hawkes.aiKey ideas
- The one idea — Logits are unnormalized scores over every candidate — the richest object in the pipeline, and almost all of it is discarded before you see a reply.
- How it is shown — Walking the skyline: peaks, rivals, the long tail; the same prompt's confident vs conflicted skylines.
- The trap to avoid — The reply you read is one roll over this landscape — judging the model by single outputs is judging a distribution by one sample.
- What it sets up — Logits as the control surface.
Before your AI answers, it weighs every possible next word and scores them all. Then the interface throws almost all of that away — you see one token.
The one idea
Logits are unnormalized scores over every candidate — the richest object in the pipeline, and almost all of it is discarded before you see a reply.
This is the model's real output. Not text — this city of scores, one tower per possible next token. Walk it with me — reading this skyline is reading the machine's mind. Read the districts. A favorite, towering. Rivals, close behind — live alternatives the model genuinely weighed. A middle district of plausible long-shots. And the tail: tens of thousands of slivers, each technically scored. The values are raw — any size, negative included — and individually meaningless. What carries information is the gaps: how far the favorite leads, how crowded the rivals, how fat the tail. And the shape is a state of mind. "The capital of France is" — a monarch skyline, one tower dwarfing the city: certainty.
How it works — the demo
Walking the skyline: peaks, rivals, the long tail; the same prompt's confident vs conflicted skylines.
"The best movie ever made is" — a squabbling council of near-equals: genuine conflict. Before any word gets chosen, the skyline already tells you whether the machine is sure, torn, or lost. Remember that; a whole trust-act episode measures it. Here's why this episode is load-bearing: the skyline is the control surface of applied AI. Forcing a model to output valid JSON? Darkening every illegal tower before the choice — you'll watch it done. Measuring confidence? Reading the favorite's lead. Steering, watermarking, constraining — all of it is operations on this city. Master the skyline and the back half of this series is just visiting machinery that edits it. And now the quiet scandal. Of this entire city — the rivals, the margins, the beautifully-shaped doubt — you receive one token.
The trap to avoid
The reply you read is one roll over this landscape — judging the model by single outputs is judging a distribution by one sample.
Why it matters — and what’s next
Logits as the control surface.
The interface discards the rest, every word, every reply. The model's most honest artifact, its full weighted candor, is thrown away at the door. APIs can expose slices of it; builders who use them hold information everyone else never sees. The trap follows directly: judging the model by a single reply. The reply is one roll over this landscape — episode two's slot machine, now with the full mechanism visible. The same skyline, rolled twice, can produce answers a critic would grade oppositely. Evaluate distributions, not anecdotes: it's the difference between measuring the machine and reviewing your luck. One thing keeps the city from commerce: the scores are raw — wild units, negatives, nothing summing to anything. You already know the machine that civilizes a mess like this; you watched it focus attention. Next episode, softmax's promised second appearance — and the skyline becomes "this word: forty-three percent."
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 2:45 of narration
This is the model's real output. Not text — this city of scores, one tower per possible next token. Walk it with me — reading this skyline is reading the machine's mind.
Read the districts. A favorite, towering. Rivals, close behind — live alternatives the model genuinely weighed. A middle district of plausible long-shots. And the tail: tens of thousands of slivers, each technically scored. The values are raw — any size, negative included — and individually meaningless. What carries information is the gaps: how far the favorite leads, how crowded the rivals, how fat the tail.
And the shape is a state of mind. "The capital of France is" — a monarch skyline, one tower dwarfing the city: certainty. "The best movie ever made is" — a squabbling council of near-equals: genuine conflict. Before any word gets chosen, the skyline already tells you whether the machine is sure, torn, or lost. Remember that; a whole trust-act episode measures it.
Here's why this episode is load-bearing: the skyline is the control surface of applied AI. Forcing a model to output valid JSON? Darkening every illegal tower before the choice — you'll watch it done. Measuring confidence? Reading the favorite's lead. Steering, watermarking, constraining — all of it is operations on this city. Master the skyline and the back half of this series is just visiting machinery that edits it.
And now the quiet scandal. Of this entire city — the rivals, the margins, the beautifully-shaped doubt — you receive one token. The interface discards the rest, every word, every reply. The model's most honest artifact, its full weighted candor, is thrown away at the door. APIs can expose slices of it; builders who use them hold information everyone else never sees.
The trap follows directly: judging the model by a single reply. The reply is one roll over this landscape — episode two's slot machine, now with the full mechanism visible. The same skyline, rolled twice, can produce answers a critic would grade oppositely. Evaluate distributions, not anecdotes: it's the difference between measuring the machine and reviewing your luck.
One thing keeps the city from commerce: the scores are raw — wild units, negatives, nothing summing to anything. You already know the machine that civilizes a mess like this; you watched it focus attention. Next episode, softmax's promised second appearance — and the skyline becomes "this word: forty-three percent."