Adaptive Article Reader

Modern web applications increasingly combine two distinct classes of on-device models: fast, non-autoregressive System-1 decision models (like window.Classifier and SemanticEmbedder) and generative System-2 language models (like LanguageModel, Summarizer, Writer, and Rewriter).

While generative models excel at drafting prose and synthesizing long documents, invoking a multi-billion parameter LLM on every keystroke or scroll event drains laptop batteries and introduces hundreds of milliseconds of latency.

By placing a 15-millisecond System-1 classifier in front of generative APIs, web apps can understand user intent on every interaction and only wake up heavier generative tools with the exact configuration needed.