AI has no shortage of money, attention, or new products. Yet a lot of the work still feels incremental to me: a better chat model, another assistant, another copy of something that already found a market.
Then Jev appeared.
I think Jev may be one of the more fundamental new directions in AI right now. It is early, and I do not know how far the idea will go. That is exactly why I am excited about it.
TypeSafe describes Jev as its first "System One Model," built for decisions inside software. Instead of generating a passage of text for a person to read, Jev answers structured questions with typed decisions, probabilities, and confidence that code can use directly. TypeSafe calls its training method Reinforcement Learning for Calibrated Decisions, or RLCD, and frames it as a different research direction from chat models trained around human preference. The practical idea is easy to grasp: software can act on a decision when confidence is high and route uncertain cases elsewhere.
For a quick first impression, watch Matija Sosic's 45-second introduction to Jev. Then read TypeSafe's own explanation of the model and the research direction behind it.
Why I built Jevfast
What interests me now is watching people test that idea. Where does Jev fit inside mature software? What new applications become possible when a model returns a small typed decision instead of a long generated answer? People are already trying it in games, databases, music tools, browser agents, and much stranger experiments.
The work is scattered across GitHub repositories, X posts, project websites, and videos. I made Jevfast as an independent place to find the good work more easily, follow each project back to its creator, and get ideas for your own experiments.
The name is deliberate. Speed may not be Jev's deepest innovation. Its typed output, calibrated probabilities, and training direction matter more. Still, speed is the first impression many people have, and it can change which ideas are practical. A decision model that can sit inside a tight loop invites applications that feel different from waiting for a chatbot to compose an answer.
That is why Jevfast has a realtime or nearly filter. It surfaces projects designed to call Jev in realtime or every few seconds.
How Jevfast works
Automation handles most of the repetitive work. It discovers public material from GitHub, X, and YouTube, cleans up and deduplicates the links, records evidence, and prepares reviewed entries for the site. Before anything appears, I check the public sources, confirm Jev's role, credit the creator, and separate demonstrated behavior from claims I could not verify.
Automated discovery will miss new or unusual work. If you built something with Jev, found a project I missed, or saw a breakthrough idea, please share the link through the submission page. You can submit a project or a substantive video. If I credited something incorrectly or got a detail wrong, use the same page or open an issue. Every submission is reviewed before inclusion.
A few places to start
Browse categories such as games, data and search, and automation. The videos page also collects announcements, tutorials, reports, reviews, and demos.
TypeSafe Mario by fhshaik reads structured emulator state, asks Jev to choose a controller action, and shows the decision through a dashboard.
pg-jev by realZachi brings Jev into PostgreSQL. Queries can filter, rank, or classify rows with a plain-language condition while still composing with familiar SQL.
Jevthoven by cocktailpeanut uses Jev for bounded choices in a musical composition, then turns them into editable notes and MIDI.
They share a model, but little else. The interesting part is seeing where each creator draws the boundary between a Jev decision and the rest of the program.
Jevfast will always be an edited snapshot. I use public evidence, but I do not run every project or reproduce every claim. I want creators to get the credit and readers to reach the original work.
If you are curious about Jev but unsure what to build, pick a category, open one project, and follow it to the source. I built Jevfast to make that first step easier, and to watch where this strange new direction goes next.
Comment on this X post to discuss this article.