The cost of AI and the work left undone

Two new practitioner essays examine the cost of reused context and the useful work broader AI access might make possible.

2 stories · Yesterday in AI

TL;DR

The short version
  1. Max Woolf’s new analysis makes cache reads a budget question: compare accepted work using actual billed categories, not a headline token price.
  2. Rao’s essay asks what people do with wider AI access; a bounded pilot can test useful requests that were previously left undone.

Big Moves

1 story

Rao’s AI essay asks who gets to discover useful work

Venkatesh Rao’s new essay shifts attention toward demand for AI. An enterprise pilot can test useful work that teams previously left undone.

As a leader: Invite workers to nominate neglected tasks and evaluate actual use and review effort alongside familiar efficiency metrics.

As a practitioner: Bring one previously unserved request to a bounded adoption pilot and identify who can judge its usefulness.

Venkatesh Rao / Contraptions ↗Practitioner perspective

Vendors

1 story

Model price cuts make cache accounting matter

Max Woolf’s new pricing analysis highlights cache reads; workload accounting determines whether lower token rates reduce the cost of accepted work.

As a leader: Compare representative accepted tasks using separate ordinary-input, cache-read, cache-write and output charges.

As a practitioner: Inspect one agent run’s billed usage categories and retain retries and correction work in the task-cost comparison.

Max Woolf ↗Practitioner perspective