I read 321 Touch Grass entries. None of them could referee a game of Simon Says. So I built one that can.
This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass Playground games rarely end because people get bored. They end when one person checks a phone, then another, and the circle on the grass turns into a row of people looking down. I wanted to give the phone the one job in Simon Says that nobody wants, the referee, so everyone else can stay in the game. Before…
The author of this submission for the Hacktoberfest Open-Source AI Challenge Week 1 sought to create a group version of the game Simon Says that could referee the game live, keeping everyone engaged without the distraction of checking phones. After reviewing 321 Week 1 entries and 130 of them being outdoor quests or games, the author found no solution that provided live monitoring of a group playing together.
The built game, named Simon Sees, functions like Simon Says outdoors, with a phone propped up on a stump or backpack, tracking the players. The game starts with everyone lining up and the phone displaying a numbered photo card for each participant, which can be named verbally or typed. After a brief pause, the phone announces a command, such as "Simon says stand on one foot." Participants then perform the action while the phone streams a few frames per second to a laptop.
The phone employs a Qwen3.5-4B I fine-tuned for the specific task of determining if a person is following the command, or Liquid's d1 decision model if the hardware constraints prevent the use of the larger model. The referee assesses each player's cropped image, answering a yes/no question about whether the individual is executing the command correctly. The game concludes when only one player remains standing.
The game is designed for any user with a phone, laptop, and some space – be it a family at a picnic, children at a birthday party, or a group at the park who have exhausted their activities. The demo video showcases a replay with AI-generated sample photos, running on a 2-vCPU, no-GPU setup, and is sped up three times to fit the demonstration within a minute. The real models, running on a similar hardware configuration, are described as providing accurate results.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
