stochasm(@stochasticchasm):headliner here and most surprising thing to me is that the model improved so much in 30 RL steps. way less steps than i'd imagine would be needed here for such a dramatic increase. one wonders why they didn't keep going, there's like no sign of a plateau

stochasm

@stochasticchasm

pretraining lead @arcee_ai • opinions my own

加入 August 2024

1.8K 正在關注 9.4K 粉絲

stochasm@stochasticchasm

2026.09.21 21:12

headliner here and most surprising thing to me is that the model improved so much in 30 RL steps. way less steps than i'd imagine would be needed here for such a dramatic increase. one wonders why they didn't keep going, there's like no sign of a plateau