Remember your first day at a new job?
You walk in with years of experience. You know how to work, how to talk to people, how to find your way around a building. But not this building.
Someone points at a floor plan by the entrance. You can see where the meeting room is. What you can't see is how to get there. So you take the wrong corridor, hit a dead end, turn around and try again. Nobody has time to give you a tour.
And then, somewhere around lunchtime, it clicks. You stop getting lost. Everything you knew before finally works here too.
That's exactly what we did with Scout
We took an AI model that learned to walk in Counter-Strike 2 and dropped it into our office. Not the real office. A 3D replica, built from our floor plans: 35.8 by 14.1 metres of rooms, corridors, doors and furniture. Scout had never seen any of it.
On its first attempts, it was lost. Out of 96 test routes, it reached the target on just 2. Then it practiced on its own for 57 minutes, with no new human examples. On the same test, it reached 44.
What Scout works with
Scout's task is simple to describe: get from point A to point B.
It gets two things to work with:
- What it sees. A small first-person view, 256 × 144 pixels.
- A simple map. 128 × 128 pixels, with only the start and the goal marked.
Just like the floor plan by the entrance, the map has no "you are here" dot and no route. Scout has to remember what it has already seen and work out where to go. About 16 times a second, it decides how to move and where to look.
Behind Scout is a neural network we built ourselves, with about 25.3 million adjustable settings (parameters). When we started, those settings were random. Scout knew nothing.
Its previous job: Counter-Strike 2
Counter-Strike 2 players move through complex maps fast, with a clear goal and under pressure. That is exactly the kind of human behavior we study. It's also where Scout got its experience, in two steps.
First, it watched. Scout learned by copying real players:
- 61,782 short clips of players moving through maps
- about 119.5 hours of movement in total
- on 5 maps: Dust2, Mirage, Inferno, Ancient and Nuke
Then, it practiced. In a 3D simulator, Scout tried routes on its own and earned a reward for getting closer to the goal and for finishing the route. This is called reinforcement learning: trial, error, and a reward when it works. We added two more maps at this stage, Anubis and Train. Across all seven maps, Scout made about 200 million moves.
The result: Scout finished 51.6% of its test routes across the seven maps.
Distance makes a big difference. On the shortest routes, Scout succeeds 85% of the time. On the longest ones, only about 5%. Long routes are still the hardest problem to solve.
First day in the office
Our office was never part of Scout's training. When we first dropped it in, straight from Counter-Strike, it finished just 2 of 96 routes. What it learned in the game didn't carry over on its own.
So we let it adapt. Unlike most new hires, Scout got no tour. It spent about 57 minutes practicing inside the office replica, making 4.4 million moves and learning only from its own attempts. No human walked those corridors to show it the way.
| Office test (96 routes) | Before practice | After practice |
|---|---|---|
| Routes completed | 2 | 44 |
| Success rate | 2.1% | 45.8% |
That's 22 times more routes completed than before practice. To keep the test fair, the goals of those 96 routes were never used during practice. Scout had to find its way to places it hadn't trained on.
40 hours vs. 57 minutes
Teaching Scout Counter-Strike took about 40.5 hours, from the first copied move to the end of practice in the simulator. Adapting it to the office took about 57 minutes. That's roughly 42 to 1.
To be clear: these are two runs we actually did. We haven't tried teaching Scout the office from scratch, so this doesn't prove that adapting is 42 times faster than starting over.
What it does show is a recipe that works. First learn from game data. Then adapt to a different environment.
What Scout can't do yet
- It lives in a simulation. The office is a simplified 3D replica. There is no robot walking through our real office.
- It still misses more often than it succeeds. 45.8% means more than half of the office routes were not completed.
- Long routes are hard. Even in Counter-Strike, success drops sharply the longer the route gets.
- The simulator helps while it learns. Scout never sees its exact position when it decides where to go. But during practice, the simulator uses that position to decide when Scout has earned a reward.
Why a game?
We are building a frontier AI lab to model and predict how humans react and make decisions under pressure.
Counter-Strike 2 is full of people moving with purpose, fast, while something is at stake. Scout shows that what a model learns there doesn't have to stay there. Like anyone on their first day, it walked into a building it had never seen. Within an hour, it was starting to find its way around.