Some notes about Jev After studying Jev for a day, the novelty it brought quickly declined, which seems like a faster universal classifier/decision-maker that LLM can fully achieve. Moreover, due to the lack of understanding of the parameter scale behind the model (which should not be very large), world knowledge may not be as comprehensive as conventional LLM. In this case, there is still a question mark on whether the decision results of its complex scenarios are truly accurate. Jev may be a good direction for some problems with finite solution space and low latency requirements, and formal output ensures correctness from a programmatic perspective. One of the most common scenarios is Computer Use, which is perfect for Jev. The Dom elements on a webpage are a finite set and can be manipulated by Jev. The loop here is equivalent to a loop of dom list ->jev action ->new dom list. However, Jev's reasoning ability and the effectiveness of computer use in long context situations are still unclear. Currently, no benchmark has been seen (intuitively, it is definitely not as good as GPT 6 Astra, but the speed is too fast). Yesterday, I tried to rewrite some modules in the Pi Agent using Jev and found that there are not many areas that can be optimized. The selection of the tool using part cannot be replaced by Jev because it is not a finite set (each step of tool using actually comes with many parameters, such as edit tool, which will have specific lines and other parameters, which need to be inferred by the model and cannot be added to Jev's decision list in advance). However, there are some places that can be optimized, such as Compaction, which can allow Jev to quickly make classifiers (LLM can also do it, but there is not much differentiation here). Another type of scenario relies on decision tree logic, for example: -Game AI -Robot -Autonomous driving Moreover, these scenarios require high real-time performance, and a major problem with traditional LLM for doing these things is that it is too slow (a large part of the scenarios lack data for training). Two demos are currently being developed: 1. Poker AI is very suitable for this scenario, and the decision tree is very long and complex. We are using Jev and other models for adversarial battles. The traditional GTO Wizard used for training is really difficult to use 2. Pokemon VGC AI， This is a serious Pok é mon doubles match, similar to a World Championship, with corresponding rules for each season. Because the data is comprehensive, it is very suitable for researching AI decisions. Plus did not have a good AI for personal training before (after completing this, I plan to use this AI in the Pok é mon Champion qualifying rounds). Both of these demo scenarios are turn based, but LLM can also be used for them.