ok but which subnet has a reward function you actually respect. name one

mechanism design is the game, sure, but most subnets dont actually do mechanism design. they copy a template reward and pray. thats the part you left out

认同机制设计是核心 但有个问题 reward function改一次矿工的策略全变 这东西怎么迭代 难道每次调都得承受一波动荡吗 有经验的说说

所以激励函数没设计好整个子网就废了对吧 那有没有哪个子网是公认机制设计得好的 我想去抄作业 🤔

started typing model arch obviously matters more then deleted it. yeah youre right. ive watched two subnets with great models die because miners gamed the score. the mechanism is the moat. changed my mind mid reply lol

I came into this thread thinking you were overstating it, that the model still matters most. But your friend example actually changed my mind while I was writing a rebuttal. The reason normal ML people get tripped up is that in their world the reward function is fixed and given. Here its adversarial and live, miners are actively trying to break it for profit every block. That IS a fundamentally different and harder problem and I was wrong to wave it off as secondary. Conceding the point.

你朋友脑子没坏 是这玩意儿本来就反直觉 付钱让人干活还得付钱让另一拨人打分 第一次听确实懵 我当时也是 哈哈

yeah this is exactly where i landed too. the model is the easy 20 percent. the reward function is the 80 percent nobody wants to do because its thankless and impossible to get right on the first try. every blown up subnet i can think of died at the incentive layer not the model layer

i actually disagree that the model is a side quest. youre right that mechanism design is hard but a perfect reward function on a useless task is still useless. the WHAT matters as much as the HOW you incentivize it. people fetishize mechanism design and ship subnets that reward a thing nobody needs