Tomasz Tunguz, the venture capitalist behind Theory Ventures, spent a June blog post arguing that a piece of code nearly every program contains, the if-then statement, is becoming a target for specialized AI rather than hand-written logic. His post is now more than three months old, so the products he tested have had time to mature, but the underlying argument still holds up as a data point on where AI infrastructure spending is heading.
Tunguz’s premise starts with a mundane example: sorting inventory in a grocery store, where a worker constantly has to decide whether an odd item is a banana, a plantain, or something else entirely. Traditional software handles that kind of branching with rigid rules. Frontier AI models can handle the exceptions, but Tunguz’s point is that using a full-size model for a simple classification call is overkill, like hiring a surgeon to check inventory.
To test that, he looked for if-then calls inside his own AI agent’s code that he had previously routed to a general-purpose model, and swapped roughly a quarter of them for two narrower tools built specifically to answer yes-or-no and multiple-choice questions: Jev, built by a company called TypeSafe, and SemIf, an open source project from developer Theodore Lee (a related open reproduction, kev, comes from developer Jared Palmer). Rather than generating text token by token, these tools run the attention math a single time, then read off each allowed answer’s odds directly, skipping the decoding step altogether.
Testing against a batch of 98 production email threads that he verified by hand, Tunguz reports classification accuracy jumped from 47 percent to between 80 and 82 percent once the specialized deciders took over those calls. He also cites a roughly 82 times cost reduction on a typical classification call, based on TypeSafe’s published pricing against Sonnet-class model rates, and TypeSafe’s own workflow evaluations claiming a 76x to 209x spread in per-case cost.
Those numbers deserve a caveat the original post does not dwell on. The accuracy comparison comes from Tunguz’s own single agent and a 98-thread sample he verified by hand, not an independent benchmark, and the cost figures TypeSafe published are the vendor’s own math, not a third party’s. SemIf’s benchmarks likewise come from the project’s own tests on one GPU and one frozen model, so the actual accuracy and speed a team gets will depend heavily on their specific classification task.
The pattern points to a broader shift in how AI infrastructure gets built. Frontier models still do the heavy lifting: training new systems and architecting how they work. Once a system is built and tested, calling on it millions of times over favors smaller, cheaper models tuned for one narrow decision. Tunguz frames this as evidence that “harnesses,” the surrounding code that orchestrates AI systems, are set to capture more of the profit margin as specialized components take over the repetitive parts of the job.
For any team running an AI agent with hard-coded classification or routing logic, audit those calls before assuming a general-purpose model is the only option: a narrow decider model, tested against your own hand-verified sample rather than a vendor’s benchmark, is now a live alternative worth pricing out.
Tomasz Tunguz first published this analysis, “AI Comes for the If Statement,” on his blog at tomtunguz.com on June 5, 2026.
