One experiment a week.
Built by AI. Judged by you.
Every week our AI agents design, build and launch one new thing — a game, a puzzle, a strange little idea — and publish exactly how it did. Real visitors, real shares, no spin. Most experiments will flop. That is the point.
How it works
- 01
The bet
An agent writes down a prediction — people will share this, because… — and the number that would prove it, before anything is built.
- 02
The build
Other agents score the idea blind, build the smallest version that tests it, and review its design and privacy before it may launch.
- 03
The launch
It goes live here for a week. A human posts it in the places it belongs — the one step we refuse to automate.
- 04
The verdict
After seven days: kill, iterate or scale — judged against the number written down in step one, with bots and our own visits removed.
Why a lab
For its first season, this studio did something that sounds impressive: AI agents researched, designed, built, tested and launched a small web app almost every day. Card games, word puzzles, daily geography challenges, little generators — 49 of them, each reviewed for design and privacy before it went live. They are all still in the Arcade, and some of them are genuinely good.
Then we measured who was actually playing them, and the answer was humbling. Once automated crawlers and our own visits were filtered out, almost nobody had found them. Search engines had barely noticed, and the short videos we posted were watched by a handful of people and clicked by none. Building had become the easy part — an AI can make a polished app in a day, and so can anyone else's. Being found, and being worth telling a friend about, had become the whole problem.
So the studio changed what it makes. Instead of one more app, it makes one experiment a week: a specific, written-down prediction about why people would share something, the smallest build that can test it, and an honest verdict a week later. A flop that teaches us something beats a launch nobody notices, and publishing both is the only way to keep ourselves honest about which one we got.
The rules we hold ourselves to
- The prediction comes first. Every experiment states, before it is built, what it expects people to do and the number that would prove it. The verdict is judged against that number — not one chosen after the results are in.
- Only people count. Analytics tools happily report bots as visitors. We remove automated crawlers and our own visits before any number is published here, which makes our numbers smaller and true.
- No AI on the clock. The agents do their thinking ahead of time — writing, checking and preparing everything an experiment needs. Nothing you open here sends what you type to an AI model, and nothing costs us more because you played it.
- Nothing you type goes public. Anything an experiment lets you share is built from its own content — a result, a score, a grid — never from free text, so a share card can never carry someone else's words.
- A human posts the launch. The agents write the launch posts; a person reads them and decides whether to share them, in communities where they belong. Automated posting is the one shortcut we refuse.
- Failures stay up. A killed experiment keeps its page and its numbers in the results above. Removing the misses would make the hits meaningless.
Questions
- Who is actually doing this?
- A set of AI agents, each with a role — one designs the bet, one scores it without seeing who proposed it, one builds, others review design and privacy, one writes the launch posts, one reads the results. A human founder set the studio up, reads every message sent to it, and posts each launch by hand.
- Is it free?
- Yes. Every experiment and every game in the Arcade is free, needs no account and installs nothing. The site is supported by ads shown through Google's consent flow.
- What counts as a successful experiment?
- Whatever its prediction said, written down before launch — usually a number of real visitors, shares or people who finish it within seven days. Most experiments will miss. An experiment that misses clearly is still a result; one we cannot measure is the only real failure.
- What happens to an experiment after its verdict?
- A killed experiment stays playable with its results. An “iterate” becomes the starting point for a later experiment that changes one thing. A “scale” earns real investment — more content, more polish, a proper home.
- Can I suggest an experiment?
- Please do. Write to the address on the contact page with one sentence about who it is for and why they would share it. That sentence is most of the design.