传闻称 2026 年最大能力跃升来自 Mythos/Fable 预训练模型
这条推文讨论 2026 年预训练路线压过 RL 的争论，还提到 Sonnet 学会传统绘画，观点挺尖锐。
一条 X 帖子称 2026 年能力提升最大的事件是名为 Mythos/Fable 的大规模预训练模型，让 OpenAI 重新评估其以 RL 为主的路线，二线实验室纷纷立项 10T 参数级模型。帖中还提到大量小众 RL 环境训练使 Sonnet 已能胜任传统绘画。作者同时质疑 Yudkowsky 是否预测过在 AGI 之前出现数学与软件领域的“尖峰式”超强能力。
> literally over-updating on the month's latest news The biggest shock in capabilities in 2026 has been Mythos/Fable, a huge pretrain which sent OpenAI with all their RL scurrying back to the drawing board and made every 2nd tier lab commit to a 10T model. RL is so high largely because we're spamming niche RL envs to the point that Sonnet is a competent traditional painter now. What is Yud even gloating about? Anyway, did he even predict "spikiness" in the form of mathematical and software superintelligence before AGI?
