Kittie Has Other Plans

Can AI play as Kittie?

I made a Game Boy game for my wife's plushie as an anniversary gift. Then I gave AI models the controls.

What impressed me was watching them play in real time. The game kept running while each model decided what to do next.

Kittie resting his paws on a ModRetro Chromatic displaying the cover of Kittie Has Other Plans
Kittie with his game.
Two successful Courtyard runs at original speed, selected near each model's median winning time. Jev completed 23/50 attempts; OpenAI completed 12/50.

Courtyard

Both models received structured game state without screenshots. Jev completed more attempts and looked more decisive.

Read the experiment →
ModelFinishedAPI cost / attempt
Jev23/50$0.0093
OpenAI Decisions12/50$0.0414

Head Goon

I added screenshots for OpenAI during the boss fight. Its observed completion rate rose, while Jev still completed more attempts.

Read the experiment →
SetupFinished
Jev11/30
OpenAI text1/30
OpenAI + image6/30

Teaching a small language model to play

Can demonstrations and reinforcement learning teach a 0.8B model to play Head Goon? I trained Qwen with behavior cloning, then PPO.

Watch the gameplay and read the study →

Original encounter · 50 games per policy
PolicyWins
Pretrained Qwen0/50
Behavior cloning39/50
PPO1547/50

A separate study from the earlier API comparison. Randomized scenarios and limitations are reported in the article.

Methods and data

How the models receive game state, choose controls and play in real time. The methods explain the timing, costs and limits of these exploratory comparisons.

Download results: Courtyard · 100 attemptsHead Goon · 90 attempts