Closed-form · Z₁
Deterministic tools.
Formatters, AST tools, lint rules, refactors with a fully-specified output. The answer is determined by the input.
microseconds · verifiable · cacheable
A What-If
What does coding-AI actually cost?
Three coding tasks. Three different kinds of work. A five-act narrative — what becomes possible when each task routes to the mechanism that fits its shape.
OpenIE · Synthesis A what-if · mapping coding work to information synthesis
ACT 01 · THE QUESTION
Closed-form work.
format selection
rename variable
sort imports
convert tabs → spaces
add type annotation
rename variable
sort imports
convert tabs → spaces
add type annotation
Constraint-bound work.
resolve dependency versions
schedule jobs on cores
verify function termination
minimize a failing test
satisfy type constraints
schedule jobs on cores
verify function termination
minimize a failing test
satisfy type constraints
Pattern-bound work.
write a test matching our style
add a route following the pattern
implement the existing interface
fix the dependency warning
add a route following the pattern
implement the existing interface
fix the dependency warning
Open-ended synthesis.
design the event schema
choose Redis vs Kafka
debug a race condition
write the ADR
choose Redis vs Kafka
debug a race condition
write the ADR
Energy expenditure
Instrument 0.0 mJ
Baseline (current AI) 0.0 J
Reduction —×
Gears engaged
Deterministic 0 0%
Verifier 0 0%
Retrieve + adapt 0 0%
Small model 0 0%
Frontier model 0 0%
Audit trail · per-task routing
Energy freed by routing
Tasks routed to lighter gears0
Tokens not needed0
Joules made available0.0
What you just watched
Three properties of the work. Three mechanisms that fit.
Pattern-bound · Z₂
Retrieve + adapt.
Tests matching your style, routes following the pattern, code following an existing interface. Known shapes, applied to new instances.
retrieval + small model · citation-anchored
Open-ended · Z₃
Frontier model.
Architecture, design, debugging the unprecedented, ADRs. No template fits. Synthesis under soft constraints.
the model engages where it fits
The frontier model is still here. Doing what it does beautifully. Alongside the simpler mechanisms that handle what they handle well. The work gets done. The output is the same. The energy difference is what becomes available — for longer sessions, smaller devices, larger fleets, more people.
Continue
The principle, on every input.
The walkthrough applies it to coding. The instrument lets the terrain run free.