模型选择器已成死胡同

Lovable 发布文章阐述其“模型独立性”理念，反对让用户手动选择 AI 模型，转而通过内部“控制平面”根据任务动态分配最合适的模型 1。该控制平面可动态将构建的不同部分分配给不同模型，而非让单一模型承担全部工作 1。

Lovable 团队针对每个模型定制指令、工具和项目上下文，研究其擅长与卡点领域 1。在优化策略上，Lovable 以最终应用效果为优化单位，每次构建运行多次，并结合人工判断与 LLM 评判进行校准 1。在一次评估中，前沿模型比前代完成任务快 15%，轮次减少 40%，得分高 2–3% 1。此外，Lovable 已开始训练自有模型，目前其自训练模型在生产中处理了“有意义份额”的应用构建工作 1。

Lovable has released an article detailing its "model-agnostic" philosophy, which opposes allowing users to manually select AI models 1. Instead of having a single model handle all the work, the company employs an internal "control plane" to dynamically allocate different parts of the building process to the most appropriate models for specific tasks 1. To support this, the Lovable team customizes instructions, tools, and project context for each model, researching their specific strengths and bottlenecks 1.

Rather than relying on benchmark tests, Lovable optimizes its system based on final application outcomes, running each build multiple times and calibrating the results through a combination of human judgment and LLM evaluation 1. During one evaluation, frontier models outperformed their predecessors by completing tasks 15% faster, reducing the number of turns by 40%, and achieving scores 2% to 3% higher 1. Additionally, the company has started training its own models, which are already processing a "meaningful share" of application building work in production 1.