Rohan Paul@rohanpaul_ai · X安德烈·卡帕西在访谈中指出，计算的未来可能从传统软件转向以神经网络为中心。传统软件需先将现实世界转化为规整符号，再由预设代码处理。而神经系统的处理顺序相反，能直接处理视频、语音、屏幕状态等“混乱”输入，实时推断关键信息并动态生成界面或行动。这不仅是用聊天替代应用，更是用能持续实时解读和渲染现实的系统取代固定流程。在此架构下，用户界面不再是稳定产品，而是为特定时刻、任务和个人临时组装的表面；传统CPU则退居类似协处理器的角色，负责算术、存储等确定性任务，而神经模型主导感知、规划和适应。
New interview of Andrej Karpathy： Says that the future of computing may be centered on neural networks instead of traditional software.
e.g. that systems could take raw inputs like video and audio and generate interfaces dynamically in real time.
Today's software assumes the world must be translated into clean symbols first， then handled by code written in advance.
A neural system flips that order.
It can take messy input like video， speech， screen state， and user history， infer what matters， and generate the interface or action on the fly.
The shift is not just replacing apps with chat， but replacing fixed pipelines with systems that continuously interpret and render reality in real time.