AI cost per successful outcome is more useful than cost per model call

Description and use

<p>AI economics should include retries, failures, supporting software and human review rather than counting only the nominal model-call cost. A cheap attempt can still produce an expensive workflow when the success rate is low or review work is high.</p> <h2> What is actually changing </h2> <p>AI economics should include retries, failures, supporting software and human review rather than counting only the nominal model-call cost.</p> <p>The useful move is to treat this as an operating constraint rather than a slogan. Make the boundary explicit, decide what evidence would confirm the decision,

Problem

<p>AI economics should include retries, failures, supporting software and human review rather than counting only the nominal model-call cost. A cheap attempt can still produce an expensive workflow when the success rate is low or review work is high.</p> <h2> What is actually changing </h2> <p>AI economics should include retries, failures, supporting software and human review rather than counting only the nominal model-call cost.</p> <p>The useful move is to treat this as an operating constraint rather than a slogan. Make the boundary explicit, decide what evidence would confirm the decision,

Outcome

<p>AI economics should include retries, failures, supporting software and human review rather than counting only the nominal model-call cost. A cheap attempt can still produce an expensive workflow when the success rate is low or review work is high.</p> <h2> What is actually changing </h2> <p>AI economics should include retries, failures, supporting software and human review rather than counting only the nominal model-call cost.</p> <p>The useful move is to treat this as an operating constraint rather than a slogan. Make the boundary explicit, decide what evidence would confirm the decision,

Evidence

AI cost per successful outcome is more useful than cost per model call

Source