Google may choose AI models by task difficulty, Lyra reports

Google may be building a system that chooses an AI model according to how difficult a request looks. Lyra (@lyraxana) described the proposed “Smart AI Routing” in a post on 3 November, Kyiv time: simpler requests would go to a faster, less expensive model, while harder tasks would go to a heavier one. The post says the system is based on “Jev”, without explaining that name.

That would put the decision about how much computing power a task needs ahead of the answer itself. The potential benefit is shorter waits for routine work while reserving the heavier model for requests that need it. Whether that works depends on the router recognizing when a seemingly simple request is actually difficult.

Lyra also claims lower cost and latency with accuracy preserved, but supplies no measurements or implementation that could establish those gains. In a reply, the account says a connection to “createDecision” is possible; that is a suggestion, not a verified code discovery. The reported routing system has no release date or access terms in the cited posts.