开发者用 Grok Bot 解释训练进度并协调 Cursor

yazin @yazins原推发布于 2026.10.02

在 X 上打开原推

that's it, it's time to HIT THE GAS i've got cursor running 24/7 to figure out how i can get correction working.. with modal spinning up 4 pairs of H100's training a bunch of different model options.. spent ~$3,000 so far (~$600 on GPUs, ~$2.5k on Cursor) training a few new models off the excellent QuranLab v3, and a few others from vanilla ZipConformer.. goal is to make correction work just as well as tracking does: ✅ high accuracy 📵 completely offline 🤏 under 100MB size to make sure this keeps going no matter what, i've got Grok Bot eli5'ing what's going on to me, and prodding the main Cursor planner (which is coordinating across all the other workers in parallel) hopefully have some results to report in a couple of days..

这条是引用着另一条写的：x.com/yazins/status/2105260872214196239