- In danger
- Sliders moved
- Goal
Show the neural network behind the sliders
Each connection and small dot is one slider.Blue is a positive weight, orange a negative one.
How the scoring works
Each slider is one weight of a tiny neural network. Your score on a level is how many sliders end up away from where they started, however often you moved them. The goal, the fewest possible, is proven: a search through every smaller change on the 0.5 grid finds no fix.
The models get the same weights and points as plain numbers, with no tools available, at high effort. They get three tries and are told which points are still wrong, but not what the goal is. You get as many tries as you like and a live map, so this is not a fair fight and not a benchmark. Every model attempt is in runs.json, and the code is on GitHub.
Under the story, a human is a point labelled 1 and a zombie a point labelled 0, and daylight is where the network's output is above 0. The game takes its idea from Sparks of Weight-Space Programming. These networks have at most 17 weights, so a good score here says nothing about editing real models. For the argument over that work, read the critique and Simon Willison's replication.