I have been asked to write this post several times by people who knew us in early 2023, when M10 was four people in a ground-floor space on Rivington Street and the gap between what we were building and what we could actually show anyone was roughly the size of the Thames. I kept putting it off because honest founding stories are harder to write than the polished retrospective versions, and I wanted this to be the honest one.
So here is what the first six months actually looked like.
The Space
The Rivington Street space is the kind of room that photographs well and works less well. The exposed brick is real. The natural light from the street-facing windows is genuinely good in the mornings. The heating system was unreliable until March. We had four workstations, a whiteboard that doubled as a project wall, and a server rack in the corner that Lena had configured across three nights. The rent was manageable because the building is not a tech campus and the landlord had no idea what going rates were for this particular street.
We chose Shoreditch partly because of proximity. Kofi was already living in Hackney. Yuki was coming from an old job in Barbican. Lena and I could get there in under thirty minutes from where we were. The area also has a density of game studios and interactive media companies that made it easy to find people to talk to when we needed advice we did not have internally. That turned out to matter more than we expected.
The First Month: What We Were Actually Doing
The first month was almost entirely preproduction, which is a clean word for a messy process. We spent the first two weeks arguing about scope. Not unpleasantly, but seriously. The game we had all agreed on in principle, a competitive shooter with adaptive AI opponents, turns out to have about fifteen different reasonable interpretations once you start trying to define what "adaptive" means with engineering precision.
Kofi's version was technically rigorous and probably two years of work for a team of our size. My version was product-forward and probably insufficiently rigorous on the AI side. Yuki's version was gameplay-first and had features we could not build with four people. Lena's version was infrastructure-forward, which was right but made the other versions look premature by comparison.
We resolved it by doing what we should have done on day one: writing a one-page design constraint document that defined what we were NOT building. Not a full game. Not a competitive title ready for external players. Not a polished AI system. What we were building was a proof of concept for one core mechanic: an opponent that reads your positioning habits over the course of a match and adjusts its approach accordingly. Everything else was secondary to answering whether that mechanic was worth building.
Months Two and Three: The Prototype
Once we had the constraint document, things moved faster. Lena set up the basic networked architecture in about three weeks, which was faster than any of us expected. The prototype we were building did not need full networked multiplayer, but the architecture decisions made at that stage would propagate forward, so we built it right from the beginning rather than cheaply. That was the right call.
Kofi started on the AI side from week five. The first version was embarrassingly simple by the standards of what we are building now, but it was functional: it tracked the last twelve positions the player had occupied and used that history to inform angle selection. In a prototype environment with a single player and a single AI opponent, it produced something that felt immediately different from scripted AI. The first time I played against it and it pre-emptively held an angle I had been exploiting, I felt the thing we were trying to build. That was about six weeks in.
Yuki spent those months building the test environment and the feedback infrastructure for gathering observation notes during internal sessions. We did not have external players yet. The sessions were the four of us playing and watching each other play and writing down what we noticed. The volume of notes from that period is extraordinary. Most of them are not useful now, but they built the vocabulary we still use to talk about player experience.
Months Four Through Six: The Hard Part
The hard part was not the technical work. The hard part was the gap between what we could demonstrate and what we were trying to explain to people outside the room. We brought in a small group of competitive players for informal sessions in month four. Their reactions were positive but cautious. They could feel that something was different about the opponents. They could not articulate why. And when we tried to explain the AI mechanics to them, their eyes glazed over in a way that told us we were describing the system and not the experience.
That was a significant realization. The thing we were building was a feeling, not a feature set. We could talk about adaptive behavior trees and observation windows and counter-strategy selection all we wanted. None of that translated until a player sat down and spent forty minutes with the opponent and started noticing, on their own, that it had learned their habits. Getting people to that moment of recognition, without prompting, became the design problem we spent most of months four through six trying to solve.
We changed the timing of the AI's first visible adaptation based on those sessions. Originally it would adapt within the first few rounds of a match. We pushed it back. We wanted players to establish habits before the adaptation kicked in, so that the moment of recognition would be grounded in something they could retrospectively identify as their own behavior. The change made the adaptation feel more earned, both from the AI's perspective and from the player's.
Where We Were at Six Months
At the six-month mark we had: a networked prototype that worked reliably for one-versus-one sessions in a single test environment, an AI system with the core adaptation mechanic functioning as designed, an observation protocol that let us gather consistent feedback across sessions, and a team that had learned to work together under low-resource conditions without anyone quitting or having a serious falling out.
What we did not have: any external validation beyond a handful of informal sessions, any sense of timeline for a playable build, any certainty that the market we were building for was large enough to sustain a studio. Those were all still open questions. They remain partially open today. But the six months in that room, heating unreliable and all, gave us the conviction that the thing we were building was real. That turned out to be enough to keep going.