Chubby♨️@kimmonismus · XAI 摘要
据早期测试，GPT-5.6（代号 Sol）在30小时内已超越 Claude Opus 64小时运行达到的加速效果，尽管其初期探索更慢、失败更多、写代码更少。Sol 采用不同策略：不使用低精度，而是借助集群/DSMEM 和创新的数值方法取得优势，后续将转向低精度并利用 Tensor Cores（TC）。当前在某个排行榜位列第7。开发者认为这种“更少暴力探索、更持久研究”的思路令人期待。
First reports about GPT-5.6；
The interesting thing about GPT-5.6 Sol is not that it looks cleaner than Opus. It apparently doesn't.
It progresses slower， fails more often， explores fewer ideas and writes far less code.
But after 30 hours it already surpassed the speedup of Claude Opus' 64-hour run.
Less brute-force exploration， more persistent research. Not prettier output， better optimization.
Sounds very promising！ Excited for next week！
30 hours in and Sol has surpassed the speedup of the 64 hour Opus run Very different approaches; Sol isn't using low precision but is leveraging clusters/DSMEM ...