StarCraft AI learns from both wins and losses — still can't beat humans

Researchers built a language-model AI system that plays StarCraft II by learning from expert game recordings and its own past mistakes. The system beat the game's hardest computer opponent, but the paper provides no evidence it outperforms existing AI players or that this method works outside this specific game.

This is a narrow capability demonstration in a controlled lab setting. The AI learned to reflect on past games and improve, which sounds interesting until you ask: does it actually learn faster or better than simpler reinforcement learning methods already used for RTS games? The paper doesn't say. The practical implication is zero until someone shows this 'reflection' approach actually solves a real problem that existing methods don't, or does it more cheaply. Right now it's a technique that works on one game in one lab.