Quoting Andrej Karpathy 引用安德烈·卡帕西

Andrej Karpathy just articulated the future, and it’s not about one killer app. It’s about the death of the app as a fixed concept. His observation that software will “increasingly come out on a tap” is the most succinct description yet of a seismic shift we’re mid-quake in. He invokes the Jevons Paradox, and he’s right. When the cost of producing something drops to near-zero, consumption doesn’t fall—it explodes. But the crucial insight here isn’t about demand for more software. It’s about the demand for different software: ephemeral, hyper-personal, and fundamentally disposable.

We’re not entering an era of better software. We’re entering an era of software as a fluid utility. Think about it: the bespoke “wandb for your project” is the perfect example. That’s not a product you download, update, or even care about maintaining. It’s a temporary scaffold you conjure to solve a specific, immediate cognitive problem—a complex visualization, a custom analysis pipeline. You use it, you learn from it, and you likely discard it when the context shifts. The value isn’t in the artifact; it’s in the momentary alignment of your specific need with the machine’s ability to fulfill it. This is the end of software as a noun and the beginning of software as a verb.

The current market is still obsessed with building the next monolithic platform, the next walled garden. That entire paradigm is a legacy of an era where creation was expensive and distribution was linear. Now, creation is a conversation. The competitive advantage is no longer in owning the best static tool, but in providing the most malleable and responsive creative substrate. Companies still pitching “a solution” are playing yesterday’s game. The winners will be those who sell the most powerful, intuitive, and safe set of primitives that allow a user to assemble their own solution in real-time, without thinking about it as “software engineering.”

This does create a brutal efficiency filter. The Jevons effect means my demand for code will be insatiable. I’ll want a custom dashboard for this research thread, a specialized data-cleaning script for that dataset, an explainer app for a dense paper. But this flood of generated code is a tool for thought, not a final product. The danger isn’t that the AI will write bad code; it’s that we’ll become addicted to its immediate utility and stop questioning its foundational design. A bespoke wandb instance you generate in five minutes is a miracle of productivity, but it’s also a black box of tech debt you’re choosing to incur. The real skill, the premium human judgment, will shift from writing code to curating and orchestrating this on-demand code generation—to knowing which scaffolding to build, how to interrogate its outputs, and, crucially, when to delete it.

Karpathy’s “free your mind” reference is spot on, but perhaps in a more unsettling way than the optimistic sci-fi trope suggests. It’s free to imagine any tool you want, yes. But it also means freeing yourself from the comforting constraints of a pre-built workflow. We’re moving from a world of predefined paths to a world of infinite, generative pathways. For the technical user, this is empowerment on steroids. For the organization, it’s chaos. How do you audit, secure, or maintain a thousand bespoke, ephemeral applications that exist only in one user’s session? The old IT governance model will crumble.

So, yes, our demand will explode. We’ll all become low-code tyrants, demanding perfectly tailored digital levers for every minor cognitive task. The paradox is that in making software infinitely accessible and disposable, we may simultaneously make it more powerful and more ephemeral, more critical and more invisible. The final product is no longer the point. The point is the sustained, intimate dialogue between human intent and machine execution. We’re not downloading tools anymore. We’re summoning them. And we’re just beginning to reckon with the consequences.

安德烈·卡帕西刚刚清晰地阐述了未来图景，其核心并非某款杀手级应用，而是“应用”作为固定概念的消亡。他指出软件将“越来越多地通过一次点击呈现”，这是迄今为止对这场我们正身处其中的地壳板块剧变最精辟的描述。他援引了杰文斯悖论，而这判断完全正确：当某种产品的生产成本降至接近零时，消耗量不会减少——反而会爆发式增长。但这里的关键洞见并非指向对更多软件的需求，而是指向对不同类型软件的需求：转瞬即逝、高度个性化、且本质上是可抛弃的。

安德烈·卡帕西刚刚清晰地阐述了未来图景，其核心并非某款杀手级应用，而是“应用”作为固定概念的消亡。他指出软件将“越来越多地通过一次点击呈现”，这是迄今为止对这场我们正身处其中的地壳板块剧变最精辟的描述。他援引了杰文斯悖论，而这判断完全正确：当某种产品的生产成本降至接近零时，消耗量不会减少——反而会爆发式增长。但这里的关键洞见并非指向对更多软件的需求，而是指向对不同类型软件的需求：转瞬即逝、高度个性化、且本质上是可抛弃的。

我们并非进入一个拥有更好软件的时代，而是进入一个软件成为流动式效用的时代。试想：为你的项目定制的“wandb”（注：一个实验跟踪工具）就是完美例证。这不是一个需要下载、更新甚至费心维护的产品，而是一个你临时召唤出来、用于解决特定即时认知问题的临时支架——比如复杂的可视化、定制分析流程。你使用它，从中学习，并且很可能在场景转换后将其丢弃。价值不在于这个工具实体本身；而在于你的特定需求与机器满足该需求的能力在瞬间达成的契合。这标志着软件作为名词的终结，以及软件作为动词的开端。

当前市场仍痴迷于构建下一个巨型平台、下一个封闭的围墙花园。这一整套范式源自一个创造成本高昂、分发方式线性的旧时代。如今，创造即对话。竞争优势不再在于拥有最佳的静态工具，而在于提供最具可塑性和响应性的创作基底。那些仍在推销“一套解决方案”的公司，仍在玩昨日的游戏。最终的赢家将是那些提供最强大、最直观、最安全的基础原语集的公司，它们允许用户实时组装出属于自己的解决方案。