In 2010 I built 1378(km) as a mod for Half-Life 2. Today the engine is free and an AI drafts a world from one sentence. A look at what changed in between.
Picture a screen in 2010. Four viewports, three of them flat grids, one a 3D camera. Walls are boxes called brushes. You press compile and wait while the map is calculated. Then you load it, walk around, and find a gap in a wall. Back to the grid.
That is the Hammer editor, the level tool of Valve's Source Engine. It is where 1378(km) was built.
For 1378(km), that loop never ended. Build, compile, test, build again.
In the end it took plenty of workarounds to make it run. Again and again the editor asked for creative problem solving and lateral thinking.

A mod, because that door was open
1378(km) was my intermediate diploma project at HfG Karlsruhe in 2010. It is a multiplayer game set at the inner-German border for up to 16 players, who play either as refugees or as GDR border guards. Technically, it is a modification of Half-Life 2 on the Source Engine.
Why a mod? The surprising part: on paper, there were alternatives. Epic released the Unreal Development Kit in November 2009, free for non-commercial use. A commercial project paid $99 up front and a 25 percent royalty on revenue above $5,000. Unity had made its basic version free in October 2009.
So the engines existed. For anyone outside a studio, they were also brand new. Unity's own focus at the time was mobile and the web.
Source had a head start. Valve released the Source SDK tools on Steam on 5 November 2004, eleven days before Half-Life 2 itself. By 2010 that meant six years of tutorials, forum threads, sample code and finished mods. The engine came with a working client-server multiplayer framework, first-person movement, physics and weapons. You started with a running shooter and changed it from there.
My reasons were practical. The young engines had little content, and building something this complex from scratch as a solo developer was hard. There were no templates or marketplaces yet. A mod was simply the easier way.
Building inside someone else's game meant building with its parts. There were many limits, above all the models you could use from the base game. I made my own, but that ate up a lot of time in other tools. Much of the world was built directly in Hammer from brushes. And a lot from the engine library got repurposed wherever it fit, even if it ended up the wrong way round.
The engine became the free part
A few years later the economics flipped. In early 2014, Epic offered Unreal Engine 4 as a subscription for $19 a month and put the C++ source code online for every subscriber. On 2 March 2015, Tim Sweeney announced that the engine was free for everyone. His blog post was titled "If You Love Something, Set It Free."
Unreal Engine 5 followed in April 2022. It brought Nanite, a virtualised geometry system that renders extremely detailed assets in real time, and Lumen, fully dynamic global illumination. In gamer terms: less time baking lightmaps, more time playing with light.
Today the engine is free for individuals and small companies. A game pays a 5 percent royalty only on lifetime gross revenue above $1 million. In 2009, it was 25 percent above $5,000.
An independent artist in 2026 does not need a host game anymore.
What I build with it now
My current work runs on Unreal Engine 5. POSOMAT Magic Move is an interaction layer for UE5. A stereoscopic AI depth camera tracks several people at once, without controllers or wearables. Their skeleton coordinates flow straight into Niagara particle systems and physics. Everything runs locally on a single computer with NVIDIA CUDA graphics, and nothing goes to the cloud. BASHBALL, a game for LED walls, is built on top of it.

And now the world builds itself
The next shift is happening outside the engine.
In August 2025, Google DeepMind announced Genie 3. Given a text prompt, it generates dynamic worlds you can navigate in real time at 24 frames per second, consistent for a few minutes, at 720p. Its developers say there is no underlying game engine. A neural network predicts every single pixel in reaction to the user's input and to what came before.
In May 2026, Google connected its Project Genie prototype to Street View imagery for places in the U.S. Google's own example: explore what the Fort Worth Stockyards in Texas might have looked like in the 1920s, in a "B&W film" style. In October 2026, DeepMind began rolling out Genie 4 as a research preview for trusted testers. According to the announcement, photojournalists are exploring how a single 2D archival photograph can become a grounded simulation of that environment in seconds. Other teams are testing how such worlds can help people experience historical spaces.
World Labs took a different route with Marble in November 2025. It generates worlds from text, images or video, and they can be exported as meshes or Gaussian splats and pulled into an engine.
The tool barrier of 2010 has largely disappeared. Today one sentence gets you a first draft of a world.
Who decides what is left out?
For artists, that sounds like pure liberation. More people can build worlds. Ideas can be tested in seconds instead of semesters.
But a historical place is more than how it looks. Every reconstruction leaves things out. In 2010, if something was missing from a level, a person had decided to leave it out, or had simply run out of time. When a world model rebuilds a place from a prompt or an old photograph, it fills the gaps with what it has seen most often.

So who decides what such a model leaves out? And would anyone walking through that world even notice that something is missing?
I do not have an answer yet. In 2010, the hard part was getting the engine to do what I wanted. In 2026, the harder part might be knowing what it decided on its own.
