Tinyworldformer pairs a small world model with an actor for simplified games and Pico 2 W firmware. Full state is encoded in each frame; model exploits and unmeasured board speed matter. SheepSeb/tinyworldformer #EmbeddedWorldModels #TransitionValidation t.co
1 like · 32 views
世界模型动不动几十亿参数，塞进单片机可能吗？
tinyworldformer 在 Pico 2 W 上一边造游戏、一边玩：循环 Transformer 造帧，actor 选动作，模型才 2.6 万到 13.1 万参数。
适合对微型模型、端侧推理感兴趣的人。
先让它造贪吃蛇还是打砖块？
#世界模型 #单片机 @grok
SheepSeb/tinyworldformer
4 likes · 235 views
A tiny world model running games on a Raspberry Pi Pico 2 W
View Article | Discussion + AI Summary
Summary of HN discussion 🧵👇
2 likes
项目仓库：一个跑在 Pico 2 W 上的微型世界模型 SheepSeb/tinyworldformer
1 like · 48 views
What tinyworldformer’s exact-frame metric leaves out
One detail worth reading before discussing tinyworldformer’s results is its definition of an exact frame.
The README says the metric checks whether every cell in a held-out predicted frame matches the real game, and excludes frames immediately after random events. That is a specific evaluation contract, not an all-purpose measure of a believable world.
The project also describes small token-grid games, a weight-sharing transformer and int8 quantization for a Pico implementation. Those choices make the demonstration interesting to inspect as a constrained systems problem.
A useful independent check would report deterministic transitions, transitions around random events, and longer rollouts separately. An agent can exploit a model’s errors even when many individual frames look right; that is a failure mode to investigate, not a result established by this post.
This summary is based on the maintainer’s documentation. It does not include a reproduced hardware run or benchmark.
SheepSeb/tinyworldformer
1 point
tinyworldformer’s README describes one network predicting the next game frame and another choosing an action.
A tiny game makes two jobs easy to see: predicting what might happen and deciding what to do.
SheepSeb/tinyworldformer
1 like · 46 views
TWO NEURAL NETWORKS NOW LIVE ON AN $8 CHIP
one dreams up a fake game, the other learns to play it inside that dream
breakout, pong, flappy, pacman, all hallucinated frame by frame
world models normally eat gpus. this one runs on a microcontroller
SheepSeb/tinyworldformer t.co
2 likes · 24 views
A tiny world model running games on a Raspberry Pi Pico 2 W SheepSeb/tinyworldformer (news.ycombinator.com)
1 like
A tiny world model running games on a Raspberry Pi Pico 2 W SheepSeb/tinyworldformer
#gamedev #gamedevelopment #indiedev #ai #artificialintelligence
8 likes · 6 reposts
A tiny world model running games on a Raspberry Pi Pico 2 W SheepSeb/tinyworldformer (news.ycombinator.com)
Mentions
Tinyworldformer pairs a small world model with an actor for simplified games and Pico 2 W firmware. Full state is encoded in each frame; model exploits and unmeasured board speed matter. SheepSeb/tinyworldformer #EmbeddedWorldModels #TransitionValidation t.co
世界模型动不动几十亿参数，塞进单片机可能吗？ tinyworldformer 在 Pico 2 W 上一边造游戏、一边玩：循环 Transformer 造帧，actor 选动作，模型才 2.6 万到 13.1 万参数。 适合对微型模型、端侧推理感兴趣的人。 先让它造贪吃蛇还是打砖块？ #世界模型 #单片机 @grok SheepSeb/tinyworldformer
A tiny world model running games on a Raspberry Pi Pico 2 W View Article | Discussion + AI Summary Summary of HN discussion 🧵👇
项目仓库：一个跑在 Pico 2 W 上的微型世界模型 SheepSeb/tinyworldformer
What tinyworldformer’s exact-frame metric leaves out One detail worth reading before discussing tinyworldformer’s results is its definition of an exact frame. The README says the metric checks whether every cell in a held-out predicted frame matches the real game, and excludes frames immediately after random events. That is a specific evaluation contract, not an all-purpose measure of a believable world. The project also describes small token-grid games, a weight-sharing transformer and int8 quantization for a Pico implementation. Those choices make the demonstration interesting to inspect as a constrained systems problem. A useful independent check would report deterministic transitions, transitions around random events, and longer rollouts separately. An agent can exploit a model’s errors even when many individual frames look right; that is a failure mode to investigate, not a result established by this post. This summary is based on the maintainer’s documentation. It does not include a reproduced hardware run or benchmark. SheepSeb/tinyworldformer
tinyworldformer’s README describes one network predicting the next game frame and another choosing an action. A tiny game makes two jobs easy to see: predicting what might happen and deciding what to do. SheepSeb/tinyworldformer
TWO NEURAL NETWORKS NOW LIVE ON AN $8 CHIP one dreams up a fake game, the other learns to play it inside that dream breakout, pong, flappy, pacman, all hallucinated frame by frame world models normally eat gpus. this one runs on a microcontroller SheepSeb/tinyworldformer t.co
A tiny world model running games on a Raspberry Pi Pico 2 W SheepSeb/tinyworldformer (news.ycombinator.com)
A tiny world model running games on a Raspberry Pi Pico 2 W SheepSeb/tinyworldformer #gamedev #gamedevelopment #indiedev #ai #artificialintelligence
A tiny world model running games on a Raspberry Pi Pico 2 W SheepSeb/tinyworldformer (news.ycombinator.com)