THE JEV GUIDE · CURATED EDITION
Launch news, explanations, demos, and community projects around TypeSafe Jev, organized by topic. Every entry links back to its X source.
Curated selection · Not exhaustive · Updated manually
Sponsors
These products pay for the directory. Want yours up here?
Introduces Jev, a fast model from TypeSafe that does not chat or think at length, but simply picks from given options. It describes a setup where a rules layer filters legal moves and Jev selects among them.
The Jev model does not generate text; it provides typed responses and probabilities to automate software decisions and classify data. $0.042 per input, free output.
The post uses a triage nurse analogy to explain Jev as a traffic controller, not the final analyst, with the LLM as a specialist only when needed.
The post explains that Jev takes the current state and questions with predefined answers, returning probabilities over those answers; labels are defined by the app, so the model cannot invent new ones.
The post explores how Jev works as a decision model: most AI systems make a single LLM sort, route, judge, reason and write, while the Jev guide proposes a split—the LLM generates and Jev decides. It also covers where the decision model sits in a system and what the guide does not prove.
One card in the feed, every sixteen, on every page of it. $50 a week, cancel any week.
Run the Jev decision model locally via Ollama 0.35. Feed a text and get suggestions on which team to assign, whether to refund, and urgency. Each output has probabilities; the score (e.g., 0.815) is a probability-weighted average, so it should not be confused with confidence.
The author explores how to simulate Jev using an open-source language model and token probabilities, breaks down how Jev works, and shares a demo running with DiffusionGemma on Cloudflare Workers AI. Link
Jev checks each request before it goes to a large model, dividing tasks to the right model in under a second at almost no cost.
TypeSafe's Jev 1.13 takes context and typed questions, returns probabilities—no paragraphs to clean up, no JSON to coax.
Jev routes requests in a laptop Jarvis, reads a million tokens for 4.2 cents, outputs free, and picks an answer from your list.
With each model's cutoff set so at most 1 in 100 let-through answers are wrong, Jev lets 93% skip human review, OpenAI 66%, while big general models reach 98% but are about 10× slower.
hev mind argues that keeping half the benchmark private ensures rankings reflect real skills rather than memorized answers, calling it the right call. Link
