What Jev is: a decision model that only outputs probabilities
TypeSafe's Jev generates nothing and only assigns probabilities to options; a video covers real uses like email classification, a Gmail add-on, agent command vetting and model routing, and notes it performs poorly on trading signals.
The most hyped AI launch of the month can't write a single sentence. It's called Jev, from TypeSafe AI, a startup founded by one of the GPT-4 paper's authors. You give it a piece of text plus a list of choices. It returns a probability for each. No paragraphs, no JSON to parse, and nothing it can invent beyond the options you set. That sounds small until you see what builders are doing with it: - Ryan Vogel ran about 1,700 emails through four checks (category, priority, spam, reply-worthy) and paid 18 cents - jevmail, an open-source Gmail add-on, sorts 1,000 messages a minute for $0.03 - agents use it as a bouncer, vetting a command like rm -rf before it executes - teams send each request to a cheap or pricey LLM instead of defaulting to the big one TypeSafe claims it's 193x faster and 444x cheaper than large LLMs on some tasks. An independent test put it at 97.1% accuracy with zero training, next to 97.6% for a BERT fine-tuned on that exact job. Then someone rebuilt the idea in the open. SemIf (formerly OpenJev) reads the answer straight from the probabilities of a small Qwen, works on a home GPU or right in your browser, and passed 4,000 GitHub stars in nine days. Jev still misses, though. Ryan also asked it for Bitcoin buy/sell signals every minute, and the results were poor enough that he advised against trading on it. Watch the 80-second video for how it works, where it shines, the fine print, and when a normal chatbot is still the better tool.
This post quotes another post: x.com/zodchiii/status/2101243146596384854