AI labs are building "world models" — AIs that understand how 3D worlds work. To learn, they need recordings of real people playing games, with the button presses included. Lots of companies have gameplay video. Almost nobody can do what we're building: find out exactly what a researcher is missing, send players to create precisely that, and prove it worked. That's a data engine. And it's step one of a bigger plan: the same engine, proven in games, is how we capture the data the internet never had — the real world. This page explains it with no jargon, then shows where every one of us fits.
Selling game data is the wedge, not the company. Everything the engine learns to do — find what's missing, steer real people to capture it, prove the delta — is practice for the same loop run on the physical world.
The internet was a one-time subsidy for AI, and it's drying up. Labs are on a trajectory toward $100B+/year of data spend, and the scarcest kind isn't text — it's the dark matter: embodied, tacit, physical-world experience that was never digitized at all. Whoever learns to collect it, on purpose, owns the bottleneck of the decade.
Games are the one place the loop is verifiable: ground truth exists, so we can measure exactly what our data improved and correct the error on the delta. And the field is converging on multi-view learning — ego + exocentric + actions, aligned — which is precisely what synchronized gameplay produces at internet scale, and almost nothing else does. The 3D spaces are common; the input side is what transfers.
The same engine, pointed at reality: guidance for humans in physical space — responders, industrial crews, accessibility, humans moving among robots. The path runs through the overlay itself: today it is deterministic (scripted quests, fixed prompts); every gamer reaction to it is a human-in-the-loop label that tunes it into a generative overlay; once it generalizes across game environments, it rolls out to the real world. Overlay today, wearables next.
Graded honestly, by our own standard: this layer is a bet — orange, not green, and the series says so. That's fine: missions are bets. But it's the bet that makes the green parts one company instead of a data shop, and it's why "sell gameplay" alone was never the plan.
The data we sell is the data we need.
Most companies in this space are warehouses: mountains of gameplay clips collected by accident — whatever players happened to record. If a researcher needs something specific, tough luck. You get what's on the shelf.
We work from an order book. A researcher says "I need footage of players building structures and looking away from them" — and our quests send real players to do exactly that, this week. Then a taste-test proves the dish was right.
The same play sessions build the spatial wiki: a living, social 3D map of game worlds where every achievement carries the trace that reproduces it. Its core scope is the bridge from isolated diaries — individual trajectories, one player's run — to a collective synchronization of knowledge and actions: the collective brain of gamers, built automatically and redistributed to gamers. That's the consumer product players actually want, which means trust we can't buy, distribution we don't pay for, and supply at hobby economics. Selling gameplay alone is a poor business; the wiki is how we win it.
Round and round. Every lap makes the next lap smarter. That loop — not any single dataset — is the company.
A researcher (or our AI stand-in for one, "Dr. R") — or the engine's own coverage map — tells us what's missing: a game, a place, a behavior. Buyers' orders are one input; the map is the standing order.
That gap becomes a quest. Players play it their own way — the variety in how they do it is the value.
We test an AI model before and after our data, on tests we wrote down in advance. No fudging.
The result updates our map of what's still missing — which writes the next quest. Repeat.
Why "prove it worked" matters so much: our entire pitch to buyers is one sentence — "our numbers are real." The tests are locked before we collect a single hour, and we publish results even when they're not flattering. The first time anyone fudges a number, the whole thing is worthless. That honesty isn't a nice-to-have; it is the product.
