Claude Tops Singularity-Bench With Five Wins and One Draw
The Singularity-Bench preprint evaluates six AI-development tasks covering training, data selection, reinforcement-learning code, inference engines, agent runtimes, and GPU cluster scheduling. New智元 and AIbase report that Claude Opus 5.5 with Claude Code was the only combination above the human baseline, with five wins and one draw, while its generated inference engine reduced P90 first-token latency from 12.029 seconds with SGLang to 0.478 seconds on matched hardware. The researchers note that humans still defined each task, so the results do not yet demonstrate end-to-end recursive self-improvement.
The benchmark provides a direct evaluation of agents performing AI infrastructure and development work, while the inference-engine comparison offers concrete evidence that agent-generated systems can outperform an established engine under matched hardware and workloads.
原创 ASI启示录 2026-11-20 00:18 北京
新智元报道
你让AI帮忙修一个登录报错。
照着它的建议改完代码,你重新跑了一遍登录测试。终端又弹出一行红字,报的还是同一个错。
你把新日志贴进对话框,再按下回车,等AI告诉你下一步查哪儿。
#Claude Opus 5.5#Claude Code#Singularity-Bench#SGLang#Carnegie Mellon University#NVIDIA