die. Every token keeps at least a filament to every other. That's a blessing for learning: a connection that was wrong today can be strengthened tomorrow, because it was never severed. File the other edge of it: a channel that never fully closes is also a channel that can leak. The security act remembers this. And the machine has a dial. Feed it bigger scores and the budget sharpens toward winner-take-most; scale them down and it flattens toward a floodlight. Attention layers set this implicitly — but remember the dial itself, because when softmax reappears at the model's output, that dial gets a famous name: temperature.
The trap to avoid
Reading attention weights as explanations — they show where information flowed, not why the model answered as it did.
Why it matters — and what’s next
You'll meet this machine again at the output.
The trap — and entire research debates live here: reading attention maps as explanations. "See, it attended to the contract clause, so that's why it answered." Careful. The weights show where information flowed, not why the answer came out. Pretty attention pictures are evidence, never verdicts. The interpretability act will give you sharper instruments. So the visit stands at: scores, sharpened into a budget. One step left — the one the whole mechanism exists for. Spending the budget on the books, and watching a word rewrite itself. Next: the blend.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:30 of narration
Raw scores are messy. Softmax turns them into a spotlight. Two moves inside the machine — and a personality worth knowing.
Move one: exponentiate every score. That does two things at once — gaps become gulfs, so leaders pull dramatically ahead, and even negative scores become small positive numbers. Nothing dies. Move two: divide everything by the total. Now the numbers are fractions of one whole. A budget.
Why force a sum of one? Because attention is a fixed budget being divided. The token has one reading-visit to spend: seventy percent on "river," fifteen on a neighbor, slivers everywhere else. Every weight is now comparable, spendable, and interpretable as a share — which is exactly what the blending step next episode needs.
Now the personality. Softmax never outputs an exact zero — exponentials can shrink forever but never die. Every token keeps at least a filament to every other. That's a blessing for learning: a connection that was wrong today can be strengthened tomorrow, because it was never severed. File the other edge of it: a channel that never fully closes is also a channel that can leak. The security act remembers this.
And the machine has a dial. Feed it bigger scores and the budget sharpens toward winner-take-most; scale them down and it flattens toward a floodlight. Attention layers set this implicitly — but remember the dial itself, because when softmax reappears at the model's output, that dial gets a famous name: temperature.
The trap — and entire research debates live here: reading attention maps as explanations. "See, it attended to the contract clause, so that's why it answered." Careful. The weights show where information flowed, not why the answer came out. Pretty attention pictures are evidence, never verdicts. The interpretability act will give you sharper instruments.
So the visit stands at: scores, sharpened into a budget. One step left — the one the whole mechanism exists for. Spending the budget on the books, and watching a word rewrite itself. Next: the blend.