hawkes.aiKey ideas
- The one idea — Softmax over the logits yields the next-token probability distribution — same machine that focused attention, now pricing the vocabulary.
- How it is shown — The city civilized: raw skyline in, clean percentage-districts out; every token kept alive with a sliver.
- The trap to avoid — "43% likely" means likely-as-next-text — not 43%-true; probability of text ≠ probability of fact.
- What it sets up — The never-zero tail = why any output is technically possible.
That forty-three percent likely inside an AI doesn't mean forty-three percent true. It means likely to be written — and popular myths get written a lot.
The one idea
Softmax over the logits yields the next-token probability distribution — same machine that focused attention, now pricing the vocabulary.
One sweep of a machine you already own, and the wild city becomes commerce: this word, forty-three percent. That word, twelve. Softmax — exactly as promised in the attention act — making its second appearance, at the exit. Same two moves you watched in episode thirty-four: exponentiate — gaps become gulfs, negatives become slivers — then normalize to sum to one. In attention, this machine split a reading budget across your context. Here it splits certainty across the entire vocabulary. One machine, two doors of the same tower. What you now hold is a genuine probability distribution: two hundred thousand entries, summing to exactly one, every single token strictly above zero. A complete betting market on what comes next — favorites, long-shots, and a dust-ring of never-quite-impossible.
How it works — the demo
The city civilized: raw skyline in, clean percentage-districts out; every token kept alive with a sliver.
And softmax's personality — never an exact zero — means something heavier at the exit than it did inside. Every token, always, holds a nonzero stake. The absurd continuation, the forbidden one, the catastrophic one: improbable, never impossible. That ember-field is a fact the security act will return to, carefully. Now the trap, and it's the deepest one at the summit: these are probabilities of text, not of truth. The market was trained on what tends to get written — so a popular misconception can hold a wide, confident district while the correct answer sulks in a sliver. Forty-three percent likely means likely-to-be-said. Whether the model even knows the difference between said and true — that question gets an astonishing answer in the trust act. Keep both objects on your bench.
The trap to avoid
"43% likely" means likely-as-next-text — not 43%-true; probability of text ≠ probability of fact.
Why it matters — and what’s next
The never-zero tail = why any output is technically possible.
Engineering interventions — masking illegal tokens, steering, measuring — operate on the raw logits, before the sweep. The chooser ahead consumes the finished probabilities. When a later episode says a tool "edits the logits," you'll know exactly where in this room it's standing. The market is priced. Two hundred thousand stakes, summing to one, waiting. One machine remains in this act — the one that finally rolls. Next: sampling. The dice you've heard about since episode two, built at last.
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
One sweep of a machine you already own, and the wild city becomes commerce: this word, forty-three percent. That word, twelve. Softmax — exactly as promised in the attention act — making its second appearance, at the exit.
Same two moves you watched in episode thirty-four: exponentiate — gaps become gulfs, negatives become slivers — then normalize to sum to one. In attention, this machine split a reading budget across your context. Here it splits certainty across the entire vocabulary. One machine, two doors of the same tower.
What you now hold is a genuine probability distribution: two hundred thousand entries, summing to exactly one, every single token strictly above zero. A complete betting market on what comes next — favorites, long-shots, and a dust-ring of never-quite-impossible.
And softmax's personality — never an exact zero — means something heavier at the exit than it did inside. Every token, always, holds a nonzero stake. The absurd continuation, the forbidden one, the catastrophic one: improbable, never impossible. That ember-field is a fact the security act will return to, carefully.
Now the trap, and it's the deepest one at the summit: these are probabilities of text, not of truth. The market was trained on what tends to get written — so a popular misconception can hold a wide, confident district while the correct answer sulks in a sliver. Forty-three percent likely means likely-to-be-said. Whether the model even knows the difference between said and true — that question gets an astonishing answer in the trust act.
Keep both objects on your bench. Engineering interventions — masking illegal tokens, steering, measuring — operate on the raw logits, before the sweep. The chooser ahead consumes the finished probabilities. When a later episode says a tool "edits the logits," you'll know exactly where in this room it's standing.
The market is priced. Two hundred thousand stakes, summing to one, waiting. One machine remains in this act — the one that finally rolls. Next: sampling. The dice you've heard about since episode two, built at last.