hawkes.aiKey ideas
- The one idea — Input is read in one parallel pass; output is written one serial pass per token; cached input shares pages — the price sheet is the physics, itemized.
- How it is shown — The price board's three rates (cached input, input, output) cabled live into the three mechanisms (shared pages, prefill sweep, decode drumbeat).
- The trap to avoid — Optimizing prompts while ignoring output length — the reply is where the meter spins fastest.
- What it sets up — Who sets these prices and when to pay them at all.
An AI reads your entire essay in one parallel gulp. Then it writes its reply one token at a time, hauling weights for each. That asymmetry is on your bill.
The one idea
Input is read in one parallel pass; output is written one serial pass per token; cached input shares pages — the price sheet is the physics, itemized.
Pull up any AI price sheet: output tokens cost several times more than input — and "cached input" costs a fraction of either. Those aren't marketing choices. The price sheet is this act's physics, itemized. Let's trace each rate to its machine. Trace the input rate first. Your prompt is read in prefill — episode forty-eight — one parallel sweep digesting a thousand tokens, the machinery's favorite work. Substantial, but shared across the whole slab. Per token, it's a thin slice of one sweep. That's why reading you is the cheap direction. Now the tall column. Every output token is decode: one full pass — all weights hauled across the conveyor — per glyph.
How it works — the demo
The price board's three rates (cached input, input, output) cabled live into the three mechanisms (shared pages, prefill sweep, decode drumbeat).
A fifty-token reply is fifty complete machine-efforts; your thousand-token prompt was one. Parallel reading versus serial writing: the whole gap — episode one-thirteen's apprentice asymmetry, denominated in dollars. And the discount column: cached input. Episode one-twelve planted it — identical openings sharing physical pages. When requests repeat the same preamble, the machine points at the stored copy and charges a referencing fee — a fraction of fresh reading. That's why products ship enormous standing instructions without going broke — and why structuring prompts for reuse is real money. Now you can read bills like blueprints. A workload writing torrents — long generations, verbose agents — lives in the tall column; no prompt-tweaking hides it. A workload that hauls big repeated context and answers tersely rides the cheap lanes. Before optimizing, ask which column your money is in. The bill already told you.
The trap to avoid
Optimizing prompts while ignoring output length — the reply is where the meter spins fastest.
Why it matters — and what’s next
Who sets these prices and when to pay them at all.
The trap: optimization theater. Teams tweezer prompt tokens while outputs gush unexamined — but the meter spins several times faster on the reply side. The unglamorous fix beats the clever one: ask for concision, cap response lengths, stop paying for prose nobody reads. In a business built on serial writing, brevity isn't style. It's unit economics. You can now read the menu. Which raises the act's last economic question: should you be ordering from it at all? Renting tokens is one strategy; owning the machine another — and the honest break-even surprises almost everyone who runs it. Chips first, then that math. Next group.
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
Pull up any AI price sheet: output tokens cost several times more than input — and "cached input" costs a fraction of either. Those aren't marketing choices. The price sheet is this act's physics, itemized. Let's trace each rate to its machine.
Trace the input rate first. Your prompt is read in prefill — episode forty-eight — one parallel sweep digesting a thousand tokens, the machinery's favorite work. Substantial, but shared across the whole slab. Per token, it's a thin slice of one sweep. That's why reading you is the cheap direction.
Now the tall column. Every output token is decode: one full pass — all weights hauled across the conveyor — per glyph. A fifty-token reply is fifty complete machine-efforts; your thousand-token prompt was one. Parallel reading versus serial writing: the whole gap — episode one-thirteen's apprentice asymmetry, denominated in dollars.
And the discount column: cached input. Episode one-twelve planted it — identical openings sharing physical pages. When requests repeat the same preamble, the machine points at the stored copy and charges a referencing fee — a fraction of fresh reading. That's why products ship enormous standing instructions without going broke — and why structuring prompts for reuse is real money.
Now you can read bills like blueprints. A workload writing torrents — long generations, verbose agents — lives in the tall column; no prompt-tweaking hides it. A workload that hauls big repeated context and answers tersely rides the cheap lanes. Before optimizing, ask which column your money is in. The bill already told you.
The trap: optimization theater. Teams tweezer prompt tokens while outputs gush unexamined — but the meter spins several times faster on the reply side. The unglamorous fix beats the clever one: ask for concision, cap response lengths, stop paying for prose nobody reads. In a business built on serial writing, brevity isn't style. It's unit economics.
You can now read the menu. Which raises the act's last economic question: should you be ordering from it at all? Renting tokens is one strategy; owning the machine another — and the honest break-even surprises almost everyone who runs it. Chips first, then that math. Next group.