Jev, SemIf AI deciders cut if-then costs 99%

Tunguz builds his argument from an everyday analogy: grocery store rules like 'if banana, then produce' and their endless exceptions. Coding, he writes, means encoding those rules and exceptions into software, and AI has been used to handle the exceptions, such as identifying an unfamiliar fruit as Musa paradisiaca. But he argues the world's most brilliant model is not needed for if-plantain-then-produce logic.

The body then moves from that intuition to a concrete test. Tunguz went looking for if-then statements in his own code that he had handed to AI, and within a few minutes replaced about a quarter of those calls in one of his agents. Using Jev and SemIf, the specialized deciders nearly doubled classification accuracy on 98 hand-verified production email threads, jumping from 47% to over 80%.

He frames this as evidence that software primitives can be optimized one by one, with the if-then statement showing nearly two orders of magnitude in cost reduction alongside higher accuracy. That leads him to describe a bifurcating economics of AI: frontier models for discovery and architecture, narrower models for hardened production workflows. The open question he raises is which other primitives will follow, and whether harnesses are about to capture a lot more margin — a claim that hinges on whether this pattern repeats.