如何在AI生成的应用程序随着时间演变时保持其一致性？
嗨，HN，
我一直在尝试让大型语言模型（LLMs）生成并持续修改小型商业应用程序（CRUD、仪表板、工作流）。第一次生成通常是有效的，但问题往往在第二次或第三次迭代时出现。
我看到的一些反复出现的失败模式包括：
- 架构漂移，悄然破坏仪表板
- 指标在不同迭代中含义变化
- 用户界面组件以不兼容的方式查询数据
- 人工智能在局部修复某些内容时违反全局不变性
令人惊讶的是，大多数人工智能应用构建者将生成视为一次性问题，而实际应用是需要安全演变的长期系统。
我正在探索的方向是将应用程序视为运行时模型，而不是生成的代码：
- 应用程序由结构化的、版本化的JSON/DSL（实体、关系、指标、工作流）定义
- 每个人工智能提出的更改在执行前都由后端进行验证
- 用户界面组件绑定到语义概念（指标、数据集），而不是原始查询
- 人工智能提出结构；运行时强制执行一致性
从概念上讲，这种方法更接近Kubernetes对基础设施的处理方式，或语义层在分析中的工作方式——但应用于完整的应用程序，而不仅仅是报告。
我很好奇：
- 这里有没有人探索过类似的模式？
- 是否有成熟的方法来控制人工智能驱动的架构演变？
- 你认为语义层应该在应用程序运行时内部，还是应该仅限于分析？
我并不是在推销什么——我真心想了解其他人如何看待人工智能与长期应用状态的结合。
谢谢。
查看原文
Hi HN,<p>I’ve been experimenting with letting LLMs generate and then continuously modify small business applications (CRUD, dashboards, workflows). The first generation usually works — the problems start on the second or third iteration.<p>Some recurring failure modes I keep seeing:
• schema drift that silently breaks dashboards
• metrics changing meaning across iterations
• UI components querying data in incompatible ways
• AI fixing something locally while violating global invariants<p>What’s striking is that most AI app builders treat generation as a one-shot problem, while real applications are long-lived systems that need to evolve safely.<p>The direction I’m exploring is treating the application as a runtime model rather than generated code:
• the app is defined by a structured, versioned JSON/DSL (entities, relationships, metrics, workflows)
• every AI-proposed change is validated by the backend before execution
• UI components bind to semantic concepts (metrics, datasets), not raw queries
• AI proposes structure; the runtime enforces consistency<p>Conceptually this feels closer to how Kubernetes treats infrastructure, or how semantic layers work in analytics — but applied to full applications rather than reporting.<p>I’m curious:
• Has anyone here explored similar patterns?
• Are there established approaches to controlling AI-driven schema evolution?
• Do you think semantic layers belong inside the application runtime, or should they remain analytics-only?<p>Not pitching anything — genuinely trying to understand how others are approaching AI + long-lived application state.<p>Thanks.