@bojie_li: I want to change my perspective on Jev-type models a bit. When it was first released, I thought it was just a classific…

I want to change my perspective on Jev-type models a bit. When it was first released, I thought it was just a classification model—existing LLMs could decode a single token and be done with it. Now I realize it's not technically difficult; the key is the System 1 thinking: the world is event-driven, making just a small, simple decision at a time. First off, previously, LLMs were used inside agent react loops, and we'd try to avoid using LLMs in the outer harness and business logic as much as possible, otherwise the costs and latency would be unbearable. Jev and the Pine Computer logic released just today suggest that as long as the model is fast enough and cheap enough, a lot of business logic that used to be patched together with if-else statements can now run on models. Thinking deeper, human real-time decision-making is actually more like the System One model, rather than autoregressive LLMs. - In voice conversations, humans make interaction judgments like interrupting, agreeing, or pausing every few hundred milliseconds—this is key to making AI speech sound human. In full-duplex models like GPT-Live 1 and TML interaction model, it only outputs a few tokens per few-hundred-millisecond time slice. - In computer use, humans decide where to click next every few hundred milliseconds; once the UI changes, they don't click the originally prepared spot anymore. Pine Computer turns computer use from turn-based to event-driven, simplifying outputs from coordinates to labels as much as possible. - In robotics, humans are planning the next action or stopping the current one every few hundred milliseconds. In interactions, the frequency of the observe-feedback loop is likely more important than IQ. Full-duplex models, Jev, and Pine Computer all point toward a new direction: SOTA background models handle deep thinking and problem-solving, while System One Models handle real-time voice interactions, computer use, business logic judgments, and even future embodied operations.