Maven 智能系统也许是你从未听过却最重要的武器系统之一：它已发现伊朗导弹、也门火箭发射器、美国南部边境移民和加勒比海毒船，并在 2022 年某一天找出了超过 260 个乌克兰潜在目标。它不仅能感知，还能融合照片、文字、无线电和电磁脉冲等情报，把数据快速转化为打击行动。

该系统在不同战区的表现差异明显：2017 年在索马里测试时，算法把云标成校车；2018 年在阿富汗又把树当成人、把石头当建筑，但后来逐步改进。一个分析员用 40 秒才发现一名牧羊人，而 Maven 在同一视频流上不到 1 秒就识别出来；2021—22 年间，超过 1,500 个算法被筛到只剩 24 个用于乌克兰。

乌克兰战争是 Maven 的关键转折点：它以每月 100 万美元的云计算费用，工业化地向乌克兰提供所谓“兴趣点”，而大型语言模型又把目标识别流程提速 5 倍，使美国每天可识别并打击 5,000 个目标。书中还指出，许多模型在阿富汗成功率为 70%，到菲律宾则降至 30%；而在乌克兰，即便部署已久，系统每评估 1 平方公里仍会产生 10 个错误检测。

Maven may be the most important weapons system you have never heard of: it has found Iranian missiles, rocket launchers in Yemen, migrants on America’s southern border, and drug boats in the Caribbean, and on one day in 2022 it located more than 260 potential targets for Ukraine. It does not only sense; it also fuses photos, text, radio, and electromagnetic pulses into action, turning data into strikes with a single human click.

Its performance varies sharply by theater: early tests in Somalia in 2017 mislabeled clouds as school buses, and in Afghanistan in 2018 it mistook trees for people and rocks for buildings, though it improved over time. One analyst took 40 seconds to spot a shepherd that Maven detected in less than a second, and in 2021-22 more than 1,500 algorithms were pared down to just 24 for use in Ukraine.

Ukraine was Maven’s pivotal moment: at a cloud-computing cost of $1 million per month, it fed Ukraine industrial-scale “points of interest,” while large-language models sped targeting up five-fold and enabled the U.S. to identify and hit 5,000 targets per day. The book also notes that models with a 70% success rate in Afghanistan fell to 30% in the Philippines, and that in Ukraine the system still produced 10 false detections for every square kilometre assessed.

Source: The AI that transformed American warfare

Subtitle: Maven not only identifies targets—it tells commanders how to attack them, too

Dateline: 5月 14, 2026 11:24 上午