Jev for computer use hands steps back to Opus when confidence drops
In a computer-use test, Jev drove the screen with a fast model for routine steps and handed a step back to Opus when its confidence dropped, forming a fast-slow model relay.
tried Jev from @typesafeai in a computer-use this week, and in comparison with it now watching Opus drive a screen is like watching your grandma use a laptop 😅 the part I liked most is what happens when confidence drops: Jev hands the step back to Opus, so the slow model only gets called when the fast one isn't sure -- I think that's the shape most agent systems end up in, a fast model for routine steps and a big one on call, and the handoff between them is exactly where I'd want the context to travel as latents rather than as a text summary of what the fast model just saw the premise underneath Jev is the same one we started @mostik_ai on: by the end of the forward pass the model already has what it needs, and decoding that into tokens is the slow and lossy step (the difference is that they read a decision straight off that state while we hand the state to another model) glad more people are building from the same premise 🔥
This post quotes another post: x.com/aimalysheva/status/2099153882081345788