A day of controlling a Mac by voice with Jev
Opening apps, switching windows and searching sites took about 700ms, but Jev only picks from given options and returns a probability, so quality depends almost entirely on how the choices are designed.
Spent a day using @typesafeai's Jev to control my Mac by voice. Honest take: It is extremely fast. Opening an app, switching to one, searching a site, all landing in around 700ms. Those same decisions were costing me six seconds before. It also refuses what it can't do in about 350ms, which matters more than it sounds when the thing is clicking inside your apps. But you can't build with it like a chat model. It doesn't write anything back, it picks from the options you give it and returns a probability. So almost all the quality is in how you set up the choices. I got it wrong four times in a row and every single time it was my question, not the model. Two options that mean nearly the same thing will split its answer and come back looking like uncertainty. Knowing when to ask for a pick between options and when to ask a plain yes or no is most of the skill. It's also early. The moment a command needs even a little reasoning, it struggles. Where I landed: computer use is a genuinely great fit for this. None of the demos going around are pure Jev either, the popular macOS one uses a small Claude call to produce the URL. That combination is the interesting part. Jev doing the fast picking, one light LLM call for the bit that needs reasoning. Cheap, and fast enough to feel instant.