Why "Tokens Aren't Fungible" - Anthropic's Angela Jiang

演讲者概述了一个渐进式的AI发展框架，该框架超越了简单的信息检索，迈向了更复杂的“执行”和“协调”层面。这种多层抽象代表了他们在使用Claude等AI模型方面的工作战略路线图。 讨论的第一个也是基础的层次是**知识层（Knowledge Layer）**。在这一层，AI的主要功能是“了解事物”——即处理信息、理解查询并提供答案。尽管这对基础理解和信息整合至关重要，但仅凭这种能力，被认为不足以满足公司所设想的更广泛的工作范围。 下一个重要的抽象层，也是团队投入越来越多时间的领域，是**执行层（Execution Layer）**。这一层超越了单纯的知识；它的目标是让Claude“执行工作”。这意味着AI执行主动任务，而不仅仅是回答问题。例如，生成特定输出，主动“在多个不同系统中编辑文件”，以及执行需要与外部环境交互的操作。这种从“知”到“行”的转变引入了显著的复杂性，需要强大的基础设施来管理这些操作。演讲者将这一层描述为由“低级线束（low level harness）加上托管基础设施（managed infrastructure）”组成。该公司目前针对这一需求的高级产品是**Claude托管代理（Claude Managed Agents）**，旨在促进这些更高层次的执行任务。 展望未来，演讲者指出了一个更先进的层次：**协调层（Coordination Layer）**。这一层被构想为“位于”执行层之上，充当一个“元线束（meta harness）”，负责协调AI活动。虽然低级线束专注于直接执行，但协调层引入了“策略”的概念。这里的核心思想是，AI计算单元或“tokens（令牌）”并不总是“可互换的（fungible）”；相反，它们可以被赋予不同的角色。例如，一些令牌可能专注于“提供建议”，另一些专注于直接“执行”，还有一些专注于“构思”或“产生创意”。协调层的目的是组合和管理这些不同的角色，使AI能够形成“协调策略”，让不同的组件协同工作以实现一个更大、更复杂的目标。 演讲者强调，这些层次是相互关联且“逐级递进”的。协调层提供全面的战略方向，决定“做什么”。这一方向随后被传递给执行层，由其执行实际任务，并利用知识层获取必要信息。这种渐进式的堆叠确保了每个组件都在一个更大、更集成的系统中发挥作用。 该公司的路线图明确表明了战略重点的转变：在他们将发布的抽象和产品方面，他们计划“越来越多地从知识层转向执行层，再从执行层转向协调层”。这一演变反映了构建AI系统的历程，这些系统不仅能够理解信息，还能主动执行复杂任务、管理错综复杂的工作流程，并智能地协调各种功能，最终实现更自主、更有能力的AI代理。

The speaker outlines a progressive framework for AI development, moving beyond simple information retrieval to more complex levels of "execution" and "coordination." This multi-layered abstraction represents a strategic roadmap for their work with AI models like Claude. The initial and foundational layer discussed is the **Knowledge Layer**. At this level, the AI's primary function is to "know stuff" – to process information, understand queries, and provide answers. While crucial for foundational understanding and information synthesis, this capability alone is seen as insufficient for the broader scope of work the company envisions. The next significant layer of abstraction, and where the team is dedicating an increasing amount of time, is the **Execution Layer**. This layer transcends mere knowledge; it's about getting Claude to "execute work." This involves the AI performing active tasks, not just answering questions. Examples include generating specific outputs, actively "editing files in a bunch of different systems," and performing actions that require interaction with external environments. This transition from knowing to doing introduces significant complexity, necessitating robust infrastructure to manage these operations. The speaker describes this layer as comprising a "low level harness plus managed infrastructure." The company's current high-level product addressing this need is **Claude Managed Agents**, designed to facilitate these higher-order execution tasks. Looking ahead, the speaker identifies a future, even more advanced layer: the **Coordination Layer**. This layer is conceived as sitting "on top of" the execution layer, acting as a "meta harness" that orchestrates AI activities. While the low-level harness is focused on direct execution, the coordination layer introduces the concept of "strategies." The core idea here is that AI computational units or "tokens" are not always "fungible"; instead, they can be assigned distinct roles. For instance, some tokens might specialize in "advising," others in direct "executing," and still others in "dreaming" or ideation. The coordination layer's purpose is to compose and manage these diverse roles, enabling the AI to form "orchestrated strategies" where different components work together towards a larger, more complex goal. The speaker emphasizes that these layers are interconnected and "ladder together." The coordination layer provides the overarching strategic direction, determining "what to do." This direction is then passed to the execution layer, which performs the actual tasks, leveraging the knowledge layer for necessary information. This progressive stacking ensures that every component serves a purpose within a larger, integrated system. The company's roadmap clearly indicates a strategic shift in focus: they plan to move "more and more from the knowledge layer to the execution layer and from the execution layer to the kind of coordination layer" in terms of the abstractions and products they will release. This evolution reflects a journey towards building AI systems that can not only comprehend information but also actively perform complex work, manage intricate workflows, and intelligently coordinate diverse functions, ultimately leading to more autonomous and capable AI agents.