Twitxy Saved

GeekPark

Verified account This account is verified. Learn more

@GeekParkHQ

We bet on tech founders, report and invest, then bring China's AI insider stories to the world.
First angel in @UnitreeRobotics.
Run by our reporter in SF.
San Francisco, CA about.geekpark.net/blog/ Joined February 2026
755 Following 1.2K Followers 1.2K Posts

Latest posts 5

GeekPark Verified account @GeekParkHQ
Wait, they made this in five days? We are more curious about how you tell an AI a sword hit doesn’t feel right.
Grummz Verified account @Grummz 20h
Seele is a company that creates AI tools for game making. Recently they invited me to participate as they took on a 5-day challenge to push their new model, Seele03, to its limits and move closer and closer to AAA with a Wuxia-style action game.

The product consists of their AI model, asset generators, and target engines such as web (three.js), Unity, or even Unreal. What's nice is that their site has everything in one place, from mesh generators and rigging to sound fx and sprite work. They also provide common templates for most types of games. Having it all work together seamlessly saves a lot of time. You can get a game running very quickly, in 20 minutes or so. (20 minutes! I remember when this took months and months.)

So what can they do in 5 days?

Their progress video, below, shows them iterating through creating a Phantom Blade Zero-style game. This is not reverse engineering, but made from scratch. Over the course of development, they sent me video updates on their progress. The 03 model is designed for iterative development, and I gave them some advice as they took it from basically nothing to the final version of the video today. The progress the model made day to day was rapid, and the traces actually factored in something that I don't see that often: game design.

Most models focus on the graphics when making a game, but Seele03 is thinking about combos, hit feedback, everything that really matters for how a game feels and plays. One day, I think that's going to be the core of what makes AI game making viable.

After you are done prompting and refining your game, it's not frozen on the web. You can download and edit the project files themselves. Since you can download the result and open it in Godot or Unity (Unreal is embedded web atm), Seele03 can give you a quality core and a way for you to take things even further.

So how did it turn out? I've tried a lot of game generators, and Seele03's all-in-one tool approach and engine targeting stand out for me. Most others only generate three.js for web, or are more world models based on generative rendering. This is the rare one I've seen that is making games directly in game engines.

They've asked me to share the full process video with you so you can judge for yourself. As I watched the team work, I sat back and wondered at how it's all changing so fast: you simply... chat... interacting with a UI that is purely text, typing in wishes and directions, and everything else unfolds around you. I think they got pretty far in their goal gameplay-wise, which is where they really focused. The key thing is that after you have the final result, you also have the files to download, and you can replace anything you want to take it even further: animations, level graphics, basically anything.

If you want to make games for real engines on PC and console, then give Seele a shot at seeles.ai. It's faster than vibing a game in Claude because of their game-design-driven model, their built-in asset generators, and the fact that you are not burning tokens re-creating an entire engine before you even get to your game. And every day they get closer to AAA... it's just a matter of time.

@SEELE_AI1125
View quoted post on X
GeekPark Verified account @GeekParkHQ
Put two interviews with MiniMax (@MiniMax_AI)'s Yan Junjie (IO) side by side. You will get a pretty good snapshot of how AI has changed in 2 years.

In 2024, he told GeekPark (@GeekParkHQ) that pushing the frontier was the hard part. Cutting costs by 10x? There was already a playbook for that. He even predicted a nearly 100x drop in two years.

Fast forward to his recent podcast with Lawrence Burns from Scottish Mortgage (@ScotMortgage_UK). Now he calls low inference costs MiniMax's moat.

Why? Because cheaper inference lets you scale RL and push the frontier further.

How the thing that used to come after a breakthrough is now part of how you get to the next one.