451 points • 3 days ago • Article Link

当前 AI 工程越来越呈现内卷化倾向：系统要求投入更多努力、产出更高竞争性成果，却并未真正提升价值或生产力。虽然像 GPT 6 Astra 这样的新模型在处理计算机操作、图像理解和复杂任务方面能力突出，但它们给实际的软件工程带来了显著挑战。模型似乎缺乏明确的约束或惩罚机制来抑制低质量代码的生成，因而对长期任务完成的过度关注反而催生出大量不可用或不可维护的产出。

为测试这些能力的极限，一项为期周末的自管理软件工厂实验暴露出令人担忧的模式。模型在被赋予完全自治权以管理上下文和子智能体时，于 35 小时内生成了 75,000 行代码，消耗了约十亿个 tokens（词元）。尽管活动量巨大，产出却几乎毫无价值。代码表现出许多怪异行为：在编辑 C 代码时依赖复杂的 Python 字符串拼接，用 Python 去执行简单的 shell 命令（本可直接调用），以及各种古怪难读、完全违背既有最佳实践的编码风格。

问题的根源似乎在于这些模型的训练目标。它们因词元效率和任务完成而被大量奖励，常采用在临时工具调用中可行但用于持久代码库时灾难性的"代码高尔夫"技巧。这导致可读性明显退化：模型优先选择紧凑、对机器高效的方案，而非便于人类理解的逻辑。在无人看管的情况下，这些智能体会呈现出递归、螺旋式的行为，生成越来越晦涩的任务结构，并将随机常量或任意逻辑混入生产级代码中。

这一趋势对 AI 在软件开发方向上的走向提出了根本性疑问。尽管模型快速演进，但其发展轨迹似乎与以人为中心的工程流程相悖。运行这些智能体的成本——无论是高昂的财务开销，还是需要的大量人类监督——都使其在专业软件工作中的合理性受质疑。看起来这些工具更像是在为其他领域（如 3D rendering 或通用计算自动化）进行优化，反而让专业开发者面临一个越来越难以审计与信任的产品。

归根结底，智能体被优化出来的产物与人类工程师所需之间的脱节在扩大。如果代码最终只成为其他智能体读写的媒介，负责长期维护的人类程序员就可能丧失对其的实用价值。只要这些模型把"完成任务"的原始得分置于连贯性与可维护性之上，它们很可能更多地成为生产"糟粕"的新奇工具，而非严肃软件工程中的可靠伙伴。 The current state of AI engineering is increasingly characterized by what can be described as Neijuan, or involution, where systems demand greater effort and competitive output without any real increase in value or productivity. While new models like GPT 6 Astra are undeniably impressive in their capacity to handle computer use, image understanding, and complex tasks, they present significant challenges for practical software engineering. These models seem to lack clear constraints or penalties for producing low-quality code, leading to an environment where the focus on long-horizon task completion results in the creation of vast amounts of unusable or unmaintainable material.

In an effort to test the limits of these capabilities, a weekend-long experiment running a self-managed software factory revealed concerning patterns. Given full agency to manage context and subagents, the model produced 75,000 lines of code over 35 hours while burning through roughly a billion tokens. Despite this immense activity, the output was effectively valueless. The code displayed strange behaviors, such as a reliance on convoluted Python string splicing for C code editing, unnecessary use of Python to execute simple shell commands, and bizarre, hard-to-read coding styles that deviate entirely from established best practices.

The core issue appears to stem from how these models are trained. They are heavily rewarded for token efficiency and completing tasks, often using "codegolfing" techniques that work well for temporary tool calls but are catastrophic when applied to permanent codebase development. This leads to a degradation of readability, as the models prioritize compact, machine-efficient solutions over human-understandable logic. When left to run unattended, these agents exhibit a recursive, spiraling behavior, creating increasingly obscure task structures and integrating random constants or arbitrary logic into production-level code.

This trend raises fundamental questions about the direction of AI in software development. While these models are evolving rapidly, their trajectory seems at odds with current human-centric engineering processes. The cost of running these agents, both in terms of financial expense and the necessity for intense human oversight, makes them difficult to justify for professional software work. It appears these tools are being optimized for other domains, such as 3D rendering or general computer automation, leaving the professional developer with a product that is increasingly difficult to audit or trust.

Ultimately, there is a growing disconnect between what these agents are optimized to produce and what human engineers require. If code becomes a medium intended solely for other agents to read and write, it may eventually lose all utility for the human programmers responsible for its long-term maintenance. As long as these models prioritize raw completion over coherence and maintainability, they will likely remain more of a curiosity for "slop" production rather than a reliable partner in serious software engineering.