Co-Design Bench: agents finish 26%, PDD Kit — AI News Nov 9
Co-Design Bench: agents finish 26%, PDD Kit — AI News Nov 9
重點導讀
- Alibaba's Co-Design Bench: the best coding agent fully finishes only 26.4% of real change requests, and why partial-credit benchmarks mislead agent evals. - Microsoft PDD Kit: proof-driven development with Verus specs for agent-written Rust code, and what stays outside the proof. - Anthropic's research models publish physics manuscripts, plus the $3.4B generator deal, DeepSeek's 1.54M daily agent users, GPT-6.2, the Anthropic IPO, Apple Siri, the NVIDIA board probe, Suno lawsuits, and AutoDNA.
AI 伴隨問答
針對這篇報導有疑問?可直接點選快捷問題或向 AI 追問背景脈絡與技術細節:
Podcast 專家音訊節目
收錄於 888news Podcast 智庫專區,包含精華摘要與上市公司即時行情
正文
- Alibaba's Co-Design Bench: the best coding agent fully finishes only 26.4% of real change requests, and why partial-credit benchmarks mislead agent evals. - Microsoft PDD Kit: proof-driven development with Verus specs for agent-written Rust code, and what stays outside the proof. - Anthropic's research models publish physics manuscripts, plus the $3.4B generator deal, DeepSeek's 1.54M daily agent users, GPT-6.2, the Anthropic IPO, Apple Siri, the NVIDIA board probe, Suno lawsuits, and AutoDNA. Follow AI Engineering Briefing and leave a rating — it's how new listeners find the show.
來源: AI Engineer (redcircle.com)
開啟原文