Hello World

This post should have been written 2 years ago. Better late than never.

For the last 2 years, GameDiffuser has been working on simulating games via neural networks. This approach has many names; real-time interactive AI video generation, neural game engine, world model… It is simply an AI model that produces new frames by taking in player inputs such as button presses and keeping track of the game history.

This may sound exactly like how we make games today already, because it is! If you make a game today, using existing tech like vulkan to render your triangular meshes, you will receive player inputs the same way but keep track of the state of the game differently. You will have an abstract representation of the state, mostly independent of the pixels on the screen, that will keep getting updated with the player inputs. Each frame you will convert this abstract representation to pixels on the screen, which is called rendering.

What’s wrong with the current game tech?

Why not simply make new games with Unreal Engine? GameDiffuser bets that we are reaching the limits of the traditional game development pipeline. Red Dead Redemption 2 remains an extraordinary feat of art and engineering, but it was released almost 8 years ago. Since then, team sizes, budgets, production timelines, and technical complexity have all grown dramatically. Yet the player experience has not improved at the same pace. In many cases, the added complexity seems to slow teams down, increase risk, and make genuinely novel games harder to build.

We want to be excited again by the new releases, each better than the last, like the games by Carmack. In the early 2000s, each new game did something new, novel and rose to new heights. Nowadays we observe the opposite.

GameDiffuser is our long term bet for full stack AI generated games.

Axioms

1. People will keep playing games, probably even more than today.
2. AI will scale.
3. Current game development pipeline will not scale as much, if at all.

The team is currently just a single person, Tarık Kaya. After early attempts to fund the company through VC, Tarık shifted the project toward academic research, where the core technical problems can be explored more deeply. This has turned out to be a feasible approach, because the technology is still new and needs time to mature to the point where it actually provides fun and engaging experiences for the players. There are a lot of challenges to be solved on all fronts, especially on the neural network side.

We will release follow-up posts to present; what has been explored, what is being worked on and how GameDiffuser’s approach is different.

See you on the other side!