Jev Wikiracer
Race the AI across Wikipedia. From the same start article, you and Jev click links to reach a target — you tap, Jev makes one real decision per hop, picking the closest of dozens of links. First to the target wins. No key, just race.
You and Jev start on the same article and race to the target using only in-article Wikipedia links. You click; Jev makes one real decision per hop. First to the target wins.
The same demo is one Jev call
- Read the current article + the target
- List the page's outbound links (dozens of candidates)
- Send them to Jev as one
choicequestion - Jev picks the link closest to the target
- Click it, load the next page, repeat
const res = await fetch("https://jevtypesafeai.com/api/v1/decide", {
method: "POST",
headers: { Authorization: `Bearer ${process.env.JEV_API_KEY}` },
body: JSON.stringify({
state: `On ${current}. Target: ${target}.`, // current + target article
questions: {
hop: { type: "choice", instructions: "Which link gets closest to the target?",
criteria: links }, // { 0: "Philosophy", 1: "Espresso", … }
},
}),
});
const { answers } = await res.json();
go(links[answers.hop.choice]); // → "Coffeehouse" (0.58)See a real decision COFFEEHOUSE · 208ms
On "Coffee". Target: "Philosophy". Links: Coffeehouse, Espresso, Ethiopia, Enlightenment, Trade, …
How Jev races Wikipedia
Wikipedia racing (or "Wikiracing") is a classic game: get from one article to another using only the links inside each page. It's the perfect test for Jev — every hop is a choice among dozens or hundreds of options, and the model has to pick the one link that's a real stepping stone toward the target without wandering or hallucinating a page that isn't there.
Here you and Jev start on the same article. You click links to navigate; on its turn Jev sends the current page's candidate links to the API as one typed `choice` question — "which of these gets closest to the target?" — and clicks its pick, with the confidence shown. It's the same high-cardinality decision Jev does in Wikiracing benchmarks, made into a race you can actually win.
What is Jev?
Jev is TypeSafe AI's first System One model. Instead of writing prose like a chatbot, it reads a situation and returns a typed, calibrated decision — a choice between options, a score, or a yes/no probability — in tens to hundreds of milliseconds, with no hallucinations. Picking the best link out of a big list is exactly the kind of fast, bounded, high-cardinality decision it's built for.
The same primitive routes support tickets, gates risky agent tool calls, scores leads and prunes agent context in real products. Wikiracing is just a fun, visible way to watch it choose.
Tips to beat Jev
- Think in stepping stones: a broad hub article (a country, a field, a famous person) usually links toward more things than a narrow one.
- Use the filter box to find a promising link fast instead of scrolling.
- If the target is directly linked on your page, you'll see it highlighted — click it to win instantly.
- Jev is quick but only hops every ~1.7 seconds — a decisive human can out-click it.
FAQ
Is this real Wikipedia?
Yes — the links come live from the Wikipedia API, so every article and every link is the real thing. Jev navigates the same pages you do.
Does Jev actually decide each hop?
Yes. On each of its turns the current page's candidate links are sent to the Jev API as one typed choice question, and Jev returns the link it judges closest to the target, with a calibrated probability. It's not a scripted path-finder.
Can I beat Jev?
Often, yes — Jev only hops about once every 1.7 seconds, so a fast, well-planned human can win. That's the fun: you're racing a real, live decision model.
Is it free?
Free, no key — each of Jev's hops is a rate-limited call to the live Jev API, a fraction of a cent. The page shows the running cost.
More: Jev Pac-Man — outrun the AI · Jev Gomoku · All decision games · Jev API docs