flappy

Documentation

Overview

Flappy Bird played by a language model, to put a number on a question: what can a model decide at a reflex game when each step is one scored question? Nothing is trained. The model reads the state as a sentence and answers through Engine.Score, so the answer is a probability and the cost is one prefill per decision; a random flapper and a one-line heuristic play the same game as the floor and the ceiling. Asked yes or no, no model plays; asked which of two numbers is larger, with the numbers as the options, a 1B plays the heuristic's game.

go run ./_example/flappy                    # Qwen2.5-0.5B-Instruct from the cache
go run ./_example/flappy -model ./x.gguf    # another model
go run ./_example/flappy -larger -rows      # the comparison asked as a choice of numbers
go run ./_example/flappy -episodes 5 -show  # print each decision
go run ./_example/flappy -larger -nobase -screen   # watch it in the terminal