Not every job deserves a flagship model with full thinking capabilities.

Way before AI, humans distilled a LinkedIn voice so thoroughly that a few-hundred-million-parameter model — running in your browser, on WebAssembly or WebGPU — can rewrite your text in that style. No remote “thinking” cluster. If a constrained tab can do this, imagine what a native app and a bit more creativity in the harness could do, instead of throwing heavier models at every problem.

Data centers are already bumping into U.S. grid limits: years-long interconnection queues, turbines on backorder, chips that want more watts than substations can promise. Prompts matter as much as parameter counts.