跳到正文
Lisan al Gaib· @scaling01 · X·· 22 小时前AI 评分58
AI 导读

作者 @scaling01 分析 METR 50% 时间跨度数据,指出易验证任务的时间跨度长于全部任务,且远长于难验证任务,高成本任务的时间跨度也低于低成本任务。引用的 METR Time Horizons 2.0 数据显示翻倍时间为 3.6 个月,GPT-6 Astra 为 40.3 小时、Opus 5.5 为 39.0 小时,95% CI 约从 11.9 小时到 110.2 天。

正文 · 原文

RL slop environments are winning, but Opus 5.5 shows some signs of life.

These two plots are very telling for how frontier models are trained today:
- METR 50% time horizons are longer for easy-to-verify tasks compared to all tasks, and MUCH longer for easy-to-verify tasks vs hard-to-verify tasks.
- time horizons based on higher-cost tasks are lower than for lower-cost tasks

引用Lisan al Gaib@scaling01
THEY ARE BACK! METR TIME HORIZONS 2.0 Doubling time of 3.6 months, similar to the 3.8 months estimate by OpenAI GPT-6 Astra: 40.3 hours Opus 5.5: 39.0 hours but with extremely wide 95% CI from around 11.9 hours to 110.2 days
在 X 查看被引用的帖子

来源:Lisan al Gaib · x.com