Andrew Carr 🤸(@andrew_n_carr):the weird part of this MiMo graph is that the judge and probe disagree 60% of the time. and for pro, disagreement goes up during training. would love to know how much of that is harder-to-judge behavior vs the model getting better at fooling the judge.

Andrew Carr 🤸

@andrew_n_carr

co-founder leading science @getcartwheel co-founder advisor @arcade_ai Past: Codex @OpenAI, Brain @GoogleAI, world ranked Tetris player

加入 July 2015

5.2K 正在關注 28.7K 粉絲

Andrew Carr 🤸@andrew_n_carr

2026.09.17 04:04

the weird part of this MiMo graph is that the judge and probe disagree 60% of the time. and for pro, disagreement goes up during training. would love to know how much of that is harder-to-judge behavior vs the model getting better at fooling the judge.