Last week, a new AI model was released, and it’s nothing like the AIs we already have. It’s called “JEV”, it’s a System One AI, and it only decides, it doesn’t generate any text. I got early access to it and I found an interesting use case: making it the Decide layer of my Assess – Decide – Do framework.
I made a video in which I show exactly how it’s done, and why this new paradigm is opening many new possibilities, check it out.
If you prefer reading, here’s what happened. JEV cannot output text, like I said, but it can answer questions, using 3 pre-defined primitives: noul (yes/no), choice (multiple options) and score (probabilities). I made a small system that builds a piece of software that you define only with one sentence. JEV picks that up, forwards it to Assess, where a LLM starts drafting the deatailed specs of what you want to build, including tests. Then it passes it back to JEV (classic ADD flow). If JEV has enough context, it passes the task forward to Do, where the LLM starts implementing. When implementation is done, it passes it back to Decide, where JEV runs the tests and decides whether or not the app is done.
In the video I made a small CLI in Python that counts the words in the files of a specific folder. But I reckon anything can be build this way. The most interesting part, at least for me, it’s the visual transition between realms, there’s something fascinating in the flashing of red, orange, green and white circles, as the action is moving between them.
I ran two tests of this system: the first one used the official JEV API, which is remote, as well as a solid LLM from OpenRouter, and it took under 2 minutes. The second one used a local implementation of JEV, and a local LLM, both way less performant, and that took almost half an hour. You can see in the video a sped up recording of that too.
I think this new paradigm of deciding with a structured output is opening a lot of new possibilities for AI. If you have any use cases, or just want more about this, please leave a comment on the video.