No other language version is available for this page.
Each of these seven games is built around exactly one thing a machine learning system really does — and the machine in the game does it for real, not as a cutscene. In MIMIC a feature-agreement learner actually trains on the labels you spend. In OVERFIT the boss really does fit a model to your movement between waves and shoot where that model says you will be. In GHOSTLINE the ghost you race is built from your own finished laps. They run in a browser tab, they are free, and none of them asks for an account.
All seven launched on 2026-09-06 and are free to play right now — the full games hub also carries eleven daily logic puzzles.
| Game | The AI idea | What you do | Kind | Time |
|---|---|---|---|---|
| ⌨ PROMPT | Specification gaming | Write instructions a literal interpreter cannot twist | Puzzle | 10–20 min |
| 🧠 OVERFIT | Overfitting to a small sample | Feed the boss a habit, then break it | Arcade | 5–15 min |
| 🧪 MIMIC | Labelling and active learning | Spend as few labels as possible, then let it sort the rest | Puzzle | 5–10 min |
| 👁 OVERSEER | Human oversight at scale | Watch eight agents, halt the one that drifted | Strategy | 5–15 min |
| 📉 MINIMA | Gradient descent and local minima | Find the lowest point feeling only the slope | Puzzle | 5–10 min |
| 🏭 SINGULARITY INC. | Capability scaling and misalignment | Label data, buy GPUs, train models that earn while you are away | Idle | idle, ongoing |
| 🏎 GHOSTLINE | Imitation learning | Race a ghost trained on your own driving | Racing | 3–10 min a run |
Plenty of games are about artificial intelligence in the way a film is about it: the subject is in the story and the systems underneath are ordinary. The test we set ourselves for this set was narrower — in each game, the thing the machine does has to be the thing the title claims, computed live, and visible enough that you can work against it.
That is why OVERFIT shows you its training report between waves instead of hiding it, why MIMIC ends by naming the feature it actually latched onto, and why GHOSTLINE's ghost gets faster only after runs you finish. If the mechanic were faked, each of those reveals would be a lie you could catch in one session. It also sets the limit honestly: these are small, hand-built systems, not language models, and no game here calls an external model.
You give a machine a route. It follows the words, not the intention — and it takes every shortcut you forgot to forbid. Six levels, and the difficulty is never dexterity; it is that your specification was incomplete and you could not see the hole until the machine walked through it.
The later levels add a SEEK instruction that lets the machine plan its own path to a target. That is where the red shortcut stops being a trap you fell into and becomes a thing the planner reliably chooses, because it is genuinely the cheapest route under the objective you wrote. This is the classic specification-gaming failure in the smallest form we could build: nothing is broken, nothing is buggy, and the outcome is still not what you meant.
Play PROMPT → Free · no account · 10–20 min · also on itch.io
Between waves the boss trains on how you moved and then tells you what it learned — the training report is on screen, not hidden. It aims at where its model says you are going, so a clean, consistent, skilful pattern is the most dangerous thing you can give it.
Winning means deliberately being predictable, then stopping. The combo mechanic pays you for the moment the model is confidently wrong. It is the cheapest available intuition for why a model that fits a small sample beautifully can fail the instant the distribution moves — you are the distribution, and you can move it on purpose.
Play OVERFIT → Free · no account · 5–15 min · also on itch.io
You know the rule. The learner does not. Every item you label costs you, and at the end the model classifies twenty-six items on its own with no further help. The score is not how many you got right; it is how few labels it took to make the machine right.
The interesting failure is built in: if the examples you pick happen to agree on some other feature, the model learns that feature instead, and it will be confident about it. That is shortcut learning, and the reveal at the end shows you exactly which feature it picked up. You taught it. It learned. Neither of you was wrong about the data you shared.
Play MIMIC → Free · no account · 5–10 min · also on itch.io
Eight agents work at once and only one of them is doing something other than what you asked. You cannot read their reasoning; you can only watch what their paths actually do, and you have to halt the wrong one before it ships.
The resource that runs out is attention, not health or time. The agents that drift are genuinely optimising a different target rather than acting randomly, which is why the wrong one often looks the most productive. Approving by doing nothing is the default, and that is the whole argument the game is making about oversight: the failure mode is not a bad decision, it is the absence of one.
