90% of AI tokens will be asynchronous | Matan Grinberg, Factory

讲者强调了一个重要且令人兴奋的趋势：AI消费的指数级增长，这正直接为AI供应商带来收入的飙升。然而，他们随即对这种消费的性质提出了一个关键区分，将当前大部分的使用归类为“同步”的。 正如所描述的，同步AI的使用涉及直接、实时、由人类发起的交互。这包括常见的场景，例如用户与Claude Code等对话式AI模型互动，调用Codex等AI编程助手，或者向Droid等AI工具分配任务。讲者提出了一个鲜明的比喻来强调这一点：如果每个人明天突然都无法使用或“生病了”，那么大部分云端代码的使用量将基本归零。这生动地说明了当前许多AI操作对即时人类输入和指令的依赖性。本质上，当前的AI主要以“副驾驶模式”运行，即在人类用户发出提示后才辅助或执行任务。 讲者预见的重点在于对近期未来的一个大胆预测：在未来12到24个月内，他们预计将发生根本性转变，绝大多数（高达90%）的AI“代币”（代表AI处理和活动的底层单位）将从同步使用转向“异步”使用。 异步代币标志着真正自主AI代理的到来，讲者将其称为“机器人”（droids）。与那些等待明确指令的同步AI不同，这些异步实体将独立、主动地运行，并且其每次操作都不需要实时的M人工干预。这标志着AI从单纯的辅助转向真正的自主导向智能。 讲者提供了这些自主机器人（droids）将如何运作的引人注目的例子： * “嘿，这是我从客户那里发现的一些信号，我们去解决它吧。”这说明了AI代理如何独立监控数据，识别关键洞察（客户信号），然后主动建议或启动纠正措施，而无需被特别要求寻找该信号。 * “我们来创建一个初步的解决方案。”这表明AI不仅能够识别问题，还能自主制定并启动初步解决方案的开发，进一步体现了其主动性和解决问题的能力。 这一演变标志着讲者所称的“代理原生应用”的开始。它代表了AI能力的成熟，其中系统本身能够根据对环境的理解和预设目标启动流程，而不仅仅是响应人类命令。 虽然讲者承认当前“智能体式”的交互——即用户给AI一个复杂任务（“嘿，帮我做这个”）并由AI以最少的来回沟通来执行——比简单的对话式查询更为复杂，但他们强调这些交互本质上仍然是由人类发起的。即使在这些先进的副驾驶场景中，也是由人类“启动”的。真正的“代理原生”未来，设想的是AI系统能够独立识别需求、发现问题并启动解决方案，从而推动AI如何融入和影响各个领域发生深刻变革。这种从反应式辅助到主动、自主操作的转变，必将重新定义AI消费的格局及其影响力。

The speaker highlights a significant and exciting trend: the exponential growth in AI consumption, which is directly translating into wild revenue increases for AI providers. However, they immediately introduce a crucial distinction regarding the nature of this consumption, categorizing much of the current usage as "synchronous." Synchronous AI usage, as described, involves direct, real-time, human-initiated interactions. This encompasses common scenarios like users engaging with conversational AI models such as Claude Code, invoking AI coding assistants like Codex, or directing tasks to AI tools like Droid. The speaker offers a stark analogy to underscore this point: if everyone were to suddenly become unavailable or "woke up sick tomorrow," a substantial portion of cloud code usage would effectively drop to zero. This vividly illustrates the present dependency of many AI operations on immediate human input and instruction. In essence, current AI largely functions in a "co-pilot mode," where it assists or executes tasks *after* being prompted by a human user. The core of the speaker's foresight lies in a bold prediction for the near future: within the next 12 to 24 months, they anticipate a radical shift where an overwhelming 90% of AI "tokens"—the fundamental units representing AI processing and activity—will transition from synchronous to "asynchronous" usage. Asynchronous tokens represent the advent of truly autonomous AI agents, which the speaker refers to as "droids." Unlike their synchronous counterparts that wait for explicit commands, these asynchronous entities will operate independently, proactively, and without the need for real-time human intervention for every action. This signifies a move beyond mere assistance to genuine self-directed intelligence. The speaker provides compelling examples of how these autonomous droids will function: * "Hey, here's some signal that I found from a customer, let's go fix it." This illustrates an AI agent independently monitoring data, identifying critical insights (customer signal), and then proactively suggesting or initiating corrective actions without being specifically asked to look for that signal. * "Let's go create a first-pass solution to this." This suggests an AI not only identifying a problem but also autonomously formulating and initiating the development of an initial solution, further exemplifying its proactive and problem-solving capabilities. This evolution marks the beginning of what the speaker terms "agent-native stuff." It represents a maturation of AI capabilities where the systems themselves initiate processes based on their understanding of the environment and predefined objectives, rather than simply responding to human commands. While acknowledging that current "agentic" interactions—where a user gives an AI a complex task ("hey, go do this for me") and it executes it with minimal back-and-forth—are more sophisticated than simple conversational queries, the speaker emphasizes that these are still fundamentally initiated by a human. Even in these advanced co-pilot scenarios, a human is "kicking it off." The true "agent-native" future envisions AI systems that independently identify needs, discover problems, and initiate solutions, driving a profound transformation in how AI integrates into and influences various domains. This shift from reactive assistance to proactive, autonomous operation is poised to redefine the landscape of AI consumption and its impact.