The map of what's missing lives client-side too: the overlay reacts during play, nudging capture toward the gaps the moment they're reachable (the Tesla trigger pattern from the architecture doc). Distribution-aware, runtime-steered capture — knowing the shape of what we have, and steering live play toward what we lack — is what makes this an engine and not a marketplace with extra steps. It's also the guidance product, prototyped: steering capture and guiding a human are the same model — and every player reaction to the overlay (followed, ignored, corrected) is a label, so the fleet doubles as the human-in-the-loop labeling engine that tunes the deterministic overlay into the generative one.
| Who | What they do | Why the engine beats it |
|---|---|---|
| Scrapers (labs hoovering YouTube/Twitch) | Take millions of hours of video without asking. No button presses, no permission, growing legal trouble. | We ask. Every session is consented and paid, with the inputs included — the part you can't scrape and the part models actually need. |
| Clip archives (Medal / General Intuition) | Billions of highlight clips players happened to record. Impressive pile — but you can't order what's missing, and they deliberately don't keep real button presses. | We're directable. A pile answers "what do you have?" An engine answers "what do you need?" — that's the question buyers are starting to ask. |
| Marketplaces (Origin Lab, Worldmodeldata) | Broker data from game studios' engines. Rights-clean, but no players — no human behavior, no intent, nothing directed. | We have the humans. Real decisions, real mistakes, real voices explaining what they're trying to do — the layer engines can't export. |
| Closed labs (everyone above, really) | Whatever they have, they hoard. Nothing to inspect, nothing to build on. | We're the open one. "We are Hugging Face; they are OpenAI." Open datasets become papers; papers bring the whole field to our door. |
Today we are small: a pilot fleet, a modest archive, big plans. The engine wins not by being biggest but by being the only one that can take an order and prove the delivery. First proof: a tiny, rigorous Minecraft experiment with a researcher who asked us for exactly this.
Game titles and publishers have rights, and we checked all of them — line by line. We only record where it's allowed, we never sell what we can't, and our long game is making publishers partners who profit, not opponents. Boring, deliberate, on purpose.
Three bands: the people who decide what to make, the people who build the machine, and the people who make the play real. If your part stalls, the loop stalls — that's what "engine" means.
The ASK and CHECK steps
Brings the researchers and buyers whose orders feed the engine — labs, publisher AI teams, the Roblox/FACEIT/GDM threads. Owns the story we tell them.
Owns the taste test: the benchmarks that decide whether our data actually improved a model. His "kept" standard becomes the quality bar every hour is measured against.
Makes sure the engine never treats players as data labor: rewards, missions, community — the reason people want to play the quests, and stay.
The researcher who asked for exactly what we're building ("red-teaming for world models"). The POC is built to his suggestion and returns to him with proof.
Our standing AI researcher-persona. Every demo, dataset, and test plan passes Dr. R's review before we spend money — so the first real reviewer we meet is never the first reviewer we've met.
The plumbing between PLAY and CHECK
Everything from a player's session to a buyer-ready file: capture, sync, upload, the whole spine. The engine's horsepower is his code.
The surface players see (overlay, quest UI) and the AI delegation loop that lets agents do weeks of research and drafting in days — the reason this whole program moves at this speed.
The engine's honesty machinery: which sessions count, the paper trail from every clip back to a consenting adult on a cleared game, and the pipelines that make it repeatable.
Designs what players are actually asked to do (MultiPOV, task structure) and keeps the future robotics option honest — priors, not promises.
The PLAY step is human beings, on purpose
The humans closest to the humans: watching how Pioneers really use the app, catching what telemetry can't, and being the face of support. When quests annoy people, they know first.
Find and vet the players, and run the server machinery (bots, roles, tickets, brand) that makes a scattered community feel like a place. Fleet quality starts at recruiting.
The reason playing for Playroll feels like belonging, not gig work — moderation, events, tournaments, partner creators. The spatial-wiki future is theirs to make loved.
The first players whose real, consented, paid play becomes the data. Every quest, every reward rule, every privacy promise exists to be worthy of them.
Research agents, audit agents, build agents — eight briefs dispatched or returned so far (the rights audit alone read 19 legal documents in an afternoon). Humans decide; agents grind.
"Players map the worlds they love. Researchers train on what players build. Everything stays open, and everyone can see."
— the vision line. The engine is how we earn the right to say it.