You're Being Promoted,
Not Replaced

Tim Menzies · Apr 2026

The Sutton Highway vs the "higher way"

Rich Sutton, 2019. The Bitter Lesson. Sixty years of evidence: brute-force compute beats human-engineered knowledge. Chess heuristics? Scale won. Edge detectors? Scale won. Phoneme models? Scale won.

His conclusion: stop baking your intelligence into the system. Just assume you can scale using tomorrow's CPU.

But there is a problem: SE isn't chess.

SE Isn't Chess

Every one of Sutton's examples shares a quiet assumption: you can assess the "What" the moment it's created. Move 47 — good or bad? Instant feedback.

Reliability? Measured eighteen months into prod. Maintainability? Ask whoever inherits the LLM-slop in 2027. Understandability? Good luck explaining it to the regulator at 2am.

Brooks 1976 knew this. The V-diagram isn't just testing — it's epistemology. After project i, you carry domain knowledge into project i+1. The code is not the artifact. What you learned building it is the artifact.

Swartout 1983, XPLAIN system. Blunt finding: jump straight to code and you lose the Why. The domain principles that drove the refinement — gone. The system can't justify itself. Forty years later, trillion parameters, same problem.

The MSR'26 Hangover

This isn't theory. Look at the posters. Pattern after pattern:

The sugar high Sutton's highway promises — followed by a maintenance crisis that lands, as always, on you.

Predicted outcome. Not a bug.

The Higher Way

Not rejection of AI. Elevation of the engineer.

Domain models are the intermediaries. Markdown skill files. DSLs that describe the why in fifty lines instead of fifty thousand. Post-conditions from specs. Fuzzing mutators carried forward from prior projects. Symbolic. Browsable. A human can challenge them.

Lethbridge: compact symbolic models actively improve LLM performance — less hallucination, tighter reasoning. Pugh: these models are where you define experiments. Does X beat Y? Now you have a testbed instead of a prayer.