Courtyard of an abandoned housing block, a frame from Linksong

Hostile Colour makes games and the machine learning tools behind them.

Two projects are in progress: a stealth game on Unreal Engine 5, and a language model architecture trained on one consumer GPU.

Projects
Linksong, Unreal Engine 5

Linksong

A stealth game in which you play a golf ball hunted by a robot golfer. The opening cutscene starts in a collapsed housing block.

Linksong, level 1

The island

Sand, shallow water and pines. The first level, where the robot sees a ball in the open from 116 metres.

Linksong, opening cutscene

The courtyard

An abandoned housing block taken over by birches and grass, seen from the rim.

Linksong, level 3

The metro

A dark line with a yellow zone deep inside and a way out. Over 1,900 hand-modelled props.

Linksong, the bar

The bar

Every asset built by hand, down to the unpaid tab. An automated loop checks each shot for broken geometry.

Linksong, opening cutscene

The nest

A summer terrace of concrete and ivy where the balls start.

Machine learning, language models

Heixing

A language model architecture with a cache 24 times smaller than GPT-2 small at one million tokens, trained on one consumer GPU.

Side projects

Fun with Claude

A Blender diorama prepared for 3D printing, a night-city engine in Python and OpenGL, and the Linksong devlog videos.

Linksong

A stealth game on Unreal Engine 5. You play a golf ball, and a robot golfer hunts you across an island course, an underground and a metro. What the robot can see and hear decides every level.

Open the devlog →
Courtyard seen from the rim Island level, shallow water Metro line

A language model that fits on a home GPU.

Heixing is a model architecture that mixes Gated DeltaNet layers with a few attention layers and sparse experts. At one million tokens of context XimO-1 Small keeps a cache of 1.5 GiB. GPT-2 small would need 36.

Read the write-up →
Gated DeltaNet, fixed-size stateAttention, exact recall

See how the work is made.