a builder's codex
codex · operators · Ethan Mollick · ins_rebaseline-quarterly-not-pin-to-snapshot

Pin AI workflows to capabilities you can re-baseline quarterly, not to one model snapshot

By Ethan Mollick · Professor, Wharton; author of *Co-Intelligence* and One Useful Thing · 2026-04-23 · essay · Sign of the Future — GPT-5.5

Tier B · TL;DR
Pin AI workflows to capabilities you can re-baseline quarterly, not to one model snapshot

Claim

Each new model release moves capabilities discontinuously — what was impossible becomes easy, and the size of leaps grows each cycle — so AI workflows must be pinned to capability targets you re-baseline quarterly, not to one model snapshot whose performance you optimize against.

Mechanism

A workflow optimized for the current model's exact behavior (its prompt patterns, context-window tricks, output shapes) accumulates dependencies that break when the next model arrives. The next model often makes the workaround unnecessary — the capability is now native — but the workflow has hardened around the workaround. Re-baselining quarterly means: define the capability target (e.g., "produce a credible 15-page strategy doc with primary research"), test it against the current model, then strip workarounds whose justification disappeared. Workflows pinned to capabilities outlive their underlying model snapshots; workflows pinned to a snapshot decay.

Conditions

Holds when:

Fails when:

Evidence

"Every few months a new model arrives... something that was impossible becomes easy, while the size of the leaps grows each new release cycle."

Honest concession: rough edges remain in long-form fiction generation. The frame applies to capability-led workflow design, not blind upgrade.

— Ethan Mollick, Sign of the Future: GPT-5.5, https://www.oneusefulthing.org/p/sign-of-the-future-gpt-55, 2026-04-23

Signals

Counter-evidence

Frequent re-baselining costs operator time and can destabilize workflows that depend on consistent behavior. For high-stakes regulated workflows, pinning to a known model snapshot is the safer move. The cadence (quarterly) is a default, not a universal.

Cross-references

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