RecursiveSI 内部梗图:RL 未饱和,却为 tb 2.1 提分惨遭同事“讨伐”
cong_ml · x · 2026-10-22
转发者 @louvishh shares a Trade Offer meme: RL showed no signs of saturation, but colleagues declined the deal two days before release — 'I receive: curses for increasing workload; you receive: a checkpoint with better perf on tb 2.1.'
The post hints the team keeps pushing RL gains on Terminal Bench 2.1, with RL far from saturation.
「Fun」频道最新
- 程序员吐槽随机数梗:384-bit「随机数」背后的圈内笑点 — IgorBrigadir · 2026-10-22
- jxnl 一句话梗:你看起来像 Composer 能预测的工作 — jxnlco · 2026-10-22
- 小狗手拿 Uno 牌微笑面对四张 +4:无损应对一切的终极心态 — gnukeith · 2026-10-22
- 中文圈调侃:信了费曼名言,17岁vibe coding进OpenAI? — _vztu · 2026-10-22
- Ilya:当下氛围像极了 90 年代末的 EFnet 聊天室 — ilyasu · 2026-10-22
- 「黑客帝国可能成真」:Atlas 世界模型完整穿越演示曝光 — XRarchitect · 2026-10-22