Aleph Alpha 新模型 Kolibri 玩《Doom》
Aleph Alpha's latest Model playing Doom
Aleph Alpha 的 Kolibri-1 模型正在直播玩《Doom》，由 Tesseracted Labs 搭建共享游戏环境，每轮结束后轮换两张地图。模型只接收结构化游戏状态而非图像，通过四次仅基于概率的评估分别选择移动、瞄准、开火和交互，再由本地控制器执行动作。页面显示的是选项标签 logits 归一化后的分数，并非校准过的成功概率。
Less talk.
More Doom.
Kolibri-1 chooses. The game responds.
One shared game, playing live for everyone.
Two maps rotate after each round.
CONNECTINGROUND —
HEALTH—
KILLS—
DECISION—
WATCHING*—
Waiting for the shared game…
Inside the decision
Four probability-only model evaluations choose movement, aim, firing and interaction. Unavailable targets are excluded before scoring. A local controller executes the chosen actions. Kolibri receives structured game state, not images.
How it works · live state and output
No generated text. We score option-label logits and normalize them over the supplied choices. These scores are not calibrated probabilities of success. The controller handles persistent exploration targets, wall clearance, steering and bounded corner recovery; failed pickup targets are temporarily excluded. These mechanics are not learned by this demo. The tactical choices still come from Kolibri.
Waiting for a model decision…
来源：Hacker News · tesseracted.com