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
01 / EXPLORE
0 means awaiting a verified post
The founder of Jev released a 1-hour masterclass breaking down how Jev actually works: LLMs → Decisions → Verification → Coding Agents, covering why LLMs alone aren't enough, the three primitives behind Jev, and turning huge AI tasks into tiny reliable steps.
This post, aimed at beginners of CodeX and Claude, introduces the basic flow of the Jev model: input state and question, process through Jev, and output calibrated probability.
Arman uses Saul Goodman and Hector's bell to contrast LLMs with Jev, noting that Jev, built by TypeSafe AI, is designed to provide small, precise structured decision signals inside software.
Grok Bot handles execution; Jev decides the next step. This 22-min tutorial explains TypeSafe's System One model that makes decisions instead of chatting.
Jev has no decode stage; the answer comes from a single prefill pass, making it about 200x faster than a traditional LLM judge.
People in the community keep explaining how Jev differs from regular LLMs.
Jev uses context and typed questions to produce bounded decisions with probabilities, not another text generator.
Jev performs a scope check on every code turn, flags hunks the prompt did not ask for, judges per hunk rather than whole PR, and offers In scope / Check scope / N unrequested options. Sessions only for now.
Jev goes beyond text generation—it analyzes, evaluates, identifies gaps, decides what needs changing, and suggests next steps instead of blindly following instructions.
TypeSafe's Diogo Almeida says even with a billion dollars he wouldn't pre-train a foundation model from scratch; instead, Jev is built by slicing, dicing, and recombining existing stacks.
Introduces TypeSafe AI's Jev model, focused on typed probabilistic decisions including Noul (Yes/No), Choice (select an option), and Score (score on a defined scale), contrasted with LLM's open-ended generation.
The post observes that much of what looks like reasoning in agent infrastructure is actually just classification in disguise. Jev makes the split explicit: it doesn't generate text at all; instead, it takes a state and a typed question.