System One models like Jev are fast general-purpose classifiers that can handle multiple tasks through prompting, similar to how large language models work. Unlike traditional classifiers built since 1958 that need separate training for each specific job, these newer models can sort email, filter Slack notifications, or do dozens of other things from a single prompt. The catch is they're slower and more expensive to run than hand-built classifiers because they have to encode knowledge about irrelevant tasks in their weights just to maintain flexibility. The real insight here is that System One models solve two practical problems for engineering teams. First, most teams lack the expertise to train custom classifiers from scratch—it's genuinely outside their skillset. Second, building a custom classifier requires assembling massive labeled datasets, which is expensive and risky before you know if a feature will actually work. A System One model lets you validate an idea cheaply by just writing a prompt and seeing if it performs well enough. Once you've proven a feature works with a System One model, the path forward becomes clear: you can trivially collect pairs of inputs and outputs from the model's real-world usage, then use that data to train a specialized classifier that's faster and cheaper for your specific task. This creates a natural progression where System One models act as a low-friction entry point, and successful use cases graduate to optimized replacements. If these models gain traction, this distillation pattern—from general to specific—will likely become standard practice.