Imagine controlling thousands of AI agents in a complex strategy game, all through simple conversations. Recent research explores this exciting possibility with HIVE (Hybrid Intelligence for Vast Engagements), a framework that lets humans command huge AI armies using the power of large language models (LLMs). Players give high-level instructions in natural language, like "Defend the bridges!" or "Exploit their archers' weakness.", and HIVE, powered by an LLM, translates these into detailed plans, assigning behaviors and targets to each individual unit in the swarm. Researchers tested HIVE with several leading LLMs, including GPT-4 and Claude variants, on a custom real-time strategy game. They discovered that LLMs can indeed handle complex coordination, exploiting enemy weaknesses, utilizing terrain, and even following instructions related to specific map markers. However, these AI generals aren't perfect. LLMs struggled with long-term strategic planning, showing a reliance on human guidance for truly effective strategies. Another fascinating discovery was their difficulty in processing visual information. When given images of the game map, the LLMs performed worse than when provided with textual descriptions. This highlights a current limitation of LLMs: spatial reasoning in dynamic environments. Even with these challenges, HIVE demonstrates the immense potential of LLMs for human-AI collaboration in controlling massive multi-agent systems. Imagine the possibilities in areas like disaster relief, urban planning, or even complex simulations. Future research will focus on enhancing the LLMs' visual and strategic abilities, potentially paving the way for truly autonomous AI commanders leading vast digital armies.