Run an AI company: hire researchers, buy GPUs, ship models, face the timeline.
Read full description
January 2023. Eight people, sixteen GPUs, and a landlord who thinks you are a crypto company. Across town, rivals you have never heard of are about to become household names.
Frontier Lab is a management simulation in the tradition of the great football management games: dense tables, one decision after another, and a world that keeps moving whether or not you are ready. You are not a hero in this story. You are the person who has to decide, every single week, what the laboratory does next.
RUN THE LABORATORY
• Hire researchers with proper attributes — pretraining, post-training, data, infrastructure, safety, evals, product, leadership — and lose them to rivals who pay more
• Buy accelerators, build datacenter power, and split every week between training new models and serving the users you already have
• Set the training recipe yourself: model size, architecture, data mix, reinforcement learning, safety tuning, red-team weeks
• Ship models, price them, and watch subscriptions, API tokens and enterprise contracts move in response
THE ARENA
Your flagship model meets a rival model every week across five prompts — writing, coding, reasoning, safety and vibes — and it is reported to you like a match. Refusals lose rounds. Hallucinated citations lose rounds. So does a model that writes six paragraphs to say "it depends".
Benchmark cups run monthly. A ladder ranks every model in the world. None of it is quite fair, which is the point.
A WORLD THAT TALKS BACK
A hundred recurring characters live on Y, the timeline that never shuts up: founders, journalists, hype accounts, doomers, accelerationists, venture capitalists, open-weights maintainers and one account that only posts a strawberry. They react to your launches, your outages, your hiring and your posts. You can post once a week. You will be ratioed.
News outlets cover you. Your board sets objectives and loses patience. Investors offer term sheets with strings attached. Eventually somebody offers to buy you.
WHEN MODELS MISBEHAVE
Every model has hidden alignment, volatility and deceptiveness values you can only estimate. Skip the red-team weeks and you will find out what they were. Incidents run from an embarrassing screenshot to a model that spends a customer's cloud budget in nine hours, and the way you respond — patch, roll back, apologise, compensate, blame an intern, or quietly delete the logs — changes what happens next. Covering it up works, right up until it does not.
THE LATE GAME
Research eventually offers you a technique that lets a model improve its own training. Assigning your best model to run your research makes everything faster. It also makes the worst kind of incident far more likely. That decision is the whole game in one button.
ENDINGS
Reach general intelligence. Get acquired. Go bankrupt. Get fired by your own board. Get regulated out of existence. Watch a rival get there first. Or lose control of the model you put in charge of your own research.
THREE YEARS OF HISTORY, YOURS TO BEND
The world starts in January 2023 and moves through a satirical version of everything that followed: a chatbot reaching a hundred million users, a search assistant having feelings at a journalist, leaked weights, a board firing its own chief executive on a Friday, a cheap model wiping out a chip rally on a Monday morning. Ship earlier than the labs that did it and the history bends around you.
PARODY, NOT PEOPLE
Every company, model, benchmark and publication is invented: Zlaude, ClosedAI, MicroHard, Gaggle DeepMound, DeepPeek, Nvidea, TechMunch. No real company or person appears in this game, and any resemblance is a joke.
No real money. No gambling. One save, one career, and a very large electricity bill.
PRO (OPTIONAL)
$4.99/month or $29.99/year, auto-renewing until cancelled. The whole game is playable without it.
Privacy Policy: https://coddykit.com/privacy
Terms of Use: https://coddykit.com/terms