When Intelligence Gets Cheap
Epoch AI estimated fixed-performance inference-price declines across six benchmarks from 9-fold to 900-fold per year, with a median of 50-fold. That is a historical analysis of benchmark-equivalent API prices, not a forecast or the total cost of accepted work. Epoch AI’s analysis
When a draft becomes cheaper, a team may attempt more work. In one hypothetical progression, a morning brief starts as a summary of one approved contract and calendar. Its assignment later expands to more context, a proposed decision, and a draft reminder. Each expansion adds context selection, source checks, a reviewer, and an authority boundary. Cheaper generation can support more candidates. It does not make one of them correct.
The coal argument is a warning, not a law
In The Coal Question, William Stanley Jevons argued that more efficient steam engines could enlarge the field of use for coal. His claim concerned coal in nineteenth-century Britain. Read chapter VII
Energy research treats rebound as conditional. UKERC describes disputed effects that depend on context, and Sorrell finds economy-wide backfire difficult to test rather than universal. UKERC’s review Sorrell’s review Applying that pressure to AI is an operating hypothesis: lower cost may invite richer context, more candidates, and more consequential delegation.
One firm’s staggered customer-support rollout increased successfully resolved issues per hour by 15% on average. The deployment also used outcome-labeled examples, top-performer knowledge, internal-document links, training, and human final authority. The study does not isolate the contribution of each complement. Brynjolfsson, Li, and Raymond
The review queue is part of the cost
Imagine the brief proposes Friday as a deadline when the linked contract says Thursday. The reviewer rejects the candidate, restores the date from the source, records the correction, and accepts the brief for planning. Sending a reminder remains a separate authorization. If the conflict cannot be resolved, the brief stays unaccepted.
That is accepted work: a result that meets a declared rubric after review and correction. Measure the review minutes, rejected candidates, retries, and exception handling beside generation cost. A fast draft that creates a larger review queue has moved the work rather than removed it.
Egress means context leaving the controlled environment. Before a cheaper route handles a consequential task, minimize the context, protect supported sensitive values, verify the transformed input, and preserve the source trail. A provider setting about model improvement does not answer whether the content was transmitted, retained, reviewed, or allowed to reach a connected tool.
Evaluate the actions behind an AI result
A correct answer is only part of a good AI workflow. Define what an assistant may do, what evidence it keeps, and how you recover when it goes wrong.
The Anatomy of a 10x Employee
High impact is accepted work plus a more capable next cycle, not message volume or rescues. What the research says about experts, managers, and handoffs.
The 10x Employer
Only 39% of firms in McKinsey's 2025 survey saw any EBIT impact from AI. Count review, rework, and recovery before you call it a 10x gain.
A Privacy Toggle Is Not a Firewall
A training toggle answers one question. Safe AI use needs four boundaries: training, transmission, retention, and tool authority. Test the action.
