Typesafe’s Jev cuts the chat from model-driven decisions

A narrower job for AI

Most current models default to a conversational interface. Jev takes the opposite approach. Its output is a probability distribution over predefined options, giving it a smaller and more predictable range of behavior than a general-purpose language model.

That makes Jev resemble deterministic programming, although it is not deterministic itself. A prompt still produces a probabilistic result, but the model is aimed at a bounded decision rather than an open-ended answer.

The distinction matters in agentic systems, where one task can generate many model calls. If every small decision goes through a general large language model, latency and cost accumulate quickly.

Fewer tokens, fewer wasted computations

Jev generates substantially fewer tokens for the same prompt than conversational models. That matters because every additional piece of text requires computation: the model turns ideas into words and then processes those words.

Technology consultant David Linthicum, cited by Anirban Ghoshal in InfoWorld, argues that companies often use general-purpose large language models as a universal intelligence layer even when an application needs only a fast, limited decision. He compares that choice with using a full enterprise service bus to answer a routing question whose possible answers are simply “yes” or “no.”

A Jev-like model could handle a slice of those decisions through probabilistic function calls, while general models remain responsible for tasks that actually require their broader capabilities.

That design could fit tasks such as:

The architecture described by GitHub user Vinnylarouge follows a compact sequence. Jev turns each option into a query vector: a short list of numbers representing its text. It then assigns attention weights to the context tokens, producing one context vector for each option. A shared dot product turns every option-context pair into a score, and softmax converts those scores into probabilities whose sum is one.

In a guide to this type of model, NNNN shows how the approach can be used for browser-state assessment, click prediction, dynamic branching and multi-tab control. Jev is trained with reinforcement learning on calibrated decisions, or RLCD.

The model that does not need to speak

The practical case for Jev is strongest where a business process needs a number, not an explanation. A silent model can sit inside a workflow and make constrained choices without generating a long textual response that no one will read.

The idea also explains the name. Jev refers to the 19th-century English economist William Stanley Jevons, associated with the Jevons paradox: improving the efficiency of coal use could increase, rather than reduce, coal consumption.

The same logic maps neatly onto the AI industry’s bet on tokens. If token processing becomes cheaper and more efficient, total consumption might rise rather than fall. That would require a token market broad enough to resemble the energy market.

That outcome is not established. Thomas Claburn writes in The Register that many people with access to AI tools either find no use for them or avoid them for moral reasons.

My read is that Jev is less a challenge to large language models than a challenge to their default placement. Using a general model for every decision is convenient, but convenience can hide an expensive mismatch between the task and the machine. The quieter issue is demand: Jev can make bounded decisions cheaper, but it cannot by itself create a reason for businesses to make more of them.

If the Jevons analogy holds, lower decision costs will multiply AI calls across business processes. If it does not, Jev may remain a precise solution to a narrow class of problems rather than the foundation of a much larger token economy.