Meta
Meta ProgramBench: AI Still Can't Build Large Programs from Scratch
Meta ProgramBench tests AI building programs from scratch. Top models failed, cooling 'AI builds software' hype and exposing benchmark score inflation
Back to home
SWE-bench
4 articles tagged with this topic
OpenHands
OpenHands Hits 40K Stars — Open Source Catches Up to Closed-Source AI Coders
OpenHands is an open-source AI coding agent in Docker sandbox with 40K+ GitHub stars. Open source rapidly closing gap with closed-source coding Agents
Claude Opus 4.7
Opus 4.7 来了，我并不建议你升级
Anthrop ic's Opus 4.7 removes temperature/top_p/top_k controls and inflates token counts by up to 1.35x.
Apr 173 min readopcnew Editorialjuejin.cn
LangSmith
Stop Chasing Leaderboards: How Berkeley Exposed Flawed AI Agent Benchmarks
Berkeley researchers reveal critical data contamination in top AI benchmarks. Learn how to validate your own agent tools, avoid overfitting, and build