The AI Task Frontier

Horizon, forgiveness, and directed AI innovation. The paper studies why the profitable direction of AI progress changes across task regions.

June 2026 Theory paper 99 math checks passed 14-page readable draft
Open PDF LaTeX source Math checker Learning log

Main Intuition

AI progress is not one scalar quality ladder. Tasks fail for different reasons. Some tasks require long chains of coordinated action. Other tasks can be short but unforgiving because a local error is costly or irreversible. The model turns that distinction into sufficient statistics for valuing frontier movement and project direction.

Sufficient Statistics

B_H values horizon progress, B_F values reliability in unforgiving environments, and C measures the value mass where both constraints bind.

Project Inversion

The same project menu can favor runtime, tools, or search in one task region and reliability or verification in another. The regional ratio B_H^R/B_F^R determines the direction.

Scalar Test

A one-dimensional quality ladder cannot generate common-project reversals without adding the missing task heterogeneity back through project effects or costs.

Revision Plan & Today's Learning

Current Plan

  • Treat this as a readable theory skeleton: one primitive, one project-value rule, one scalar/rank test.
  • Keep competition, data flywheels, stopping, and infrastructure outside the main paper unless they directly serve the core result.
  • Use any next pass to add only selective intuition or examples, not another layer of mechanisms.
  • Keep the appendix limited to proofs of results stated in the main text.

Today's Learning: June 28-29

  • Rounds 54-91 clarified the core object: task-side project value, not a generic AI competition story.
  • Round 91 rewrote the abstract around the economic puzzle, the horizon-by-forgiveness primitive, and the project-value rule.
  • Round 92 compressed the paper from a technical compendium into a 13-page readable theory draft.
  • Round 92 removed competition, data-flywheel, stopping, private-social, and infrastructure extensions from the main PDF.
  • Round 93 added readability scaffolding: a cleaner abstract, clearer scope, intuition before formulas, and examples before the main theorem.
  • The appendix now contains only core proofs for the main-text results.
Open full learning log from today

This panel is synced from the project-level learning.md log for the June 28-29 revision sequence.

Task Regions

High forgiveness
Low forgiveness
Low horizon
Product recommendation, ad copy, search help. Low-cost systems and imitation can be strong.
Driving maneuvers, medical dosage, payments control. Reliability and verification dominate.
High horizon
Coding, debugging, data analysis, research assistance. Runtime, tools, retries, and workflow integration matter most.
Drug discovery, critical infrastructure, autonomous organizations. Full-stack frontier advantages can persist.

Paper

The PDF preview is unavailable in this browser. Open the paper directly.

The compiled PDF, LaTeX source, and math checker in this folder are synchronized with the current revision.