Open-source decision models sold their own base in Red Alert 2 — then learned until no human could beat them

heathercmiller · x · 2026-10-27

A team gave open-source decision models two RTS games with rough results: in Red Alert 2 the base model sold its own base and lost; in Tank War v1 spun in circles while v2 found the fire button and shot itself.

They then ran a play → train → upgrade harness → repeat loop; after a few rounds the models could evolve without human intervention. Put into an arena against each other, no human has beaten the trained models so far.

Their takeaway: a decision model isn't finished at release—real environments keep throwing new situations at it, and frozen models can't adapt; continuous learning is required. The end-to-end training pipeline is open source, and they plan to turn it into training infrastructure for any decision model. You can play against their best model online.

Original post →

Open-source decision models sold their own base in Red Alert 2 — then learned until no human could beat them

More from Fun