NanoJev 是一个 0.6B 参数的并行决策模型：输入游戏状态与候选动作，一次 Qwen3-0.6B 骨干前向就直接输出完整概率分布， 零输出 token 解码。本文拆解它的核心三件套——Qwen3-0.6B 骨干、集合注意力头与标量决策头，看看它如何一次前向完成批量决策，并用同一检查点打赢迷宫、贪吃蛇和两款射击游戏。

A collection of tools and algorithms for developing traditional roguelikes. Such as field-of-view, pathfinding, and a tile-based terminal emulator.

This time, we will create a "maze-like field" in Minecraft Education that changes its shape every time it is run. Once completed, glass walls will be randomly placed within a 15x15 field. The start is ...

- Select `worlds/maze_world.wbt`.

Following the explosive adoption of generative AI, the next massive wave of technology attracting intense attention from ...

In-person attendance is by invitation only. The event will stream live on this page.