Mustafa Suleyman· X:Mustafa Suleyman(Microsoft AI CEO) (@mustafasuleyman)·· 2 小时前AI 评分38
Microsoft AI 的 Mustafa Suleyman 称用 1/12 训练算力在代码上追平 DeepSeek-V4-Pro-Base
AI 导读
Microsoft AI 的 Mustafa Suleyman 表示,在留出代码评测上,其团队用少 12 倍的训练算力追平了 DeepSeek-V4-Pro-Base 的预训练分数。他将此归因于可复合的效率提升对 scaling laws 的改写。
正文
Scaling laws, scaling... with compounding efficiency!
On held-out code, we matched DeepSeek-V4-Pro-Base's pre-trained score using 12x less training compute.
来源:Mustafa Suleyman · x.com