When AlphaStar first went up against pro StarCraft II players like TLO and MaNa, it didn’t just win—it crushed them. Perfect micro. Instant reactions. Stalkers blinking with machine precision. But something felt off. It never scouted properly. It didn’t adjust when the unexpected happened. And when it was forced to use normal human camera controls? It lost.
That match left me with a burning question—was AlphaStar actually intelligent, or just a machine executing commands faster than any human ever could?
A recent study I found interesting suggests there might be another way.
Enter TextStarCraft II, a testbed environment designed to see if AI can win by thinking, not just reacting. Instead of controlling units directly, Large Language Models (LLMs) like ChatGPT play StarCraft II by reading text-based game updates and making strategic decisions—powered by something called Chain of Summarization (CoS), which helps AI think strategically rather than reactively.
ChatGPT interacting with the game through TextStarCraftII and CoS
Ok, so imagine you’re playing SC2 blindfolded. Instead of seeing the map, you have a teammate feeding you info: “We have 50 Marines, the enemy has Mutalisks, and our economy is strong.” That’s how TextStarCraft II works. The LLM doesn’t micromanage—it thinks.
What do we have? (Minerals, gas, army size)
What’s the enemy doing? (Unit compositions, tech choices)
What’s the move? (“Enemy has air units—build anti-air.”)
ChatGPT reacting to Roach-Ravager attacks with defence
Instead of drowning in raw game data, the AI uses CoS to filter out the noise, keeping only the most important details. Unlike AlphaStar, which analyzes every frame in isolation, LLMs build a running summary of the game, constantly refining its understanding and adapting as new information comes in. It’s not reacting—it’s strategizing, adjusting on the fly. AlphaStar, on the other hand, is a speedrunner, optimized to execute the perfect moves at breakneck speed. But throw it into an unfamiliar meta, and it struggles—because it doesn’t actually think.
LLMs, on the other hand, are strategists. Like chess players, they assess, adapt, and adjust mid-game. LLMs don’t have superhuman APM, but they see the bigger picture and react accordingly.
For us building StarCraft II bots, possible use cases:
Better decision-making. Move beyond rigid scripts—your bot can actually think about its next steps.
Explainability. Instead of black-box decisions, LLMs tell you why they made a move, making debugging easier.
Opponent modeling. Instead of reacting, your bot could predict what the enemy will do next—and counter it.
LLMs may not have AlphaStar’s reflexes, but they bring something new to the battlefield—the ability to think. And that could be a game changer.
Code Ideas
You can give TextStarCraft II a shot yourself, when the paper was first done it was done with ChatGPT 3.5 turbo, but now with something like 03 mini or Deepseek, there are alot more options to make it more practical 🔗 LLM TextStarCraft II
