Sestina
Blog Talk

Non-linearity and the jagged frontier

A talk on why AI's effect on output is non-linear, and how to navigate the jagged frontier in real work.

Jason CornwallFounder, Sestina · August 2026

AI's effect is not a lineA jagged output curve shows two nearly identical asks producing opposite outcomes: approximately thirty-three times output and one tenth output.what you expected≈33×0.1×effortoutputthe same ask…AI's effect is not a lineA mobile composition of the jagged output curve showing approximately thirty-three times output and one tenth output from nearly identical asks.what you expected≈33×0.1×effortoutputthe same ask…
Two asks, a hand-width apart—one returns ≈33×, the other a tenth. Averages hide this; the shape is the story.

AI can turn a year of engineering into eleven days. It can also make a single slide take ten times longer than doing it by hand. Neither result explains the other away. Both are points on the same jagged frontier.

Capability does not advance evenly, and returns do not belong to the model alone. They emerge from the fit between the task, the current frontier, the quality of the context, the cost of checking the work, and the operator's judgment. Move one of those things and the multiplier can change sign.

This page is a companion to the talk. The deck follows the argument through three real workflows, then turns it into a practical way to decide where AI belongs in your own.

Continue with the full talk.Open the deck
Sources & further reading

Reading list

A short path through the research and case studies behind the talk. Full citations appear in the deck.

  1. 01Navigating the Jagged Technological FrontierDell'Acqua et al. · Harvard Business School / BCGThe field experiment that named the frontier and found opposite effects on tasks either side of it.
  2. 02Centaurs and Cyborgs on the Jagged FrontierEthan Mollick · One Useful ThingAn accessible account of two successful patterns for dividing work between people and AI.
  3. 03Rewriting Bun in RustJarred Sumner · BunThe million-line, eleven-day rewrite—and the unusually strong verification system that made it possible.
  4. 04Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer ProductivityBecker et al. · METRA randomized trial where experienced maintainers expected a speedup and instead took 19% longer.
  5. 05AI and Engineering Velocity: A Longitudinal AnalysisAbi Noda · DXData from more than 400 organizations on why local gains translate into smaller changes in end-to-end throughput.
  6. 06The Shape of Things to ComeSteve YeggeA vivid account of agent fleets, bespoke harnesses, and workflows that become part of the product itself.