Roll for AGI 掷骰到AGIRoll for AGI  掷骰到AGI

You're a startup. The goal is AGI.

Company assets: one sheet of graph paper, one ballpoint pen, one very small model. Roll the dice, move around the board, build on empty lots. Houses make tokens, and tokens times parameters is what a roll scores. Each level is an eval day. Come up short and the company folds, and you start over.

Making the numbers go up

• Five kinds of houses: text, code, image, video, audio.

• Data centers. Land on a street that has one and the whole street goes to work.

• MCP links, drawn as red yarn between two data centers. Each one sends the street at the other end to work again. There's no limit; in testing we've had runs with over fifty.

• GPUs bolt onto houses and pay out every time that house works.

• 69 upgrades that stick onto houses or dice faces, or fire once as commands. Stuck with the wrong one? sudo delete it.

One of our test runs ended with almost 1,000 cash unspent, because the player got tired of clicking MCP links.

Five eras, thirteen bosses

Text, Code, Multimodal, Agent, AGI. Each new era hands you new stuff and changes the rules. Every third level is a boss eval: Data Wall caps every house at 30 tokens, Pop Quiz gives you only 4 rolls, Copyright Lawsuit shuts down your image, video and audio houses, and then there's the Turing Test.

Boss evals are run by rivals with names like deepsick and Greg, and they fight back. Each one has a move. Snip goes after your biggest bundle of MCP links. Buyout takes some of your houses for the level. Copycat adds part of your best roll to the target. Headhunt poaches your Skills, and you only get them back by beating it again on your next model. Every time you ship a new model, they return to the same level, one version stronger.

Chance cards

Sixteen of them. Whatever you flip, you take.

"The investors looked at your pitch deck. They didn't understand it. They invested."

"The investors looked at your financials. They understood them. They pulled out."

"Your model's written apology went viral."

AGI isn't the end

Fifteen levels make one model. Clear the last eval and you release it: give it a name, pick up to three houses to distill into the next version, and keep whatever cash is left. Then it's a fresh board, back to the text era, and v2 begins. Targets keep climbing and there's no score cap. Each model is scored by its parameter count, and the leaderboard ranks models, not runs.

We wrote a bot to play for us. Out of 2,000 runs, 11% shipped a v2 and 1.7% made it past level 30.

Drawn in a founder's sketchbook

Houses are paper cutouts propped up on the page. Data centers are server racks colored in with marker, their screens forever reading "Training...". Someone has scribbled in the margin: "Q4 goal: AGI".