What the source published

Wonderhop (wonderhop.io) is a small browser game built in about two months by two developers from Montreal. It's free, needs no account, and refreshes daily: everyone on the site plays the same set — 4 rounds of 7 open-ended questions each day, like "name a food that a lot of people hate", with 30 seconds per question (40 in Abyss mode).

Each of the four modes has its own goal. In Sky you guess the answer most people would give — the more mainstream your answer, the higher you climb toward space. Abyss is the mirror image: guess an answer almost nobody gives, and the rarer it is, the deeper you sink toward the trench. Past and Future are about timelines: give the oldest or the most recent answer, and the game compares yours against other players' answers — a good one sends you centuries through time. Every answer scores up to 100 points, 700 per round, and there's a daily leaderboard.

In their Show HN post the authors shared the technical details. Answers are typed freely in 10 languages, so they had to cope with typos, plurals, nicknames and different ways of saying the same thing. Their solution is an AI referee: wired to knowledge APIs like Wikipedia, it checks whether an answer really exists and whether it's factually right, averaging under a second per judgment. Percentages start as estimates but converge quickly as real players answer; if the question bank misses a good answer, a few different players giving it gets it added automatically. A large part of the code was written with AI tools, they say, "but the artistic direction, the concept and every design choice are human-made" — and the game hides plenty of secrets and pop-culture references.

Sources: [1] [2]

Try it: play one round in each mode

  1. Open wonderhop.io (the /en/ path for English) — no sign-up, just start playing. Remember everyone gets the same daily set, so a missed day is gone for good; that's also what makes the daily leaderboard work.
  2. Play Sky first: forget being right and write whatever most people would write. National-level answers like "peanuts" or "Michael Jackson" are the template for a perfect Sky score.
  3. Switch to Abyss and invert your thinking: come up with something real but extremely obscure. "Real" is the key — made-up answers fail the AI referee, which actually checks; obscure-but-verifiable is what sinks deepest.
  4. If your answer is rejected, argue with the referee before retyping: one player convinced it that CompuServe counted as a file-sharing service by explaining "CompuServe had file sharing", and the answer was accepted. Appeals are part of the game.
  5. Finish with Past and Future, which test your sense of time — dig into history for "the oldest X", scan recent news for "the most recent". Seven questions, a 30-second timer, then check your rank on the daily board and come back tomorrow.

Sources: [1]

The full picture: how the AI referee works

The most interesting engineering here is how you judge answers to open-ended questions. The classic approach is a fixed answer bank, but with free-typed answers in 10 languages the bank can never keep up with players — too many typos and nicknames. So they handed judging to a large model: it runs an existence check (is this a real thing?), a factual check (hooked to knowledge sources like Wikipedia), and returns a verdict in under a second. Commenters found its weakness fast: describing an answer instead of naming it ("the city at the southern end of the Shinkansen") could sneak through, and the author replied they'd fix it within the hour.

The scoring system is worth a look too. The starting percentages are estimates, and every answer players give refines them — so your answer doesn't just decide your score, it helps shape the question bank. One player reported a mobile bug where the keyboard covered the text box; the author fixed it quickly. That "author online, iterating live" Show HN rhythm is a recurring plus in the 97-comment thread, and several players singled out the pixel-art look for praise.

Sources: [1] [2]

How to read this

Wonderhop's clever move is productizing "AI as referee". Plenty of people use large models to write code or prose, but using a model as the enforcer of the rules is rarer: it turns open-question judging — something that used to need human review — into a scalable, real-time service. For readers, the takeaway is that the judging step is no longer the bottleneck for lightweight interactive products like polls, guessing games or quiz toys — plug in a model, connect a knowledge source, and it runs.

The other signal worth noting is the division of labor. The authors are blunt: "a large part of the code was written with AI tools, but the artistic direction, the concept and every design choice are human-made." That's the archetypal indie-developer setup right now — AI turns ideas into working code, humans decide what to build and what it looks like. Two months, two people, one small game people come back to daily: that scale is the sweet spot for AI-assisted development.

Sources: [2]

Key takeaways

  • Four rounds of 7 questions a day, the same set for everyone, 30 seconds each (40 in Abyss); free, no account.
  • Four modes: Sky guesses the most mainstream answer, Abyss the most obscure, Past the oldest, Future the most recent; up to 100 points per answer, 700 per round.
  • An AI referee wired to knowledge sources like Wikipedia verifies each answer in under a second; rejected answers can be appealed by arguing with it.
  • Percentages converge with real players' answers, and good missing answers get added to the bank automatically.
  • Built in about two months by a two-person team in Montreal; AI tools wrote much of the code, humans made the art and design calls.

Sources: [1]