What it does
The verdict is about latency rather than intelligence: Laya won 5–1 on decision volume, while Jev needed a third of the corrections but at around 400 milliseconds was acting on staler information.
Apps & data pipelines
Two System One models fight a real Doom deathmatch — open-weight Laya locally against hosted Jev — from the same compressed state and the same questions.
Only you can see your notes.
Recording unavailable. Open the experiment ↗
The verdict is about latency rather than intelligence: Laya won 5–1 on decision volume, while Jev needed a third of the corrections but at around 400 milliseconds was acting on staler information.
Sketch out the specific tactical choices your game character should make, such as picking a target or deciding whether to retreat. A developer can write code that turns the game surroundings into simple descriptions, then asks an AI to pick actions from that list.
Decide whether to run a model directly on your local computer or over the network. As shown in this experiment, waiting for answers across the internet can make fast-paced game actions out of date compared to running a smaller, faster model locally.