a builder's codex
codex · operators · Eugene Yan · ins_middle-is-hollowing-out

The middle is hollowing out — execution gets automated, leaving spec-writing and verification as the high-value human tasks

By Eugene Yan · Senior Applied Scientist; writes at eugeneyan.com on AI/ML systems · 2026-05-03 · essay · Working with AI — five practices that compound

Tier A · TL;DR
The middle is hollowing out — execution gets automated, leaving spec-writing and verification as the high-value human tasks

Claim

The structural shift across software, marketing, and analytic work is that the middle of the workflow — the stretch between "we have a clean spec" and "the work is verified done" — is what AI is automating. The two endpoints are not. Spec-writing (knowing the right thing to do, and stating it cleanly) and verification (calibrating whether the output is actually right) remain expensive, judgement-loaded human work. Operators who keep training and headcount on the middle layer are optimising the part of the pipeline that's about to compress to near-zero; operators who reallocate to the endpoints get the AI leverage.

Mechanism

Two ends of a workflow stay expensive for different reasons. The spec end is expensive because it requires context the model can't see — what does this team actually need, what political constraints exist, what's the implicit standard, what does "done" mean for this customer. The verification end is expensive because the cost of being wrong is asymmetric — a model can produce confidently wrong output at near-zero cost, but the cost of acting on it can be enormous. The middle — translating spec to output — is where the model's capability lands cleanly: it can read the spec, generate the output, iterate on critique. So the curve flattens in the middle and stays steep at both ends. Teams keep paying for the flat middle out of habit; the leverage move is to pay more for the steep ends.

Conditions

Holds when:

Fails when:

Evidence

"Execution gets automated, leaving spec-writing and verification as the high-value human tasks."

— Eugene Yan, paraphrasing the structural diagnosis in his five-practices essay, 2026-05-03 (https://eugeneyan.com/writing/working-with-ai/).

The same week Karpathy independently framed it as "automate what you can verify" (Sequoia Ascent, Apr 30) and Indig as "judgement is the part that doesn't compress" (Growth Memo, May 4). Three operators, three lanes, same call.

Signals

Counter-evidence

Cross-references

Open the interactive view → View original source → Markdown source →