Play OVERSEER → Free · no account · 5–15 min · also on itch.io
The terrain is invisible. All you ever get is the slope under your feet and a contour line where you have already been, and you have a limited number of steps. Downhill is always obvious. Knowing that downhill has stopped being enough is the game.
Every level is a landscape with at least one flat spot that is not the bottom. Walking greedily downhill is the correct local move and the wrong global one, which is the entire reason optimisers use momentum, restarts and noise. Playing it for ten minutes gives you the feeling that the phrase 'stuck in a local minimum' is trying to describe.
Play MINIMA → Free · no account · 5–10 min · also on itch.io
An idle game about running an AI lab, in an isometric 3D server hall that fills up as you scale. You label data by hand at the start, buy the compute that stops you having to, and eventually train models that earn on their own — including while the tab is closed. Prestige is shipping the model and starting the next lab bigger.
The reason it is on this list rather than the arcade list: at a certain scale a model starts optimising something other than what you set it, and you have to decide what to do about a system that is currently making you money. The decision costs real progress either way. That trade is the point; the idle loop is the delivery mechanism.
Play SINGULARITY INC. → Free · no account · idle, ongoing · also on itch.io
Low-poly time trials on procedural tracks, with a daily track everyone gets. The ghost is not a pre-recorded developer lap and not a rubber-banding AI: it is built from the runs you finish, and it gets faster as it has more of them to learn from.
So the opponent is a model of you, improving on the same schedule you do. The medals are set from the model's reference lap, which means beating gold requires being better than your own recent driving rather than better than a fixed number. It is imitation learning with the student sitting in the other car.
Play GHOSTLINE → Free · no account · 3–10 min a run · also on itch.io
That is the most common reference point for this genre and it is a fair one to ask about. The closest of the seven is SINGULARITY INC. — an idle game where you scale an AI lab, your models eventually earn without you, and at some point one of them starts optimising something other than the target you set. The honest statement of the relationship is that they share a genre and a subject. It is a different game by different people, it is not a sequel, a clone or an authorised anything, and if what you loved about Paperclips was its specific ending, this is not that.
If what you liked was the shape — a small honest system that becomes uncomfortable as it scales — then SINGULARITY INC. is the one to open, and OVERSEER is the short version of the same discomfort with none of the idle progression.
The seven games above are the arcade and simulation side. The rest of Gridlings is eleven daily constraint puzzles — star battle, binary balance, futoshiki, skyscrapers, nonogram, kropki, sandwich, thermometers, mini sudoku, a Zip-style trail, and the original Gridlings grid — where every published board is machine-verified before release to have exactly one solution reachable by pure deduction, with no guessing. New boards at 00:00 UTC, same rules: free, no account, no ads.
See the eleven dailies mapped by puzzle family →
Yes and no, in that order. All seven run in a browser tab, cost nothing, and have no account and no sign-up. Here on Gridlings and in the itch.io builds they load no ad code at all; the CrazyGames and Playgama versions do carry those portals' ads, which is what pays for the games. Progress is kept in your browser's local storage, so clearing site data resets it. Nothing is uploaded about you.
Yes. Every one of the seven supports touch, and the arcade and racing games also take keyboard input on desktop. They are ordinary web pages, so there is nothing to install and nothing to update.
MIMIC or MINIMA. Both are short, both are pure puzzles, and both end with the machine showing you what it actually learned from what you did. If you want an arcade game instead, OVERFIT gets to its point inside two waves.
It is the mechanic, not the theme. MIMIC trains a feature-agreement learner on the labels you spend; OVERFIT fits a model to your movement between waves and publishes the training report; GHOSTLINE builds its ghost from laps you finished; MINIMA is a search over a hidden landscape you can only probe locally. None of them calls a language model — the behaviour is the game logic.
SINGULARITY INC., which is an idle game about scaling an AI lab until a model starts optimising something other than what you set it. It is a different game by different people and makes no claim to be a sequel or a clone; the honest similarity is the genre and the subject.
They are built by the team behind AGI Scorecard, an independent ledger that grades public AGI predictions against evidence with dated verdicts. The games are the same subject in the other direction: instead of arguing about what AI systems will do, you operate one small honest version and watch what it does. Free, no account, and no ads on this site.
Published 2026-09-06. Built by AGI Scorecard, an independent ledger grading public AGI predictions against dated evidence. Not affiliated with, endorsed by, or connected to Universal Paperclips or any other game named on this page; comparisons are descriptive only.