⟨ H, D ⟩
Human writes a script
A=H
σ=D
Authoring ~30 min @ $75/hr
Per execution ~3 sec, ≈$0
Tools used pdftotext, grep
First-run total $37.50
Next 500 PDFs: free. Next 50,000: free.
Cost
The substrate doesn’t care. The bill does.
Both routes solve the same task on the same substrate. The substrate is neutral to selection. The five things that pay the bill — energy, time, money, compute, complexity — are not. The choice of cell determines the trajectory.
Worked example · Same task, four cells
Extract invoice numbers from 500 PDFs.
A representative Z₁-shaped problem: deterministic input, deterministic output, no judgment required. The substrate’s agency permits all four routes. The economics do not.
The task, in formal terms
∀ pdf ∈ {1..500} · extract invoice_number(pdf) → CSV
⟨ S, D ⟩
Model writes the script
A=S
σ=D
Emission (one Z₂ call) ~$0.02
Per execution ~3 sec, ≈$0
Tools used pdftotext, grep
First-run total $0.02
Next 500 PDFs: free. Authoring already amortized.
⟨ H, S ⟩
Human does it by hand
A=H
σ=S
Per PDF ~20 sec attention
Total time ~2.8 hours
Cost @ $75/hr $208
First-run total $208.33
Next 500 PDFs: pay again. Reproducibility: none.
⟨ S, S ⟩
Agent loops on each PDF
A=S
σ=S
Iterations / PDF ~3 model calls
Total calls ~1,500
Cost @ $0.005/call $7.50
First-run total $7.50
Next 500 PDFs: pay again. And again. And again.
Five axes pay the bill
The same task, five costs.
Energy, time, money, compute, and complexity. Each cell weighs differently on each axis. A single dollar figure hides four of the five.
| Axis | ⟨ H, D ⟩ | ⟨ S, D ⟩ | ⟨ H, S ⟩ | ⟨ S, S ⟩ |
|---|---|---|---|---|
| Energy joules dissipated | | | | |
| Time wall-clock to result | | | | |
| Money dollars per batch | | | | |
| Compute transformer FLOPs | | | | |
| Complexity moving parts, failure modes | | | | |
The dollar figure for ⟨S, S⟩ is small per batch. The compute and complexity figures are not, and they recur on every batch.
Cost trajectory
Ten batches of 500 PDFs.
Log scale on the y-axis. Linear scale would flatten both deterministic-output cells to invisible zero — which is the point: they don’t move.
⟨H,D⟩ human writes a script ⟨S,D⟩ model writes the script ⟨H,S⟩ human by hand ⟨S,S⟩ agent loop
After 10 batches: ⟨H,D⟩ = $37.50 · ⟨S,D⟩ = $0.02 · ⟨H,S⟩ = $2,083 · ⟨S,S⟩ = $75.00
After 100 batches: ⟨H,D⟩ = $37.50 · ⟨S,D⟩ = $0.02 · ⟨H,S⟩ = $20,833 · ⟨S,S⟩ = $750
The deterministic-output cells flatten. The stochastic-output cells climb linearly forever. ⟨S,D⟩ is 3,750× lower-cost than ⟨S,S⟩ after ten runs, and the gap widens with every batch.
Why the bill compounds
The nested agent.
The stochastic-output cell pays per iteration. Each iteration is a fresh draw, and many iterations call models that call models that call tools that call models. The doll opens, and there is another doll inside, and another. The bill is the depth of the nest times the count of items times the price per call.
A human in the ⟨H, S⟩ cell does the recursion in their own head — one doll, opened many times, paid in attention. The ⟨S, S⟩ cell does the same recursion in silicon, paid in tokens, and the doll is opened thousands of times more often.
Both routes solve the same task on the same substrate.
The substrate does not care which one is chosen.
The bill does.
The lesson is not that one cell is good and another bad. Each cell has a place. ⟨H, S⟩ is where novelty is found — design, judgment, the prose that no template would have written. ⟨S, S⟩ is where exploration happens — possibility-space search, brainstorming, drafts at speed.
The lesson is that for any task whose output is deterministic in shape — an invoice number, a sum, a CSV column — the right cell is one of the two deterministic-output cells. Authoring once, executing forever, at the substrate’s floor. The Z₁ artifact ⟨S, D⟩ emits is the lowest-cost agent ever built.
When a deterministic-output task is routed to a stochastic-output cell, the cost recurs on every batch instead of amortizing once. The substrate accepts both routes. The bill is what tells the two apart.
Continue