在二月，澳大利亚初创公司Cortical Labs公布其“生物计算机”展示：一款装载在硅芯片上的2,0,000个（20万）人体神经元模型成功玩了《Doom》，此前其已让神经元集合玩过更简单的《Pong》；该公司声称这类生物计算单元未来可放入商用数据中心标准机架，与过去半个世纪主导计算的晶体管硅芯片并存或竞争。其目标并非游戏本身，而是让活体神经元在真实世界AI任务中成为可扩展的计算组件。

Cortical的系统核心是由数千个微电极构成的阵列：神经元源自人类供体干细胞培养，阵列可让传统计算机读取其电脉冲并向其刺激信号；通过供氧供养与废物排出循环系统，细胞可存活最长6个月。该设备被封装成适配标准服务器机架的机箱，以便未来在数据中心直接部署，形成“生物芯片+液体养护”一体化运行模式。

其技术论证强调能效与架构优势。人类大脑约900亿神经元只耗约20瓦，而当前AI模型常用“数百万瓦”；而记忆芯片厂商Micron估计，传统AI处理器高达一半能耗用于数据搬运而非计算。与只具备通断两态的晶体管相比，神经元受膜电位、信号间隔等多变量影响，能更自然地处理噪声、连续的模拟信号，缓解数据与计算分离带来的“拥堵”。

In February, Australian startup Cortical Labs announced a biological computer made from 200,000 human neurons on a silicon chip that could play Doom, after previously training neurons on the much simpler Pong, and argued these living units could eventually sit in server racks beside conventional transistor chips rather than remain a science curiosity. The ambition is not game entertainment but a new class of compute substrate for AI workloads in real-world settings.

The platform uses thousands of microelectrodes to interface stem-cell-derived human neurons with conventional computers, with perfusion systems that supply oxygen and nutrients and remove waste for up to six months of operation, and it is packaged to fit standard commercial data-centre racks. The efficiency argument is explicit: a human brain of roughly 90bn neurons uses about 20 watts, while AI models can draw millions of watts, and Micron estimates that up to half of an AI processor’s energy budget can be consumed moving data between separate storage and compute units.

Cortical and chief scientific officer Brett Kagan frame this as a potential advantage for noisy, analog tasks, including Moravec’s paradox scenarios where abstract math is easier for machines than motor tasks like making a cup of tea. Adoption is still experimental: 5,500 users have already accessed internet-connected systems; a March 10 agreement with DayOne will place 20 bio-computers at the National University of Singapore; and despite heavy industry investment of hundreds of billions in standard electronics, DARPA’s March 3 biological computing program signals interest in low-energy biological processing units for autonomous drones and messy physical-world AI tasks.

Source: Why a startup is teaching human brain cells to play “Doom”

Subtitle: They could herald a new type of computing

Dateline: 4月 01, 2026 03:22 上午