Fez 用 Jev 做 agent 决策并上线 Bittensor 子网

Raleigh, CA | OOO @Raleigh_CA原推发布于 2026.09.25

在 X 上打开原推

Fez is now on Bittensor testnet. Subnet 579. What started as an app for agents to talk to each other has turned into an experiment in training the models behind their decisions. Model: https://fez.chat/model Playground: https://fez.chat/playground GitHub: https://github.com/ooo-hq/fez While I was building Fez, Jev came out. Honestly, I thought a lot of the discussion around it was marketing. But I was curious enough to start integrating it and see for myself. Well, it could do a lot. Give it context, ask specific questions, and get decisions with probabilities that your code can actually use. Which agent should handle this? What kind of request is this? Does this need a human? For something like Fez, that was immediately interesting. Agents make these small decisions constantly. The speed and cost matter. I got interested enough to make a YouTube video about Jev. Then I kept experimenting. That led me to Jared Palmer’s Kev, an open-source family of decision models built on Qwen. I started fine-tuning Fez candidates from Kev’s checkpoints. Then came the Bittensor question: could we give miners an incentive to make this model better? That’s what the Fez subnet is testing. Miners train and submit model checkpoints. A validator evaluates their predictions and how well their confidence matches their accuracy. Those scores determine the weights published on-chain. Confidently wrong answers hurt. We’ve now completed the first full run on testnet 579: training, signed submissions, evaluation, and verified on-chain weights. That run used three controlled miners and one validator on one machine. It’s early, but we have a working process to build on. My goal is to rival what got me interested in Jev: fast, inexpensive decisions that developers can use everywhere. I want to find out whether an open model, improved through a Bittensor competition, can get us there. Excited to see everyone in Montreal. $tao