Early observations on Jev: the value and limits of a multiple-choice primitive
The author shares seven early observations on Jev: it is purpose-built for multiple choice and is cheaper and faster, with the most promising use cases being low-cost computer use and search.
Early, raw thoughts on Jev: 1. LLMs & Agentic Systems today approach everything as a free-response question. They can technically answer multiple-choice questions effectively. But Jev was purpose-built only for multiple-choice so it can do it much better, faster, cheaper. 2. This is a really important primitive because it changes what's economical to build. Small products get more delightful and easier to scale. Consumer apps can afford generous free tiers for AI features. 3. I expect this to kick off a wave of specialized models for other common primitives & use cases. They won't compete on general intelligence. They'll compete on speed and cost for being really, really good & reliable in one narrow area. 4. The open question is where Jev clearly wins today. In some cases, a classifier trained on your data or plain if/else rules often work better. Jev's edge is that it works on any question you define, with no training data, and messy input. 5. The hard part is asking the right question & giving it sufficient evidence to answer that question. LLMs taught us to cram everything into one prompt but I don't think thats going to work with Jev. 6. Most compelling use case today: cheap computer use. It also seems to be powerful for search: faster, more accurate results and a snappier experience for users. 7. Important unlocks: (1) Fine-tuning so you get the accuracy of a custom classifier with the flexibility of an LLM (2) Multimodal input so it can make decisions from screenshots, images, and documents